The existing workflows remain. The same decision rights remain. The same layers remain. The same data problems remain. The same teams continue owning overlapping pieces of the same customer or process.

Only now everyone has copilots.

I do not think that is how this plays out.

AI is unusually unforgiving of organisational ambiguity. It works best when the objective is clear, the data is trusted, the process is understood and authority is explicit. Most large enterprises contain rather more ambiguity than their transformation decks suggest.

That is why one of AI’s most important effects may have very little to do with automation.

It will expose how much organisational debt companies have been carrying around for years.

AI performs beautifully inside a clear system

Give an AI system a well-defined objective, reliable information and sensible boundaries, and it can be astonishingly effective.

It can monitor far more signals than a person can. It can identify patterns quickly, recommend actions, automate repetitive work and operate continuously at a speed no committee could reasonably match.

The problem is that enterprises rarely present work in such a clean form.

Ask a business who owns pricing and the answer may be “commercial”, until marketing runs a promotion, finance intervenes on margin, the country team has local authority and the CEO has a view because somebody mentioned a competitor in the meeting.

Ask which inventory number is correct and three systems may offer different answers, each supported by a team that has spent several years learning when its particular number should be ignored.

Ask who owns the customer and the room becomes noticeably quieter.

Humans cope with this remarkably well. They learn the exceptions. They know who to call. They understand that the official process is not always the actual process. They know that a particular approval technically takes five days but can be obtained in twenty minutes if the right person is copied.

AI does not arrive with that institutional folklore.

It arrives asking what the rules are.

That can be a very revealing conversation.

The first thing automation does is force the organisation to explain itself

One of the most useful exercises in any serious automation programme is simply mapping how the work actually happens.

This is often more difficult than expected.

There is the process documented in the policy.

There is the process people describe in workshops.

There is the process the systems technically support.

And then there is what experienced employees actually do when something needs to get done.

These can be four entirely different things.

Over time, organisations accumulate workarounds because workarounds are rational responses to imperfect systems. Someone creates a spreadsheet because the core platform cannot provide the answer quickly enough. Another team adds an approval after one expensive mistake. Somebody builds a manual reconciliation because two systems disagree.

Eventually the workaround becomes the process.

Years later, nobody remembers why step seven exists, but removing it feels irresponsible because it has been there for a long time and therefore must presumably be important.

Then AI arrives.

An agent cannot reliably execute a workflow built on unwritten exceptions, personal relationships and assumptions nobody has formalised. The process has to become explicit.

That is when the real transformation begins.

Not when the AI automates the workflow.

When the business is forced to decide what the workflow should actually be.

Broken data is often an organisational symptom

Enterprise AI discussions almost inevitably become data discussions.

Correctly.

Poor data will constrain AI faster than most model limitations will. If the product information is inconsistent, customer identities fragmented or inventory unreliable, putting a better reasoning layer on top does not magically create truth.

But the phrase “data problem” can hide what is really happening.

Why are there three definitions of the same metric?

Why do different functions maintain separate customer records?

Why does nobody know which system is authoritative?

Why has the source field been wrong for three years?

These are rarely purely technical questions.

They are questions about ownership.

Somebody needs to define the metric. Somebody needs to own the source. Somebody needs to decide which system wins when two disagree. Somebody needs the authority and incentive to fix the problem rather than compensate for it downstream.

AI exposes this because a model cannot indefinitely rely on the kind of caveat humans use without thinking.

“Use that number for reporting, but not for planning.”

“This field is correct for online, except in two markets.”

“The customer record is accurate unless they purchased in-store before 2022.”

Large organisations are full of sentences like these.

They are a form of organisational debt.

Humans have simply become good at servicing the interest.

The bottleneck will often move somewhere else

This is one of the patterns I expect companies to discover as AI adoption becomes more serious.

AI makes one part of a process dramatically faster and the organisation celebrates the productivity gain.

Then everybody notices that the end-to-end cycle time barely moved.

An analysis that took three days now takes ten minutes, but approval still takes two weeks.

Customer service can identify the right resolution immediately, but the agent lacks authority to execute it.

A pricing system recommends an action in real time, but the organisation reviews prices every Thursday.

AI can identify an inventory imbalance continuously, but transferring stock still requires three functions to agree whose number will be affected.

The technology removed one bottleneck.

The operating model politely moved the queue somewhere else.

This is why isolated productivity measures can be misleading. A function may become considerably more efficient without the customer, supplier or P&L experiencing much difference.

The value appears only when the entire decision flow changes.

That requires looking beyond the task AI improved and asking what still sits between insight and action.

That space is where operating-model problems live.

Decision rights will become more important than use cases

A great deal of enterprise AI activity currently begins with use cases.

Where can we automate? Where can we generate content? Where can we assist employees? Where can we improve forecasting?

Those are sensible questions, particularly in the early stages.

But once AI starts moving from recommendation into action, another question becomes more important:

What is the machine actually allowed to decide?

That sounds like a governance issue.

It is also an operating-model issue.

Consider a customer-service agent. Perhaps AI can identify an unhappy high-value customer, understand the history and recommend compensation. Does it have authority to issue that compensation?

Up to what amount?

Does customer value matter? Does the reason for the failure matter? What if the policy says one thing but preserving the relationship suggests another?

The same questions appear in pricing, merchandising, marketing, supply chain and finance.

At what confidence level does the system act automatically?

When does a person intervene?

Who owns the outcome?

Who can override the recommendation?

And perhaps most importantly, who has the authority to redesign those decision rights in the first place?

Many organisations will discover that AI is technically capable of doing substantially more than the governance and operating model are prepared to permit.

The gap between those two things will become one of the defining constraints on enterprise AI value.

AI has little respect for organisational boundaries

Customers already have this problem with enterprises.

They experience one brand while the organisation experiences departments.

AI will experience something similar.

A useful customer agent does not care that product data belongs to merchandising, loyalty sits in marketing, transactions sit in commerce, service history sits somewhere else and identity is technically a technology problem.

It needs context.

The same is true for an AI system optimising demand. It may need inventory, pricing, promotions, customer behaviour, supplier constraints and margin.

The useful decision cuts horizontally across an organisation designed vertically.

This is why AI may place more pressure on functional silos than previous technology did.

A dashboard could tolerate fragmentation because a person eventually interpreted the different sources.

An agent expected to make or execute a decision needs the organisation to reconcile them.

That can feel like a technology requirement.

It is usually a leadership requirement wearing technical clothing.

Some work should disappear rather than become faster

There is another uncomfortable question AI will force companies to confront.

Why does some of this work exist at all?

Not every process deserves automation.

A surprising amount of enterprise activity is generated by complexity created elsewhere in the enterprise. Teams reconcile data because systems are fragmented. Managers compile reports because information cannot be accessed directly. People sit in meetings because decisions cannot happen at the point where the information originates.

AI can make all of those activities faster.

That does not mean it should.

Using an extremely capable AI agent to prepare a weekly report nobody needs is not transformation.

It is artisanal bureaucracy at machine speed.

The better question is whether the work can disappear.

This is harder because eliminating work can challenge existing roles, responsibilities and control mechanisms. Improving a process allows everyone to remain involved. Removing it entirely forces the organisation to ask who still needs to be.

That is where efficiency conversations become political.

And it is why some of the largest AI productivity gains will require leadership courage rather than model capability.

Managers will be forced to manage differently

Operating-model change will also reach management itself.

Many management layers exist partly because organisations historically needed people to aggregate information, coordinate activity and transmit decisions through the hierarchy.

AI reduces some of that friction.

If teams can access information directly, analysis can be produced on demand and routine exceptions can be resolved automatically, some coordination work becomes less valuable.

That does not make management obsolete.

It changes what good management is for.

Judgment becomes more important. So does setting direction, resolving ambiguity, developing people, making trade-offs and deciding when a machine-generated answer is technically correct but commercially wrong.

Managers whose primary value came from controlling information may find their position less comfortable.

Managers who improve decisions become more valuable.

This distinction has existed for a long time.

AI may simply make it harder to hide.

The most dangerous response is layering AI onto complexity

Large organisations have a natural tendency to add.

A new problem appears, so a new process is created.

A risk emerges, so an approval is added.

A technology arrives, so a centre of excellence appears.

Eventually there are committees coordinating the committees created to simplify coordination.

AI could easily become another layer.

Agents sitting on top of legacy platforms. Copilots added to existing workflows. New governance sitting beside old governance. Human approvals validating machine approvals because nobody quite trusts either one yet.

Some of that duplication will be necessary during transition.

But if it becomes permanent, AI may make the enterprise more complicated rather than less.

This is the danger of treating AI as an overlay.

The real opportunity is substitution.

Which step can disappear?

Which layer can disappear?

Which handoff no longer needs to exist?

Which decision can move closer to the person or system with the information?

If nothing is removed, the organisation should be cautious about claiming transformation.

Complexity has an extraordinary survival instinct.

AI should not become its newest habitat.

The companies that win will redesign around intelligence

The most valuable AI organisations will not necessarily be the ones with the most advanced models.

Most enterprises will eventually have access to capable models.

The more durable difference will come from what sits around them.

Clean enough data.

Clear enough ownership.

Fast enough decision rights.

Processes designed around action rather than historical precedent.

Governance proportional to risk.

Teams willing to remove old work rather than merely automate it.

And leaders prepared to change the operating model when the technology reveals that the old one no longer makes sense.

That last part matters.

It is easy to approve an AI programme.

It is harder to tell two functions that one of them no longer owns the decision.

It is easy to celebrate automation.

It is harder to remove the approval layer automation made unnecessary.

It is easy to ask employees to adopt AI.

It is harder to redesign jobs, targets and incentives around what AI makes possible.

The technology is often the least political part of transformation.

The value sits in everything it forces the organisation to confront afterwards.

AI does not create organisational clarity

There is a tendency to anthropomorphise AI as though sufficiently intelligent systems will somehow navigate the enterprise around us.

They will figure out the data.

They will understand the process.

They will optimise the workflow.

Perhaps eventually they will become remarkably good at compensating for organisational mess.

That still does not make the mess sensible.

A sophisticated system can learn that three different customer IDs refer to the same person. The better question is why the company needs three customer IDs.

AI can learn that an approval is usually overridden. The better question is whether the approval belongs in the process.

It can reconcile conflicting data.

The better question is why the enterprise allows conflicting versions of the same truth to persist.

The more capable AI becomes, the easier it may be to use intelligence to mask structural problems rather than solve them.

That would be a mistake.

Efficiency applied to dysfunction produces efficient dysfunction.

The exposure is the opportunity

This is why I think leaders should view AI exposing operating-model problems as a benefit rather than an inconvenience.

If an agent cannot execute because ownership is unclear, clarify ownership.

If AI cannot recommend reliably because data definitions conflict, resolve the definitions.

If decisions remain slow despite instant analysis, examine the decision rights.

If productivity improves but the end-to-end process does not, find the next bottleneck.

If employees keep working around the system, understand what they know that the process does not.

AI gives organisations a very powerful diagnostic tool.

It shows where intelligence is being constrained by structure.

That is useful information.

The temptation will be to fix the AI.

Often, the organisation needs fixing instead.

The operating model is now part of the AI stack

We tend to think of the AI stack in technical layers.

Models. Data. Infrastructure. Applications. Security.

I would add another.

The operating model.

Because that is the layer that determines whether intelligence becomes action.

A brilliant recommendation arriving inside an organisation that cannot decide is just a faster form of analysis.

An agent with perfect context but no authority is still waiting.

A model trained on excellent data cannot compensate indefinitely for incentives pointing teams in opposite directions.

AI can make organisations substantially more capable.

But it cannot make them coherent on their behalf.

That remains a leadership responsibility.

And perhaps that is the real opportunity in this next wave.

AI will undoubtedly automate work, improve productivity and change how enterprises compete.

But before it does all of that, it may perform an even more valuable service.

It will show leaders exactly where the organisation has been getting in its own way.

The companies that act on what it reveals will transform.

The ones that do not will simply automate the evidence.

Click and collect. Ship from store. Endless aisle. Mobile point of sale. Clienteling. Single customer views. Distributed inventory. Apps that know where you are, what you bought and, on a good day, whether the product is actually available.

The technology has moved considerably.

The customer has moved even further.

What has often moved least is the organisation.

Customers do not experience “digital”, “stores”, “CRM”, “customer care” and “supply chain” as separate functions. They experience a brand. The fact that the business behind that brand has divided itself into channels, departments, budgets and reporting lines is largely irrelevant to them.

Until those divisions get in the way.

That is the real omnichannel problem now.

It is no longer primarily about connecting channels.

It is about removing the organisational seams between them.

The customer journey is horizontal. The enterprise is not.

This is the structural tension underneath most omnichannel programmes.

A customer might discover a product on social, research it on the website, visit a store to see it physically, ask an associate for advice, purchase later through the app, collect from another location and contact customer service when something goes wrong.

From the customer’s point of view, that is one journey.

Inside the business, it can involve six teams, four systems, multiple budgets and several different definitions of ownership.

Digital sees traffic and conversion. Stores see footfall and sales. Marketing sees acquisition. CRM sees engagement. Customer care sees contacts. Supply chain sees fulfilment.

Everyone sees their part of the elephant.

The customer is sitting on top of it.

This is why omnichannel transformation so often reaches a frustrating plateau. The systems become connected, but the business still thinks functionally.

You can build a single customer view while maintaining separate customer incentives. You can create shared inventory visibility while retaining local ownership rules. You can enable a store associate to sell an item from another location while still rewarding them only for what leaves their own till.

Technically omnichannel.

Behaviourally not.

We connected the technology before the incentives

This is where the problem becomes much less glamorous.

Connecting systems is hard, but it is at least a familiar kind of hard. There are architectures, APIs, roadmaps, vendors and programme plans.

Changing incentives is political.

A store manager has a target. The eCommerce team has a target. Marketing has a target. Commercial has a margin target. Customer care has service levels. Supply chain has cost and availability measures.

Each of those measures may be individually rational.

Together, they can create irrational customer outcomes.

Consider a simple example.

A customer walks into a store looking for a product that is unavailable locally but available online. The associate can place the order for them.

Great omnichannel experience.

But what happens next?

Does the store receive credit for the sale? Does the associate? Does online lose anything? Who funds fulfilment? What if the product is returned to the store? Does the return damage the store’s performance even though the original transaction was recorded elsewhere?

Those questions sound administrative.

They are not.

They determine whether the associate enthusiastically helps the customer or quietly explains that they might have better luck checking the website themselves.

The technology may make the right behaviour possible.

The incentive decides whether it happens.

Attribution is not the same thing as value creation

Retail organisations spend a remarkable amount of time debating where revenue belongs.

There are good reasons for this. Leaders need accountability. Channels need P&Ls. Stores need targets. Investments need to be evaluated.

The problem begins when the accounting logic starts determining customer behaviour.

Imagine a customer researches extensively online, visits a store, speaks with an associate and later buys through the app.

Which channel created the sale?

The technically correct answer depends on the attribution model.

The commercially useful answer may be: all of them.

That does not mean abandoning channel economics. It means recognising that the place where revenue is booked is not always the place where value was created.

This becomes even more obvious in categories where physical and digital reinforce one another. A store may drive confidence, trial and service. Digital may provide range, convenience and continuity. Customer care may preserve the relationship after a service failure.

Trying to reduce all of that to a single channel owner creates precision at the expense of reality.

There is a point where measurement stops explaining the business and starts distorting it.

Good omnichannel leadership knows the difference.

Stores are becoming part of the digital operating system

One reason I think the old channel model is increasingly unhelpful is that the role of the store itself has changed.

A store is still a place where products are sold.

It can also be a fulfilment node, a returns centre, a service environment, a content location, a customer-acquisition channel, an inventory pool and a human interface into the brand.

That makes the store part of the digital system whether the organisation chart acknowledges it or not.

A clienteling associate using customer history, product availability and remote inventory is participating in digital commerce. A store fulfilling an online order is participating in eCommerce. A customer returning an online purchase physically is not “switching channels”. They are using the business in the way that is most convenient.

The distinction feels increasingly artificial.

This is why I think the next phase of retail transformation will involve much more than giving stores better technology.

It will involve redefining what the store is economically responsible for.

If the store helps create demand, fulfil demand and preserve customer relationships across channels, then measuring it only on transactions passing through a particular till misses part of its value.

That does not make store economics easier.

It makes them more honest.

The same customer should not become a stranger when the channel changes

This is probably one of the simplest tests of whether an organisation is genuinely omnichannel.

Does the business remember the customer when the context changes?

If a customer has spent years with the brand online, does the store associate know enough to recognise the relationship appropriately?

If a customer had a service problem yesterday, does marketing know not to send a cheerful promotional message this morning?

If the customer returns an online order in-store, can the colleague see what happened without asking them to reconstruct the transaction from screenshots and emails?

If loyalty status matters digitally, does it matter physically?

These are not futuristic experiences.

Customers increasingly consider them basic competence.

What makes them difficult is that the relevant information often lives across multiple teams and systems. The issue is rarely the absence of data. It is whether that data can move across organisational boundaries quickly enough to be useful.

This is where “single customer view” can become a misleading milestone.

A single view is valuable.

A single view that nobody can act on is simply a more elegant way to observe fragmentation.

The real test is whether customer context changes the next decision.

AI will make organisational fragmentation more visible

AI adds another layer of urgency.

Most serious AI use cases in commerce depend on connected context.

A customer-facing assistant needs product information, order history, inventory, loyalty, service policies and perhaps personalised recommendations. A store colleague using an AI-enabled clienteling tool needs many of the same signals. Customer care needs transaction history and service context. Merchandising systems need demand and inventory data.

AI does not remove those dependencies.

It amplifies them.

A fragmented organisation can still operate because experienced people learn the workarounds. They know which system is trustworthy, which team to call and who can unblock something outside the formal process.

AI is less tolerant of that kind of institutional folklore.

If three systems disagree on inventory, the model does not magically know which political compromise to apply. If ownership is unclear, an agent cannot infer the unwritten rules people have spent years navigating.

That means AI may expose omnichannel debt faster than previous technologies did.

The organisation that wants an intelligent customer experience first needs a coherent one.

Omnichannel is really a decision-rights problem

Once the technology is connected, many omnichannel issues reduce to a set of surprisingly basic questions.

Who owns the customer?

Who owns inventory?

Who can promise what?

Who can resolve a service issue?

Who gets credit?

Which metric wins when customer experience and channel economics conflict?

Those are decision-rights questions.

And organisations often avoid answering them because the current ambiguity allows everybody to retain some ownership.

That works until the customer needs the business to act as one company.

An associate cannot offer the best solution if they lack authority. A customer-care agent cannot resolve the issue if policy requires another function. An eCommerce team cannot optimise across the network if local inventory incentives work against it.

Technology can surface the answer.

Someone still has to be allowed to make the decision.

This is why I increasingly view omnichannel as an operating-model discipline rather than a digital capability.

The hard part is no longer making the systems talk.

It is making the organisation behave as though they do.

Shared customer economics matter more than shared terminology

Many companies have adopted omnichannel language without changing much of the underlying economics.

The word appears in strategy decks. Teams are renamed. Roadmaps become “connected”. Customer journeys are drawn across channels.

Useful, but insufficient.

The more meaningful shift is to introduce measures that cut across channels.

What is the lifetime value of an omnichannel customer?

How does store engagement affect online frequency?

What does clienteling do to retention?

What happens to conversion when customers can access inventory across the network?

How does easy cross-channel returns behaviour affect loyalty, margin and future spend?

Which customers become more valuable because they use multiple touchpoints, and which simply become more expensive to serve?

These questions create a different conversation because they force the organisation to look at the relationship rather than the transaction.

That is where the business begins behaving more like the customer sees it.

There is still a place for channel accountability

None of this means organisations should dissolve every channel P&L and hope customer centricity takes care of the numbers.

That would be convenient.

It would also be terrible management.

Stores still need to be productive. Digital still needs to convert. Fulfilment costs still matter. Marketplace economics still matter. Leaders need to know which investments are producing returns.

The answer is not less accountability.

It is better accountability.

Channel metrics tell you whether individual parts of the machine are working.

Customer metrics tell you whether the machine is working as a whole.

Both matter.

The problem is when one replaces the other.

If a channel is performing brilliantly while total customer economics deteriorate, the organisation should be suspicious of the brilliance.

The best omnichannel experience may eventually be invisible

For years, retailers marketed omnichannel capabilities because the capabilities themselves were novel.

Buy online, pick up in store.

Return anywhere.

Check availability.

Order from another location.

Eventually, these stop being differentiators.

They become expectations.

The customer no longer admires the complexity required to provide them. They simply notice when they do not work.

That is probably where omnichannel should end up.

Invisible.

Inventory is accurate. Customer context travels. The brand remembers. The associate can help. Products move through the network. Returns happen without drama. Service understands what came before.

Nobody congratulates the retailer for integrating five systems behind the scenes.

Nor should they.

The sophistication is hidden inside the simplicity.

That is what good operating models do.

The next omnichannel transformation will be organisational

Retailers have already spent years building the connective technology.

There is more to do, of course. Legacy systems remain. Inventory accuracy is uneven. Customer identities still fragment. Plenty of architecture needs work.

But the next frontier is less comfortable.

Targets.

Ownership.

Decision rights.

P&Ls.

Incentives.

The boundaries between teams.

These are harder to modernise because they involve power rather than platforms.

Yet that is where the remaining friction increasingly sits.

Customers have already reorganised commerce around themselves. They move across channels without asking permission and expect the business to follow.

The enterprise still has a choice.

It can continue organising around the way retail used to be managed.

Or it can begin organising around the way customers actually shop.

The customer stopped shopping in channels years ago.

The question is how long the organisation wants to keep doing so.

Media costs rise. Conversion softens. The easy audiences have already been reached. Promotions work, but only if they become more generous. Finance begins asking why marketing is spending more to produce roughly the same result. Marketing responds by asking for better creative, better targeting, better attribution and, ideally, a little more budget while those things improve.

The instinct is understandable. If customer acquisition cost is rising, fix acquisition.

But that is only part of the problem.

I have increasingly come to think that many businesses diagnose CAC too narrowly because acquisition is one of the easiest parts of growth to isolate and measure. The spend sits in one place. The traffic appears somewhere else. The conversion can be counted. The result fits comfortably into a dashboard.

What happens after acquisition is much less tidy.

Retention sits with CRM. Returns sit with operations. Service has its own economics. Merchandising influences repeat purchase. Pricing changes margin. Product experience affects frequency. Loyalty may sit somewhere between marketing and customer experience. The P&L, unfortunately, receives all of them.

So when CAC rises, the more useful question is not simply how to acquire the next customer more cheaply.

It is whether the customer is valuable enough once acquired.

A cheap customer can be very expensive

Digital growth created a generation of businesses that became exceptionally good at acquiring customers.

There is nothing wrong with that. Acquisition matters. Every business needs new customers eventually, and efficient demand generation remains one of the most powerful commercial capabilities a company can build.

The mistake is assuming that a low acquisition cost automatically means good growth.

Imagine two customers.

The first costs $20 to acquire, buys once using a 25 percent discount, returns part of the order, generates a service contact and never comes back.

The second costs $70 to acquire, buys full price, shops again four months later, gradually expands into other categories and stays with the brand for several years.

Which acquisition was expensive?

The answer is obvious once you look at the whole relationship.

Yet many organisations still manage growth through metrics that make the first customer look more attractive at the moment the decision is being made.

That is the danger of optimising a component of the economics instead of the economics themselves.

CAC is useful. It is simply not enough.

Acquisition became too detached from the customer P&L

One reason this happens is organisational.

Marketing often owns acquisition targets. CRM owns retention. Commercial teams own margin. Customer care owns service cost. Operations owns fulfilment and returns.

Each function can therefore perform well while the customer relationship performs badly.

Marketing hits its acquisition number by leaning into a heavily discounted campaign. Conversion looks strong. The new-customer target is achieved. Everyone is pleased.

Three months later, the cohort has poor repeat behaviour and low full-price penetration.

Those outcomes may sit in somebody else’s dashboard by then.

This is not a criticism of marketing teams. Most functions respond rationally to the metrics they are given. If you pay people to maximise new customers, they will find new customers. If you reward revenue without enough regard for margin, they will find revenue too.

Organisations tend to get remarkably good at producing exactly what they asked for.

The problem is whether they asked for the right thing.

If the objective is profitable customer growth, acquisition, retention, frequency, margin and cost-to-serve cannot remain entirely separate conversations.

The customer experiences one economic relationship with the business.

The organisation should probably do the same.

Rising CAC is often telling you something else

Acquisition costs can rise for straightforward reasons. Competition intensifies. Platforms become more expensive. Privacy changes reduce targeting efficiency. A category becomes crowded. The brand moves beyond its early enthusiasts into customers who are naturally harder to persuade.

But rising CAC can also be a symptom rather than the disease.

Perhaps the proposition is becoming less distinctive.

Perhaps the brand is relying too heavily on performance marketing because organic demand is weak.

Perhaps repeat purchase is too low, meaning the business constantly has to replace customers it should have retained.

Perhaps promotions have trained customers to wait.

Perhaps the experience after purchase is quietly destroying the value created before it.

If a business has to keep buying the same customer back, the problem is not necessarily acquisition efficiency.

It may be that the relationship is not strong enough to survive without paid intervention.

That distinction matters because a better Facebook campaign will not fix it.

Neither will another attribution model.

The obsession with the top of the funnel can hide a leaking bottom

Growth teams naturally spend enormous energy on acquisition because that is where the activity is visible.

Campaigns can be launched. Audiences can be tested. Creative can be changed. Bids can be adjusted. There is always another lever to pull.

Retention is less theatrical.

A customer who quietly comes back six months later does not create the same sense of operational excitement as a successful launch campaign. A reduction in unnecessary discounting rarely gets the same internal applause as a surge in traffic. Improving service recovery is unlikely to get its own countdown clock.

Yet those things can materially change the economics of growth.

If retention improves, the business does not need to replace as many customers.

If frequency improves, the acquisition cost is spread across more transactions.

If full-price mix improves, the same customer generates more contribution.

If returns fall, revenue becomes more valuable.

If service improves, the relationship survives moments that would otherwise create churn.

None of these metrics makes CAC itself cheaper.

They make the CAC more affordable.

That is a much more powerful lever.

The next marketing dollar is a better question than the total marketing budget

One question I find useful is deceptively simple:

Where should the next marketing dollar go?

Not where did the last million go. Not what percentage belongs to paid social versus search. Just the next unit of investment.

Should it acquire someone new?

Bring back someone who has lapsed?

Prevent a valuable customer from leaving?

Increase frequency among an existing cohort?

Improve conversion?

Improve product discovery?

Reduce a service failure that causes churn?

The answer will not always sit inside the marketing department.

That is exactly why the question is useful.

It forces the business to compare opportunities across the customer lifecycle rather than assuming that growth begins at acquisition and everything afterwards is retention management.

Sometimes the highest-return growth investment is media.

Sometimes it is inventory availability.

Sometimes it is a better returns experience.

Sometimes it is improving the product itself.

Growth is an outcome of the business system, not a media channel.

That can be inconvenient in organisations where everyone would prefer the growth problem to belong to one team.

Not every customer deserves the same investment

There is also a more uncomfortable truth that many customer strategies avoid.

Not every customer is equally valuable.

That does not mean treating people badly. It means recognising that customer economics vary.

Some customers buy full price and remain for years. Others appear only during promotion. Some explore multiple categories. Some have unusually high returns. Some require expensive servicing. Some introduce new customers through advocacy. Some are highly engaged but economically marginal.

The historical solution was segmentation.

AI and richer customer data should allow the business to become much more precise about where incremental investment creates value.

Who actually needs an incentive?

Who would have purchased without it?

Who is likely to lapse?

Which customer is worth winning back?

Which high-spend customer is less profitable than they appear?

Who is at the beginning of a potentially valuable relationship rather than simply making an isolated transaction?

Those are commercial allocation decisions.

And they are much more interesting than asking which demographic segment should receive 10 percent off this weekend.

Personalisation should eventually include economics

This is where the next generation of personalisation becomes more useful.

Retail has spent years personalising content. Products are reordered. Recommendations change. Emails differ by segment. Homepages respond to behaviour.

Useful work.

But the bigger opportunity is to personalise investment.

Two customers can see the same product and require entirely different commercial treatment.

One may need a small incentive to convert. Another would buy anyway.

One may be worth paying substantially more to reacquire because the future relationship is strong. Another may be unprofitable even if the reactivation itself looks efficient.

A customer who has just experienced a service failure may need acknowledgement rather than another promotion. A highly loyal customer may value access or convenience more than discount.

This is where customer intelligence starts moving out of communications and into business decisions.

Pricing.

Service.

Loyalty.

Media allocation.

Benefits.

Retention intervention.

That is a much more consequential form of personalisation because it changes the economics, not just the message.

LTV can become just as misleading as CAC

There is a caveat here.

Once businesses become more sophisticated about customer economics, the instinct is often to replace CAC obsession with LTV obsession.

That can create a different kind of false confidence.

Lifetime value is not a fact.

It is an estimate.

It depends on assumptions about retention, frequency, margin, future behaviour, discounting and cost-to-serve. A model can tell you that a customer is worth $800 over three years while quietly assuming the customer remains loyal, margins remain stable and your organisation does nothing particularly foolish in the meantime.

That does not make LTV useless.

It makes it something leaders should understand rather than simply admire.

The ratio between CAC and LTV can be helpful, but only if both sides are economically honest. Gross revenue is not customer value. Repeat purchase generated entirely by discounts is not necessarily healthy retention. A customer who returns half of every order should probably not be valued using the same logic as someone who does not.

The model should reflect the business you actually operate.

Not the business case you would like to present.

The best acquisition strategy may be building something customers return to

This sounds almost embarrassingly old-fashioned, but it matters.

One way to reduce dependence on increasingly expensive acquisition is to build a proposition customers actively want to revisit.

Good product.

Good service.

Reliable fulfilment.

Fair value.

Useful loyalty.

Distinctive experience.

Trust.

None of these sounds as sophisticated as AI-driven media optimisation, but they change acquisition economics indirectly because they create repeat behaviour, advocacy and organic demand.

A strong brand does something similar.

If customers already know and trust the business, the organisation does not have to purchase quite as much persuasion every time it wants to sell something.

That is one reason CAC should not be treated solely as a media metric.

It is influenced by the strength of the entire proposition.

Marketing can make a weak business grow for a while.

It cannot make the weakness disappear.

Eventually, acquisition cost begins reflecting the amount of persuasion required.

AI will make acquisition more efficient. Everyone else will have AI too.

There is an understandable belief that AI will solve some of the current acquisition challenge.

It will certainly help.

Creative can be produced and tested faster. Media allocation can become more intelligent. Customer segments can become more dynamic. Predictive models can improve bidding, targeting and personalisation.

The complication is that these tools will not belong to one company.

If every competitor can generate better creative, optimise campaigns faster and identify audiences more intelligently, those capabilities gradually become the new baseline.

The advantage moves elsewhere.

Into proprietary customer knowledge.

Into proposition.

Into brand.

Into retention.

Into operational execution.

Into knowing which customer to acquire rather than simply how to acquire them.

AI can improve the efficiency of the acquisition machine.

It cannot decide whether the economics behind the machine make sense unless the organisation has defined them properly.

Once again, the quality of the technology eventually runs into the quality of the operating model.

Growth needs one version of customer economics

This is where I think many organisations will eventually have to change.

Customer acquisition, CRM, loyalty, service and commercial planning cannot remain entirely separate if the objective is customer lifetime economics.

They do not necessarily need to sit in one department.

But they do need shared measures.

What is a new customer worth after margin?

How does value develop after 30, 90 or 365 days?

Which acquisition sources produce customers who actually stay?

Which cohorts become promotion-dependent?

Where does service failure affect retention?

Which customer segments generate profit rather than just revenue?

These questions create a different conversation.

Marketing starts caring more about what happens after conversion. Commercial teams care more about who is buying rather than only what is selling. CRM becomes connected to acquisition strategy. Service becomes visible as a commercial lever.

The organisation begins optimising the relationship rather than the transaction.

That is the operating shift rising CAC should provoke.

The goal is not cheaper customers

There will always be pressure to improve acquisition efficiency.

There should be.

Businesses should negotiate media costs, improve conversion, test creative, optimise targeting and stop wasting money. Poor acquisition discipline should not be excused by invoking lifetime value.

But the obsession with getting CAC down can become strategically limiting.

A company can halve acquisition cost by acquiring worse customers.

It can increase conversion by discounting more heavily.

It can improve short-term ROAS while reducing long-term margin.

The metrics improve.

The business does not.

The better objective is not the cheapest possible customer.

It is the most attractive relationship the company can build at an economically sensible cost.

That requires a wider view of growth.

Acquisition matters.

Retention matters.

Margin matters.

Frequency matters.

Experience matters.

The customer does not know which department owns any of them.

And the P&L does not particularly care either.

CAC may continue rising.

The answer is not always to spend less acquiring the customer.

Sometimes it is to create a customer worth spending more to acquire.

That is usually enough to get the flywheel moving. More supply attracts more demand, more demand attracts more supply, and the marketplace grows because it has made finding and completing a transaction materially easier than whatever existed before.

But the most interesting marketplaces rarely stop there.

Once enough buyers, sellers, transactions and behavioural data accumulate, the business begins noticing something else. The transaction is only one small part of the problem its ecosystem is trying to solve.

Merchants need payments. They need logistics. They need advertising. They need financing, analytics, customer-service tools, inventory visibility and software to help them run the business around the marketplace itself.

Gradually, the platform moves from facilitating commerce to supporting the machinery that makes commerce possible.

That is when the business becomes much harder to displace.

The marketplace endgame, in other words, is not necessarily a bigger marketplace.

It is infrastructure.

The transaction is where the relationship starts

Early marketplace strategy is dominated by liquidity.

Can I find enough sellers? Can I attract enough buyers? Can I match them efficiently enough that both sides have a reason to return?

That is the right problem to solve first. Without liquidity, everything else is decoration.

But once a platform achieves scale, the transaction begins revealing adjacent problems. A seller who needs customers also needs to collect money, move goods, advertise products, understand demand and occasionally borrow against the future they are trying to build.

Each of those needs creates another potential business.

Payments are the obvious example. If money already moves through the platform, helping settle that transaction more efficiently makes sense. Logistics follows for similar reasons. Advertising emerges because sellers are already paying to acquire demand elsewhere, and the marketplace has something unusually valuable to sell them: access to customers who are already in shopping mode.

Then come data products, financing, fulfilment, merchant software and, increasingly, AI.

None of these adjacencies is random. They accumulate around the friction already visible inside the core transaction.

That is the important part.

The strongest platform businesses do not keep inventing unrelated revenue streams. They keep solving problems their ecosystem is already paying somebody else to solve.

This is why the economics can compound

A pure matching business can be powerful, but it is relatively easy to understand.

A buyer arrives. A seller transacts. The platform earns a commission.

Infrastructure changes the relationship because the marketplace can now earn value from multiple moments around the same customer or merchant relationship.

A seller may pay for fulfilment, payments, advertising and software before the underlying transaction fee is even considered. A buyer may use the marketplace because the logistics promise, returns process and payment experience feel more trustworthy than dealing with an unknown merchant directly.

The pieces begin reinforcing one another.

Better logistics improves customer confidence. Better confidence improves conversion. More transactions generate better data. Better data improves advertising and forecasting. More advertising attracts additional supplier investment.

The flywheel stops being only about buyers and sellers.

It becomes a flywheel of services around commerce itself.

This is why some marketplace businesses eventually begin to look economically less like retailers and more like infrastructure companies with a retail surface attached.

The distinction matters because infrastructure tends to be sticky.

A merchant can list products somewhere else relatively easily. Moving payments, fulfilment, advertising, data, reviews, workflow and customer acquisition away from a deeply embedded platform is a very different decision.

The danger is that usefulness can turn into dependence

This is also where marketplace power becomes more complicated.

The more infrastructure a platform provides, the more valuable it can become to sellers. The same integration that creates convenience also creates dependency.

There is a fine line between building an ecosystem merchants do not want to leave and building one they feel they cannot leave.

Those are not the same thing.

Platforms occasionally forget this because the economics of dependence can look very attractive for quite a long time. Fees rise. Advertising becomes harder to avoid. Paid placement starts influencing visibility. Sellers are encouraged to adopt more services inside the ecosystem because each one improves performance somewhere else.

From the platform’s perspective, this looks like monetisation depth.

From the merchant’s perspective, it can eventually feel like rent.

The long-term strength of a marketplace therefore depends on whether the infrastructure keeps creating enough value to justify the economics attached to it.

That is a trust question as much as a pricing question.

Ecosystems are strongest when participants believe the platform is making them more successful, not simply becoming more efficient at extracting value from their success.

There is usually a point where those two things begin to look uncomfortably similar.

Good marketplace leadership knows where that point is.

Advertising is really a signal of the larger shift

Retail media and marketplace advertising are useful examples of how this evolution happens.

At first, advertising looks like a straightforward adjacency. Sellers already want more visibility. The marketplace already has traffic. Allow merchants to pay for prominence and create a new revenue stream.

Nothing particularly mysterious there.

But advertising becomes much more powerful when it is connected to actual transaction data. The platform can see which products convert, which customer groups respond, which categories are growing and where promotions create genuine incremental demand.

Suddenly the product is no longer only media.

It is intelligence.

The same pattern appears elsewhere. Payments generate transaction data. Logistics generates operational data. Financing reveals merchant health. Software reveals workflow. Together, these signals allow the marketplace to understand the commercial system around the transaction far more deeply than a pure intermediary ever could.

That understanding is where another layer of value appears.

The platform is no longer simply helping somebody sell.

It is helping them decide.

Infrastructure changes the nature of competitive advantage

This matters because marketplace competition is often discussed in terms of assortment, pricing and traffic.

All important.

But once infrastructure matures, the competitive advantage begins moving underneath the interface.

How good is the fulfilment network?

How reliable are payments?

How quickly can sellers launch?

How well does the platform understand demand?

How much friction has been removed from operating the business?

How useful are the tools once the transaction is over?

These are not necessarily the things customers notice directly.

That is usually a sign infrastructure is working.

The customer sees a product arrive quickly and reliably. They do not see the inventory allocation, payment settlement, fraud controls, routing logic and fulfilment orchestration underneath it.

The merchant sees sales increase. They may not spend much time thinking about the hundreds of services and data exchanges making that possible.

Invisible capability is still capability.

Often, it is the most valuable kind.

AI will make the plumbing more important, not less

The current excitement around AI sometimes creates the impression that conversational interfaces will make much of traditional commerce infrastructure obsolete.

I think the opposite is more likely.

An AI agent can recommend a product, compare alternatives or even initiate a purchase. But it still needs reliable product information, inventory, pricing, identity, payments, fulfilment and returns.

Someone still has to know whether the product exists.

Someone still has to move it.

Someone still has to collect the money.

Someone still has to resolve the problem when the beautifully intelligent shopping agent orders the wrong size.

The interface may become dramatically more sophisticated while the underlying requirements remain gloriously operational.

In fact, agentic commerce makes infrastructure more important because machines need dependable systems to act against. An AI agent can tolerate less ambiguity than a human customer who is willing to ring a call centre when something goes wrong.

If pricing is inconsistent, inventory unreliable or policies unclear, the agent does not experience “brand charm”.

It experiences bad data.

This is where marketplace infrastructure becomes strategically interesting again. Platforms that already connect product data, payments, logistics, identity and merchant services are well positioned to serve an environment where commerce is increasingly initiated by software rather than only by humans.

The AI layer may feel new.

The advantage underneath it will often belong to whoever built the plumbing first.

Retailers should pay attention to what marketplaces have learned

This is not only a lesson for marketplace businesses.

Large retailers are accumulating many of the same ingredients.

They have customer relationships, supplier ecosystems, payments, logistics, loyalty, media inventory, data, physical locations and technology. Historically, most of these capabilities were built to support the retail business itself.

But once they become sufficiently capable, a strategic question appears.

Does this capability remain an internal cost centre, or could it become a service others value?

A retailer with sophisticated fulfilment can move only its own inventory, or potentially support others. A strong customer-data capability can power only internal marketing, or become the foundation for privacy-safe insight and media services. A robust payments capability can remain invisible plumbing or become part of a broader ecosystem.

Not every retailer should do this.

That caveat matters.

There is a peculiar strategic habit of seeing somebody else build a high-margin adjacency and immediately deciding the organisation must have one too. Plenty of retailers would create more shareholder value by becoming better retailers rather than aspiring to become technology platforms.

Infrastructure businesses require scale, reliability, governance and an operating discipline that is very different from simply exposing an internal capability to external customers.

The question is not “Can we monetise this?”

It is “Are we genuinely good enough at this that somebody else would choose to depend on us?”

Those are very different thresholds.

The real platform advantage is accumulated trust

Infrastructure sounds technical, but its deepest advantage is often trust.

Buyers trust that products will arrive.

Sellers trust that payments will settle.

Brands trust that measurement is credible.

Merchants trust that the platform can create demand.

Developers trust the interfaces.

Partners trust the data.

Once enough of that trust accumulates, the marketplace becomes more than somewhere a transaction takes place.

It becomes part of how participants run their business.

That is a far more powerful position, but also a more demanding one.

An outage is no longer an inconvenience. It can interrupt somebody else’s revenue.

A policy change is no longer a product decision. It can alter another company’s economics.

A ranking algorithm is no longer simply an experience feature. It can determine which businesses are visible.

Infrastructure creates power.

It also creates responsibility.

The platforms that forget the second part eventually put the first at risk.

The endgame becomes difficult to see from the outside

This is why the most mature marketplaces can be misunderstood if we look only at the consumer-facing experience.

The homepage may still appear to be about products.

The economics underneath it may increasingly come from everything surrounding those products.

Payments. Advertising. Fulfilment. Software. Data. Financing. Services.

The visible marketplace becomes only the surface through which the deeper infrastructure is accessed.

This pattern is not unique to commerce. Many of the most valuable technology businesses eventually move downward into infrastructure because infrastructure sits closer to recurring economic activity.

What makes marketplaces particularly interesting is that they can begin with the customer relationship and then build downward.

That is a formidable starting position.

The marketplace may eventually disappear into the transaction

There is a certain irony in all of this.

Marketplaces were originally successful because they created destinations. Buyers went somewhere specifically to find sellers.

The infrastructure endgame may make the destination less important.

Commerce becomes embedded in other experiences. Transactions are initiated through social platforms, AI assistants, partner ecosystems and devices. Customers may increasingly consume marketplace capabilities without consciously thinking of themselves as “visiting a marketplace”.

What survives underneath is the infrastructure.

The identity.

The trust layer.

The product data.

The payments.

The fulfilment.

The seller network.

The intelligence.

That may ultimately be the strongest position of all.

The marketplace wins not because everybody visits it.

It wins because more and more commerce becomes difficult to imagine without the systems it built underneath.

An impression becomes a click. A click becomes a visit. A visit becomes an add-to-cart. A purchase appears at the other end, and an attribution model attempts to explain what happened in between.

None of this has ever been as precise as the dashboards suggested, but it was visible enough to create confidence. We could see the search query, the traffic source, the campaign, the landing page and often the transaction.

AI is beginning to disturb that arrangement.

A customer can now describe what they want to an AI assistant, ask it to compare products, summarise reviews, identify trade-offs, exclude unsuitable options and recommend a shortlist. Much of what used to happen across search results, retailer websites and comparison pages can happen inside a conversation the brand never sees.

The customer still discovers. The retailer simply loses visibility into how.

And that creates a much bigger problem than another attribution gap. It challenges the measurement architecture marketing has spent twenty years building around observable journeys.

We are moving from search behaviour to interpreted intent

Traditional search gave marketers something enormously useful: a reasonably visible expression of intent.

The query itself told us something. “Running shoes” was broad. “Best lightweight running shoes for hot weather” was considerably more useful. The customer clicked results, moved between pages, looked at products and left a trail that could be observed, optimised and bought against.

AI changes the interaction.

A customer might now say: “I’m travelling for three weeks, I’ll be walking a lot, I want something light enough for the heat, smart enough to wear casually, and I don’t want the pair everyone else has.”

That is a remarkably rich commercial signal. It contains context, preferences, constraints and probably more genuine intent than half the behavioural data sitting inside a conventional marketing funnel.

The irony is that the retailer may receive none of it.

The assistant interprets the request, consults whatever information it can access, narrows the choices and perhaps recommends three products. By the time the customer reaches the brand, much of the discovery and consideration has already happened somewhere else.

The website visit may look direct. The decision was not.

This is not the first time measurement has been incomplete

Marketing should resist the temptation to behave as though AI has suddenly destroyed a previously perfect system.

It has not.

Attribution has always involved compromise. Customers see advertising they do not click. They walk past stores. They hear recommendations from friends. They watch creators, read reviews, receive emails, compare products on marketplaces and occasionally make a purchase for reasons that do not feature anywhere in the media plan.

Digital simply gave us more observable behaviour than we had before, and the industry understandably became attached to it.

The danger came when visibility started being confused with causality.

A channel that can be measured precisely is not automatically more valuable than one that cannot. A click appearing immediately before a transaction does not necessarily deserve all the credit for creating it.

Most experienced marketers know this. Most organisations still find it easier to budget around what the dashboard can see.

AI-mediated discovery will make that tension harder to ignore.

The last click may become even less informative

Imagine a customer spends fifteen minutes discussing options with an AI assistant.

The assistant explains the category, compares brands, eliminates unsuitable choices and recommends one product based on price, availability, reviews and the customer’s stated preferences. The customer then clicks directly to the product page and purchases.

A conventional analytics system may record that as direct traffic, referral traffic or perhaps a new AI-source category if the referral is visible.

But what exactly should marketing learn from that?

The last click tells us where the customer arrived from. It says very little about why the product made the shortlist.

The useful signals may have existed much earlier: product attributes, customer reviews, editorial coverage, brand reputation, pricing, delivery promises, returns policy, availability or structured product information the AI was able to understand.

Suddenly, things traditionally treated as separate disciplines begin contributing to acquisition.

Product data affects discoverability.

Customer service affects reputation.

Inventory quality affects recommendation.

Returns policy affects confidence.

Reviews become training material for somebody else’s interface.

The boundary around “marketing” becomes much harder to maintain. That is probably healthy.

Product data is becoming part of the brand

For years, product information has often sat somewhere between merchandising, eCommerce, technology and content operations.

It mattered, obviously. Good titles, descriptions, imagery and attributes improved search, navigation and conversion.

But in an AI-mediated world, product information begins playing a different role. It becomes part of how machines understand the brand.

If an AI system is trying to decide whether a product is appropriate for a customer, vague or inconsistent product information becomes a commercial disadvantage. Missing attributes are no longer merely a catalogue-quality problem. They can become a discovery problem.

The same applies to availability, price, specifications, sizing, ingredients, compatibility, delivery options and policies.

This is one of the more interesting consequences of AI commerce because it shifts influence toward parts of the organisation that rarely considered themselves responsible for demand generation.

The person maintaining product taxonomy may suddenly be helping shape acquisition. I suspect they will be delighted to learn this once somebody explains the additional accountability.

Marketing measurement will have to become less platform-dependent

The obvious response to a new discovery channel is to instrument it.

We will undoubtedly build dashboards for AI traffic. Platforms will develop new referral classifications. Attribution vendors will create models promising to reconstruct the invisible journey. There will be conference slides.

Some of that will be useful. But I think the more important shift is to become less dependent on deterministic channel attribution in the first place.

If discovery becomes distributed across AI assistants, social platforms, creators, marketplaces, communities and environments the brand does not fully control, the question “Which channel caused this sale?” becomes increasingly difficult to answer with confidence.

That does not mean marketing becomes unaccountable.

It means accountability needs to move closer to incrementality and customer economics.

Did the activity create demand that would not otherwise have existed?

Did acquisition improve profitably?

Did new customers repeat?

Did branded demand grow?

Did conversion improve?

Did retention change?

Did customer lifetime value justify the cost?

Did the business sell more profitably than it would have without the intervention?

Those questions are less satisfying than a perfectly attributed conversion. They are also closer to the truth.

The organisation will need to relearn experimentation

This makes experimentation more important.

When direct observation becomes weaker, controlled testing becomes more valuable. Holdouts, geographic tests, incrementality experiments and carefully constructed interventions can tell a business what changed because of marketing rather than merely what happened after marketing.

This is not new thinking. What changes is the urgency.

For years, organisations could tolerate weak experimentation because platform reporting provided enough apparent certainty to keep budgets moving. If AI discovery reduces the visibility of the journey, leaders may have to become more comfortable with methods that offer stronger causal evidence but less day-to-day precision.

That is not always an easy sell.

Executives have become accustomed to dashboards updating continuously. An experiment that says “we will know more accurately in six weeks” can feel strangely unsophisticated beside a screen containing twelve decimal places.

The decimal places are seductive. They are not necessarily true.

Brand may become easier to underestimate again

There is another consequence.

If AI systems increasingly mediate product discovery, brand strength may matter in ways that are difficult to attribute directly.

An AI assistant making a recommendation may consider reputation, reviews, authority, availability and how consistently a product appears across trusted sources. Customers may also be more willing to accept a machine-generated recommendation when the brands presented already feel credible.

Some of that influence happens long before the customer clicks anything. This creates a familiar problem.

Activities that build mental availability, trust and reputation can become commercially important precisely because they operate outside the measurable transaction path.

Marketing has dealt with this tension before. Brand activity has always been harder to attribute cleanly than performance media. AI could widen that gap again just as boards and CFOs were becoming comfortable with the idea that digital made everything measurable.

It did not. It made more things visible. Those are different achievements.

AI discovery also creates a new optimisation problem

There is a temptation to turn all of this immediately into “AI optimisation”.

How do we get recommended more often? How do we appear in AI answers? How do we optimise product information for agents?

Those are sensible questions.

But the danger is repeating the early SEO playbook and optimising for the intermediary rather than the customer.

The objective should not be to trick an AI system into mentioning the brand. It should be to make the brand genuinely easier to understand, evaluate and recommend.

That means accurate product information. Strong reviews. Clear policies. Credible expertise. Reliable inventory. Useful content. Consistent information across the web.

In other words, many of the same things customers value.

That is reassuring. The best optimisation strategy may once again turn out to be making the underlying proposition better.

A disappointing conclusion for anyone hoping for a secret prompt-engineering budget line, but probably a healthier one for the business.

Customer journeys are becoming less like funnels

The funnel has survived largely because it is useful.

Awareness. Consideration. Conversion.

Easy to draw. Easy to explain. Fits beautifully into presentation software.

Customers have always been less cooperative.

Today they can discover something on social, ask AI for an explanation, investigate Reddit, visit a store, check a marketplace, forget about it for three weeks, receive a recommendation from a friend and eventually purchase directly.

Trying to reconstruct this as one linear journey becomes increasingly artificial. The more distributed discovery becomes, the more useful it may be to think in terms of signals and probabilities rather than paths.

What signals suggest demand is strengthening?

Which customer groups are becoming more interested?

Which content or experiences increase consideration?

Where does confidence fall?

Which interventions materially change behaviour?

This is a different style of marketing management. It accepts that the customer may not leave behind a convenient breadcrumb trail.

Measurement will need more judgment, not less

This is perhaps the uncomfortable conclusion.

For years, marketing technology promised to reduce the need for judgment by increasing the amount of data available. Better tracking would supposedly make investment decisions increasingly objective.

AI may now create the opposite situation.

We will have more information about customers, more sophisticated models and more powerful optimisation tools, while simultaneously losing visibility into parts of the journey where decisions are being formed.

Leaders will have to interpret imperfect signals.

They will need to distinguish measurement from causality, attribution from incrementality and precision from accuracy. That requires analytical maturity, but it also requires commercial judgment. The dashboard cannot make the decision for you if the dashboard cannot see the whole journey.

Perhaps it never could. AI is simply making that harder to pretend.

The customer journey is not disappearing. Our view of it is.

This is why I do not think AI discovery is fundamentally a search problem.

It is a measurement problem, a data problem, a product-information problem and eventually an operating-model problem.

Marketing will need stronger relationships with merchandising, technology, customer experience, data and commerce because all of those functions increasingly influence whether a product can be discovered and recommended.

The organisations that adapt will stop asking only how to attribute every transaction and become better at understanding what actually creates demand.

That may make marketing measurement less neat. It may also make it more useful.

The journey will continue. Customers will still discover, compare, hesitate, decide and buy.

We may simply know less about every step they took to get there. And the businesses that learn to make good commercial decisions anyway will have the advantage.

Own the customer. Own the data. Own the margin. Remove the intermediary. Build the relationship directly. If the economics looked difficult, the assumption was often that scale would eventually fix them.

There was logic behind it. Digital lowered the barriers to reaching customers, social created new forms of discovery, and many brands were understandably frustrated by how much control they surrendered to wholesalers and marketplaces.

Then reality became more expensive.

Customer acquisition costs rose. Paid media became less predictable. Returns remained painful. Fulfilment remained stubbornly physical. Marketplaces kept growing. Stores turned out to be useful in ways spreadsheets had occasionally forgotten. And many brands discovered that owning the customer journey also meant paying for every part of the customer journey.

The reaction now risks swinging too far the other way. DTC is dead, we are told.

I don’t think it is. What is dying is the idea that one channel needs to win.

Customers were never participating in the channel debate

This is the part that has always made the industry’s channel arguments slightly strange.

Customers do not wake up thinking, “Today I will participate in the direct-to-consumer ecosystem.” They discover a product on social, compare it on a marketplace, visit a store, ask a friend, read reviews, check the brand’s website, wait for payday, buy through an app and perhaps return it somewhere completely different.

From their perspective, this is simply shopping.

Inside the enterprise, the same journey can trigger a minor constitutional crisis. Which channel gets the sale? Which team owns the customer? Which budget funded the acquisition? Did the store influence digital, or did digital drive the store visit? Does the marketplace transaction count as good growth if we could theoretically have captured the customer directly?

These are legitimate management questions. The problem begins when the answers start shaping the experience more than the customer does.

I have seen this repeatedly in omnichannel businesses. The technology may be connected while the incentives remain separate. The customer experiences one brand; the organisation still experiences several competing P&Ls.

That is why I think the channel war is ending. The customer already won; They chose all of them.

Direct still matters. Just not for all the reasons we once claimed.

There are very good reasons to build a direct relationship with customers.

A brand-controlled experience matters. First-party customer data matters. Product storytelling matters. Testing, learning and feedback matter. So does the ability to build a relationship without depending entirely on somebody else’s algorithm, commercial rules or tolerance for your margin.

Owned commerce can also create attractive economics.

But “can” deserves emphasis.

A direct sale is not automatically a better sale once the real costs are included. Customer acquisition, fulfilment, returns, service, payment fees, technology, content, warehousing and discounting all have a habit of appearing after the gross-margin slide has already been circulated.

That was one of the more useful lessons from the first DTC wave: removing an intermediary does not remove the work the intermediary was doing.

Sometimes it simply sends you the bill.

Marketplaces have the opposite problem in strategic discussions. They are often framed as margin dilution or loss of customer control, which can certainly be true. But they can also provide discovery, trust, traffic, payments, logistics and geographic reach at a scale many brands would struggle to reproduce efficiently themselves.

Wholesale can still matter for distribution and credibility. Stores can still acquire customers, build trust, provide service and make categories easier to understand physically.

The question is not which channel is inherently superior. It is what job each channel should perform.

Channel economics can encourage very strange behaviour

This becomes especially obvious when organisations measure channels as if customers live permanently inside them.

Imagine a customer discovers a product through paid social, visits a store to try it, later buys online and eventually returns another item to the store.

Which channel “won”?

There is no particularly satisfying answer because the question itself is incomplete. Digital booked the revenue. The store may have created the confidence. Marketing created awareness. The physical return may have preserved the relationship. The customer could then make several more purchases across multiple channels over the next year.

Yet organisations frequently attempt to allocate the value of that relationship into separate boxes because the reporting structure requires it.

Some attribution is obviously necessary. Leaders need accountability, and an unprofitable store does not become profitable merely because we declare it part of an omnichannel ecosystem.

But when channel measurement creates incentives that reduce total customer value, the measurement framework has become part of the problem.

A store colleague should not hesitate to help a customer buy online because the revenue lands elsewhere. An eCommerce team should not optimise conversion in a way that creates avoidable store returns. A marketplace team should not chase GMV while ignoring what those customers are worth after fees, promotions and service costs.

The business should care about where value is created, not merely where the transaction happened to be recorded.

Obvious in theory. Messier when bonuses are involved.

The better unit of strategy is the customer relationship

This is where the next generation of commerce strategy becomes more interesting.

Instead of asking whether DTC, stores or marketplaces should dominate, ask what combination creates the strongest customer economics.

Where is discovery most efficient? Where does trust form? Which environment converts most naturally? What lowers returns? What increases frequency? Where does service matter most? Where can the business build first-party understanding without forcing customers into a journey designed largely for organisational convenience?

The answers vary enormously by category and market.

Luxury behaves differently from grocery. Furniture behaves differently from cosmetics. A customer who is entirely comfortable buying fashion through an app may still want to see a watch before spending several thousand dollars on it. A marketplace that works beautifully for one product line may destroy economics in another.

That variability is precisely why universal channel ideologies eventually fail.

Strategy becomes orchestration. The business decides what role each channel plays, how the economics work and how the pieces reinforce one another around the customer.

That is a considerably more mature conversation than “digital versus stores” or “owned versus marketplace”. It is also harder because there is no slogan to hide behind.

Stores are not the opposite of digital

Retail spent years positioning stores as the legacy part of the business while digital represented the future. In the early stages of eCommerce, that language was probably useful. Organisations needed to force investment into capabilities that had been underfunded for years.

It became less useful once customers began moving freely between both.

A store can be a customer-acquisition channel, a fulfilment node, a returns centre, an experience platform, a source of local trust and, in categories that benefit from expertise, one of the most powerful conversion tools the business has.

A good associate can still do something no recommendation engine consistently manages: understand the hesitation behind the question.

Digital provides different advantages. Range, convenience, memory, comparison, availability, personalisation and continuity before and after the store visit.

The real value comes when the two stop behaving like neighbouring countries with a complicated visa arrangement.

The customer should be able to discover digitally, experience physically and transact wherever makes sense. Inventory should move rather than forcing the customer to move. The associate should be able to access the relationship rather than only the transactions that happened inside that particular building.

That is not “store versus digital”. It is commerce.

Marketplaces are becoming harder to treat as optional

Marketplaces deserve the same nuance.

There are good reasons brands worry about them. Economics can be demanding. Customer data can be limited. Price transparency increases. Platform dependency is real.

But marketplaces increasingly sit where customers already are. That matters.

If a customer begins discovery on a marketplace because they trust the reviews, logistics and checkout experience, insisting that they should instead begin on the brand’s own website is a strategy built around organisational preference rather than customer behaviour.

The answer is not to surrender the relationship. It is to become much more deliberate about what the marketplace is for.

Perhaps it is primarily acquisition. Perhaps certain assortments belong there and others do not. Perhaps it extends geographic reach while the owned environment carries richer storytelling and loyalty. Perhaps the economics work beautifully in one category and terribly in another.

The same marketplace can be strategically useful and commercially dangerous depending on how the business uses it. Again, orchestration beats ideology.

AI will make channel ownership even less meaningful

The next shift may make much of the old debate look even more dated.

AI is beginning to change where discovery starts. A customer may increasingly ask an assistant to find the best product for a particular need, compare options, interpret reviews, check availability and explain the trade-offs.

Much of the consideration process could therefore happen before the customer visits a retailer, brand or marketplace at all. That changes what “owning the customer journey” means.

You may own the website without owning discovery.

You may own the transaction without owning the recommendation.

You may own the customer account while an external AI interface shapes the decision.

Commerce becomes more distributed.

This does not make owned channels less important. Brands will still need trusted destinations, strong product information, first-party relationships and differentiated experiences. But the centre of gravity shifts.

The brand needs to be understandable wherever the decision is being formed.

Product data starts behaving like marketing. Availability influences discovery. Reputation travels between surfaces. Returns and service policies may become inputs into machine-generated recommendations. The boundaries between marketing, merchandising, technology and channels become less convenient.

Once again, the customer ignores the organisation chart.

The brand still needs somewhere it owns

There is a danger in taking the orchestration argument too far.

If customers can be reached everywhere, why invest heavily in owned commerce at all?

Because dependency is still dependency.

A brand that relies entirely on paid platforms, marketplaces or external discovery layers is building its economics on rules it does not control. Algorithms change. Fees change. Access changes. Customer relationships can become mediated by businesses whose incentives are not identical to yours.

Owned commerce therefore remains strategically important. Its role simply becomes more realistic.

It does not need to be the place where every customer must transact. It is where the brand can build the richest relationship, understand customers most deeply, test quickly and express the experience without another platform deciding how much of it survives.

That is a valuable asset. It just does not need to win every sale to justify its existence.

Channel strategy is becoming portfolio strategy

The companies that manage this well will increasingly think about channels as a portfolio of capabilities rather than competing sales pipes.

Some channels generate demand. Some create trust. Some provide distribution. Some convert efficiently. Some improve retention. Some produce valuable data. Some deliver superior margin.

Very few are best at everything.

The executive task is therefore to understand both the economics and the strategic role of each channel, then design the system so the channels reinforce rather than cannibalise one another.

That requires better measurement. It requires shared customer KPIs alongside channel KPIs. It requires incentives that do not punish employees for doing the right thing across boundaries.

And it requires leadership to accept that not every transaction can be attributed with the precision the organisation would ideally like. There is a certain maturity in admitting that sometimes the customer bought because the whole system worked.

DTC was never really the destination

Direct-to-consumer was valuable because it forced brands to become closer to customers.

That lesson remains. The mistake was assuming closeness required exclusivity.

The strongest consumer businesses will continue building direct relationships, but they will also meet customers in stores, marketplaces, social environments, partner ecosystems and increasingly AI-mediated interfaces.

They will worry less about forcing the customer into the “right” channel and more about making every interaction strengthen the total relationship.

That is a very different operating philosophy. It also makes the word omnichannel feel increasingly inadequate. The ambition is no longer simply to connect a collection of channels. It is to build a business where the customer barely needs to know those channels exist.

DTC is not dead. The channel war is.

And the businesses that understand that will spend a lot less time arguing over where the sale belongs and a lot more time earning the next one.

The investment case gets approved, a programme is mobilised, consultants arrive, steering committees appear, integrations are built and eventually there is a go-live date. There may even be cake. A few months later, somebody asks the awkward question that should probably have been asked much earlier:

What changed in the business because we built it?

That is where many transformations become less impressive.

I have seen organisations implement genuinely capable platforms and still struggle to produce the economics promised in the original business case. The technology worked. The architecture was sound. The features existed. Yet people continued making decisions in spreadsheets, teams preserved old processes alongside new ones, managers asked for the same reports they had always asked for and frontline employees discovered inventive ways of avoiding the new workflow.

None of this means the technology failed. It means we confused technical deployment with business transformation.

Technology creates capability. Value appears only when people, processes and decisions change because that capability exists.

The distinction sounds obvious. In practice, companies violate it constantly.

Go-live is an engineering milestone, not a business outcome

One of the more misleading phrases in transformation is “go-live”.

It creates a sense of completion at precisely the moment the difficult part begins. The system is available. Now the organisation has to behave differently.

This distinction becomes particularly visible in large commerce and customer transformations. A new CRM platform can technically provide a single customer view, but that is irrelevant if store colleagues do not use it, customer care does not trust it or commercial teams continue making decisions using their own data sets. An analytics platform can produce beautiful dashboards, but the value remains theoretical if executives still ask an analyst to recreate the numbers in Excel before they believe them.

You can automate a workflow and discover that the organisation keeps the old manual version running “for safety”. At which point you have not automated anything. You have simply financed two processes.

These situations are often described as adoption problems, but even that language can be too forgiving. It suggests the technology has finished its job and the humans now need to catch up.

Sometimes they do. Sometimes the technology team has solved the wrong problem.

People do not always resist change. Sometimes they resist bad change.

“Resistance to change” is one of those wonderfully useful corporate diagnoses because it explains almost everything while holding nobody designing the transformation particularly accountable.

Employees are often portrayed as naturally reluctant adopters. Give them something new and they supposedly retreat toward familiar processes, spreadsheets and workarounds because change makes them uncomfortable.

There is some truth in that. Habit is powerful. But people also reject systems for perfectly rational reasons.

Perhaps the new workflow takes longer than the old one. Perhaps the system removes discretion they genuinely need to serve a customer. Perhaps employees are measured on outcomes that conflict with the new process. Perhaps the data is unreliable. Perhaps a manager who spent six months telling everyone to adopt the new dashboard still requests the old PowerPoint every Monday morning.

Or perhaps the product is simply unpleasant to use.

Technology teams occasionally underestimate this last possibility because once millions have been spent implementing something, usability criticism can begin to sound suspiciously like insubordination.

The most successful transformations I have seen take employee workarounds seriously. A workaround is not automatically evidence of resistance. It is often evidence that the operating model has discovered something the project plan did not.

If hundreds of intelligent people consistently avoid a process, leadership should resist the temptation to conclude that hundreds of intelligent people are the problem.

Adoption is not training

This is another distinction worth making.

Transformation programmes often measure adoption through training completion, active logins or feature usage. Those are useful indicators, but none tells you whether the technology has changed the economics of the business.

A salesperson can log into CRM every morning and still manage every meaningful customer relationship from WhatsApp.

A manager can open the analytics platform every week and still make the final decision using intuition and an offline spreadsheet.

A team can use an AI assistant hundreds of times while producing exactly the same amount of work, because the organisation has added the AI-generated output to the existing process rather than replacing anything.

Usage is activity. Adoption is behavioural change.

The useful questions are therefore harder. Which decision is now better? Which task disappeared? What cycle time reduced? What customer experience improved? What cost was removed? What revenue became possible? What risk declined?

If those things do not move, celebrating logins is rather like measuring gym memberships instead of fitness.

It is easier. That does not make it useful.

Every technology business case contains hidden behavioural assumptions

This is where the problem begins much earlier than implementation.

Most technology business cases contain assumptions about behaviour that are rarely written down with the same precision as the financial numbers.

A CRM investment assumes employees will capture customer information consistently and then use it differently.

An omnichannel programme assumes store and digital teams will cooperate around shared inventory and customer journeys.

An AI initiative assumes employees will trust recommendations enough to change decisions.

A planning platform assumes teams will stop maintaining parallel spreadsheets.

A new ERP assumes processes will standardise rather than recreate every historical exception inside a more expensive system.

The ROI model depends on these behaviours happening. Yet project governance often spends far more time debating architecture, timeline and budget than asking whether the organisation has actually created the conditions for those behaviours to change.

That is how a perfectly competent programme can deliver every technical milestone and still miss the commercial case.

The software was never supposed to create the value by itself. The behaviour was.

The operating model usually fights back

The more strategic the technology, the more likely it is to collide with the operating model.

Consider something as apparently straightforward as giving store colleagues access to a customer’s complete shopping history and allowing them to sell inventory from anywhere in the network. Technologically, this is solvable. Customer identity, product information, inventory visibility, payments and fulfilment can all be connected.

But then the organisational questions begin.

Who gets credit for the sale? Does the store target recognise an online order initiated by an associate? Who owns the customer? What happens when fulfilment comes from another location? Does the associate have permission to see the customer information? Which function funds the capability? Who owns the resulting service issue?

Suddenly the transformation is not about an interface. It is about incentives, ownership and economics.

This pattern repeats everywhere. Technology tends to expose boundaries that organisations have learned to tolerate. Once systems become connected, the logic of keeping the operating model disconnected becomes harder to defend.

That is why some of the hardest transformation decisions are not technical at all. They involve changing who owns what, how success is measured and which old behaviour the organisation is genuinely prepared to stop.

Deleting the old process is often more transformational than launching the new platform. It is also considerably more politically difficult.

AI is about to make the adoption gap much more visible

The current wave of AI makes this problem particularly interesting because the technology is advancing faster than most organisations can redesign work around it.

It is relatively easy to give thousands of employees access to an AI tool. It is harder to answer what work should now disappear.

If AI can prepare the first draft of an analysis in five minutes, but the company retains every review step built around the old three-day process, the technology creates speed and the organisation promptly gives it back.

If an AI agent can identify a customer issue and recommend a resolution but still requires multiple functional approvals before acting, the bottleneck was never intelligence.

If coding assistants make developers substantially more productive but release governance, testing processes and decision rights remain unchanged, the improvement may simply move the queue somewhere else.

This is why I am sceptical when companies announce the number of AI licences they have deployed as evidence of transformation.

It tells me how much software they bought. It tells me very little about what the enterprise can now do differently.

The AI opportunity is real, but the value will accrue disproportionately to companies willing to redesign workflows rather than simply insert intelligence into them.

Otherwise we will use extraordinarily capable technology to accelerate organisational habits that probably needed removing anyway.

There would be something almost poetic about that.

The transformation team should own value, not delivery

A deeper issue is how transformation programmes define success.

Large programmes naturally become obsessed with delivery. Scope, budget, milestones, risks, dependencies and go-live dates dominate governance because they are tangible and manageable. Somebody needs to ensure the platform actually gets built.

But once the project is organised almost entirely around delivery, business value can become somebody else’s responsibility.

Technology delivers the platform. The business is expected to adopt it.

Benefits are then tracked through a separate mechanism months later, often when the programme team has already moved on.

I think this separation needs to become much less comfortable.

A transformation programme should be able to articulate not only what capability will exist, but which behaviours will change, which old processes will stop, which decisions will move and which commercial measures should respond.

And those benefits should have named executive owners. If nobody owns the behavioural change required to produce the ROI, the ROI is not a business case.

It is a hope with a spreadsheet attached.

Real adoption occasionally requires making the old world impossible

There is a practical lesson here that organisations often avoid because it feels aggressive.

Sometimes adoption improves only when the old route disappears.

If the old system remains available indefinitely, people will use it.

If the old spreadsheet continues to be accepted in the executive meeting, somebody will maintain it.

If managers tolerate manual approval outside the new workflow, the manual approval will survive.

If employees can ignore the new customer view and still achieve their performance targets, many will.

This is not because people are malicious. They are usually optimising for the shortest path to getting their work done.

Transformation therefore requires choices.

Which system is authoritative? Which process ends? Which metric changes? Which behaviour is no longer acceptable? Which decision now happens differently?

There is an important difference between encouraging adoption and designing for adoption. The latter requires leadership to remove ambiguity.

That can feel uncomfortable because organisations often want the productivity benefits of transformation without accepting the disruption required to produce them.

Unfortunately, the technology rarely negotiates that compromise on our behalf.

The most useful question comes six months later

If I were reviewing a major technology investment six months after launch, I would be much less interested in the original feature list than in what the organisation has stopped doing.

What reports disappeared?

What manual work disappeared?

Which decision happens faster?

Which customer interaction improved?

Which cost left the P&L?

Which layer of coordination became unnecessary?

Where did conversion, productivity, service or margin actually move?

The answers reveal whether technology became embedded in the operating model or merely attached to it.

This also changes the executive conversation around transformation. Instead of asking whether the programme was delivered, leadership begins asking whether the business absorbed the capability.

That is a higher standard. It is also the only one that ultimately matters.

Enterprises will continue buying extraordinary technology. AI will make that technology more capable, more accessible and, in many cases, genuinely transformative. But capability sitting unused inside an organisation has no economic value.

The transformation happens when people trust it, processes change around it, leaders remove the old behaviours and the business begins making different decisions because the capability exists.

Technology does not create value. What the organisation learns to do with it, does.

The danger is that we describe the opportunity too narrowly.

If retail media becomes simply another advertising business, retailers will eventually compete on inventory, audience and price in much the same way media businesses always have. The more interesting advantage sits somewhere else: retailers know what people actually buy.

That sounds almost embarrassingly obvious. Yet it changes the economics of the entire proposition.

Retail media is not ultimately about putting more advertising around commerce. It is about turning commerce itself into intelligence.

The transaction changes what media can know

Traditional advertising has always lived with a gap between influence and outcome. A customer sees something, perhaps clicks something, perhaps remembers something and eventually, somewhere, may buy something. The industry has built increasingly sophisticated ways of estimating what happened in between.

Retailers sit much closer to the truth.

They can often see the search, the product view, the basket, the transaction and the repeat purchase. With loyalty or identified customer data, they may also understand how that behaviour develops over time.

That creates a fundamentally different media proposition.

A brand no longer needs to ask only, “Did my campaign generate impressions?” It can ask whether the campaign produced incremental sales, whether it attracted new customers, whether they purchased again, whether the promotion simply subsidised existing demand, and whether the resulting customer was economically attractive.

Those are not advertising questions. They are commercial questions.

And this is where I think retail media starts becoming much more strategically important than the label suggests.

Closed-loop measurement is valuable because most marketing is still messier than we admit

Anyone who has spent enough time around digital marketing eventually develops a healthy suspicion of attribution.

The dashboards are usually very precise. Reality less so.

Customers encounter brands across stores, marketplaces, search, social, creators, email, apps and countless small interactions that resist being squeezed neatly into a conversion path. Yet organisations still make large investment decisions based on models that sometimes confuse visibility with causality.

Retail media improves that situation because the media exposure and the commercial outcome can sit inside the same ecosystem.

That does not magically solve incrementality. A customer who clicked a sponsored product may have bought it anyway. A promotion can move demand forward without creating new demand. A beautifully attributed sale can still be economically poor once discounting and margin are included.

But retailers have something most publishers do not: the ability to connect advertising much more closely to actual commerce.

Handled properly, that should make the conversation between retailers and brands considerably more intelligent.

Handled badly, it will simply produce better-looking dashboards.

There is a difference.

The temptation will be to monetise every available surface

This is where the economics can become dangerous.

Retail media is attractive precisely because the incremental margin can look much better than the underlying retail business. Once executives see that, an entirely predictable thing happens: previously innocent pieces of digital real estate begin developing revenue targets.

Search results become inventory. Category pages become inventory. Email becomes inventory. Apps become inventory. Screens in stores begin looking suspiciously under-monetised.

Eventually someone will probably attempt to monetise the loading spinner.

The commercial logic is understandable. The customer logic needs more care.

A shopper comes to a retailer because they want to find and buy something. Advertising can improve that experience when it helps discovery, introduces relevant products or gives brands useful visibility. It damages the experience when paid placement begins overriding relevance.

That distinction is critical.

Retailers have something far more valuable than advertising inventory: customer intent. They should be very careful not to degrade it while trying to monetise it.

If the highest bidder consistently beats the best product, retail media becomes a tax on customer experience.

The short-term P&L may enjoy it. The customer eventually notices.

The real opportunity is helping brands make better commercial decisions

The most interesting conversations in retail media are therefore not about CPMs.

They are about questions brands genuinely struggle to answer.

Which customer groups are expanding? Which products are bringing new shoppers into the category? Which attributes are beginning to appear in search before they show up in sales? Which promotion generated incremental demand rather than moving an existing purchase forward by two weeks?

How does a new product behave after launch? Which customers repeat? Which combinations appear together in baskets? Where is price sensitivity shifting? What happens to customers after they first enter the brand?

The retailer can potentially answer these questions because it sees commerce at a level of detail few other partners can.

At that point, the relationship starts to change. The retailer is no longer saying to the supplier, “Buy more impressions from us.” It is saying, “We can help you understand your customer and your category better.” That is a much more defensible position.

Media inventory is eventually comparable. Intelligence is harder to commoditise, particularly when it comes from proprietary transaction behaviour.

This also changes the supplier relationship

Retailers and suppliers have always had a commercially complicated relationship.

They negotiate price, margin, promotional funding, placement, assortment, inventory and increasingly access to customer attention. Retail media adds another economic layer to that relationship.

It can be genuinely valuable for both sides. Brands gain access to high-intent audiences and better measurement. Retailers create an additional revenue stream and can fund improvements in customer experience and technology.

But it can also become a sophisticated version of trade spend if nobody is careful.

Suppliers may feel compelled to purchase media simply to preserve visibility they once received organically. Merchandising decisions can become entangled with advertising commitments. Commercial negotiations can become harder to separate from media budgets.

At some point, retailers have to decide what kind of business they are building.

If retail media is simply another mechanism for extracting more money from suppliers, it may generate revenue without creating much additional value.

If it helps suppliers understand demand, reach the right customers and invest more intelligently, the ecosystem becomes stronger.

Those are very different propositions, even if the invoice sometimes looks similar.

The operating model is more complicated than the revenue line suggests

Retail media also has an awkward habit of crossing organisational boundaries.

Marketing sees media. Commercial teams see supplier monetisation. Digital sees valuable customer surfaces. Data teams see first-party audiences. Finance sees margin. Technology sees another platform to integrate. Customer-experience teams see a new group of people asking whether the homepage really needs all that empty space.

Everyone is correct. Which is precisely the problem.

A serious retail media business cannot sit comfortably inside one traditional function because its economics depend on several of them working together. Customer data has to be trusted. Inventory needs to be governed. Measurement needs credibility. Supplier relationships matter. Customer experience needs protection.

This is why I am wary of retailers treating retail media as a side project owned by whoever happened to sell advertising first.

Once the business becomes material, governance becomes strategic.

Who decides how much sponsored inventory is too much? Who arbitrates between relevance and revenue? Who owns measurement methodology? Who decides whether a supplier’s media investment should influence commercial treatment elsewhere?

These sound like operational details until the revenue becomes large.

Then they become political details.

AI will make the intelligence layer more valuable

AI will make advertising production cheaper and optimisation faster.

Brands will be able to generate creative variations, personalise messaging and adjust campaigns at speeds that would have required significant teams only a few years ago. Buying and optimisation will become increasingly automated as well.

If execution becomes easier for everyone, proprietary insight becomes more valuable.

That is where retailers have an advantage.

They can combine customer behaviour, product information, transactions, pricing, promotions, availability and potentially service data. AI can make those signals easier to interpret and convert into recommendations.

Imagine moving beyond campaign reporting toward continuously answering questions such as: Where is demand emerging? Which customers are becoming less price-sensitive? What product attributes are driving repeat behaviour? Where is a brand acquiring customers but failing to retain them?

That begins to look less like advertising technology and more like commercial intelligence. And it could become one of the more interesting data businesses sitting inside modern retail.

Retailers should not confuse a new revenue stream with a new strategy

There is a broader lesson here.

Every retailer is understandably interested in additional profit pools. Retail is not famous for excessive margins, and finding businesses with different economic characteristics is strategically attractive.

But the most valuable adjacencies usually strengthen the core business rather than simply sit beside it.

Payments can make transactions easier. Logistics can improve fulfilment. Loyalty can deepen customer understanding. Retail media should ideally do something similar.

It should make the retailer more useful to customers and suppliers while creating a new source of economics. If those benefits separate, the business becomes fragile.

A retailer that maximises media revenue while making product discovery worse is extracting from the core. A retailer that uses commerce intelligence to improve supplier investment, customer relevance and its own economics is strengthening it.

That distinction will matter as retail media matures.

The name may eventually become the least interesting part

The industry will continue calling it retail media because advertising is where the business started.

I suspect the most advanced versions will eventually feel much broader.

Media will remain important, but around it will sit measurement, audience intelligence, product insight, supplier analytics, customer understanding and perhaps entirely new commercial services that do not fit comfortably into an advertising budget.

That evolution makes sense. The retailer’s unique asset was never the banner position on the category page. It was the transaction underneath it.

For years, retailers treated that transaction largely as the end of the customer journey. Retail media reveals that it can also be the beginning of an intelligence business.

The retailers that understand that distinction will build something much more valuable than another advertising network.

They will build a better understanding of commerce itself.

I think that framing is already becoming too small.

Once AI begins researching, writing, analysing, forecasting, prioritising, negotiating, answering customers, recommending actions and increasingly executing them, it is no longer behaving like a conventional piece of enterprise software. It is participating in the work.

That distinction matters because most companies are trying to introduce a new form of productive capacity without materially changing the organisation around it.

We are adding intelligence to the workforce faster than we are redesigning the workforce itself.

We are automating tasks while preserving the jobs around them

The first wave of enterprise AI has been very use-case driven. Marketing gets content generation. Technology gets coding assistants. Finance gets automated analysis. Customer care gets conversational agents. Merchandising gets forecasting. HR gets screening and summarisation.

There is nothing wrong with starting there. In fact, it is usually the sensible way to prove value without redesigning the company before anyone knows whether the technology works. The problem comes when those pilots succeed.

If AI can absorb 20 or 30 percent of the work inside a role, the natural enterprise response is often to celebrate the productivity improvement and leave the role exactly as it was. The same reporting line remains, the same approvals remain, the same meetings remain and, quite often, the same expectation that a human will continue producing the thing the AI just produced in case anyone asks.

That is how you end up with AI preparing the analysis in three minutes and an organisation spending three weeks discussing it.

Technically impressive. Organisationally untouched.

Over time, this becomes increasingly difficult to defend. If a manager previously spent half a day assembling information and AI now does that almost instantly, the more interesting question is not how many hours were saved. It is what that manager should be doing with the capacity that has just been created.

Enterprises have historically been much better at adding work than removing it. AI is going to test that habit.

The more useful unit of redesign is the decision

The conversation about AI and jobs tends to become emotional very quickly, and understandably so. Jobs are how organisations are structured, how people are paid and how careers are built.

But jobs may be the wrong place to begin. I think the better unit of analysis is the decision.

Every business contains thousands of recurring decisions. Which customer receives an offer. Which product gets reordered. Which service case needs escalation. Which campaign should stop. Which supplier needs intervention. Which price should move. Which stock should be transferred. Which applicant deserves a second look.

Some of those decisions should remain primarily human. Some will eventually be better made by machines. A large number will sit somewhere in between, with AI handling the analysis and recommendation while a person retains judgment or authority.

That is the interesting territory. The enterprise of the future is unlikely to be divided neatly between “human work” and “AI work”. It will contain thousands of different combinations of machine speed and human judgment, and those combinations will vary depending on consequence, confidence, context and risk.

A pricing recommendation for a low-volume commodity is not the same as a decision affecting the positioning of a luxury brand. A customer-service agent resolving a delivery query is not the same as one deciding compensation after a serious service failure. A forecast is not a negotiation.

The technology may be similar. The decision rights should not be.

Human judgment becomes more valuable, not less

There is a temptation in AI discussions to assume that if machines become better at intellectual work, human judgment becomes less important.

I suspect the opposite will happen in many areas.

AI is extremely good at processing volume, recognising patterns and producing answers quickly. It does not become tired, distracted or quietly decide that the spreadsheet can wait until tomorrow morning because lunch has become urgent.

That is useful. But speed and intelligence are not the same as judgment.

Commercial decisions carry context that is difficult to reduce to one objective function. A model may correctly identify that a promotion will increase conversion and still be wrong for the brand. It may optimise inventory efficiently while damaging a strategic supplier relationship. It may recommend the customer most likely to respond to an offer without understanding that repeatedly discounting to that customer is teaching exactly the wrong behaviour.

Humans are not immune to making these mistakes, obviously. We have achieved quite a respectable track record of bad judgment without any assistance from artificial intelligence.

But the point is that AI changes where the premium sits. When average analysis becomes cheaper and faster, the value of interpretation rises. When everybody can generate an answer, choosing which answer matters becomes more important. When information is abundant, taste, context and restraint become more valuable.

The strongest organisations will therefore not simply automate judgment away. They will become much more deliberate about where judgment is required.

Management may be more disrupted than many individual jobs

One of the less discussed consequences of AI is what it does to management.

A great deal of managerial work exists because information has historically moved slowly through organisations. Managers collect it, interpret it, package it, escalate it, translate it and carry it from one meeting into another. Entire reporting structures have evolved around the friction involved in moving knowledge through a large company.

AI attacks some of that friction directly. If information can be assembled continuously, anomalies identified automatically and recommendations produced on demand, some layers of coordination become less necessary. That does not mean managers disappear, but it should change what good management looks like.

The manager who built value by knowing where every piece of information lived may find that advantage eroding. The manager who creates clarity, develops people, understands customers, resolves ambiguity and makes good calls under uncertainty becomes more important.

That is a healthier definition of management anyway. It also means AI transformation will eventually collide with organisational hierarchy. Companies cannot keep removing informational friction while assuming that every layer created to manage that friction must remain untouched.

This will not happen neatly. Organisation charts have emotional lives of their own.

AI also exposes how much work exists because the enterprise is complicated

There is another uncomfortable possibility. Some work exists because the organisation itself created the problem.

People reconcile systems that do not talk to one another. They prepare reports because another team cannot access the source data. They chase approvals because ownership is unclear. They sit in meetings because the decision cannot be made where the information originates.

Over time, these behaviours become jobs, processes and occasionally departments.

AI can automate some of that activity, but there is a danger in celebrating the automation too early. If we use AI simply to make organisational bureaucracy faster, we may improve the efficiency of work that should not exist.

That would be a very enterprise outcome. Before asking where an agent can be inserted into a workflow, leaders should probably ask why the workflow has so many steps in the first place.

AI is useful precisely because it forces the organisation to describe its own work clearly. Once you try to automate a process, the exceptions, duplicated approvals and political compromises that humans have been quietly navigating suddenly become visible.

The machine has not broken the operating model. It has merely stopped pretending the operating model makes sense.

The workforce will become blended before the organisation admits it

I use the term symbiotic enterprise because I think it describes where this is heading more accurately than “AI-enabled organisation”.

An AI-enabled organisation sounds like the old company with some better tools.

A symbiotic enterprise is different. Human and machine capabilities are designed around one another. The human does not merely use AI. Work is intentionally divided according to what each side does well.

Machines provide scale, memory, pattern recognition, monitoring and speed. Humans provide context, empathy, creativity, ethics, relationships and judgment where the answer cannot be reduced cleanly to probability.

The boundaries will move constantly. A task that requires human oversight today may become reliably automated in two years. Another may remain stubbornly human because the cost of getting it wrong matters more than the efficiency of getting it right. That means organisational design cannot remain static while AI capability keeps changing.

The old approach of redesigning the operating model every five years will begin to look increasingly strange when the productive capability inside the operating model is changing every six months.

This is really a leadership question

Most executive AI conversations still begin with capability.

What can the technology do?

A more consequential conversation begins with organisational design.

Which decisions should move closer to the customer? Which activities should disappear rather than simply become faster? Where does human judgment create genuine value? Which managers need different skills? What happens when an AI agent can act without waiting for the meeting where everybody would normally discuss whether it should act?

Those are harder questions because they eventually touch power.

AI changes access to information. It changes who can make decisions. It changes how much coordination is required. It changes the value of expertise that previously depended on controlling knowledge or navigating process.

Technology transformations become political the moment they change who needs whom. This is why I do not think the companies that win with AI will necessarily be the ones that deploy it fastest.

The advantage will come to organisations willing to redesign around what the technology makes possible, including the parts of the organisation that leaders would rather leave alone.

The organisation eventually has to notice

We have seen this pattern before.

Mobile was initially treated as another channel before it changed customer behaviour. Cloud was treated as an infrastructure decision before it changed how technology organisations operated. Digital was treated as a department before it spread into almost every part of the business.

AI will follow a similar path, but I suspect much faster. At first it will sit beside the organisation. Then inside its workflows. Then inside its decisions.

Eventually, trying to distinguish between “AI work” and “normal work” will sound as peculiar as asking which part of the business uses the internet. That is when the real transformation begins. Not when the enterprise owns AI.

When the enterprise has redesigned itself around a new reality: intelligence is no longer scarce, and some of the productive capacity inside the company is no longer human.

The question is not whether organisations will become symbiotic.

They already are. The question is how long it takes the organisation chart to catch up.

And yet the uncomfortable question remains: are customers actually becoming more loyal, or have we simply become better at rewarding transactions?

I think those are two very different things.

Most established retailers now know more about their customers than at any point in history. They know what people bought, what they browsed, when they last visited, which categories they prefer, how frequently they shop, what they returned, which communications they opened and, increasingly, where they sit on some beautifully constructed propensity curve.

The problem is no longer the absence of customer data. The problem is whether any of that knowledge makes the next experience meaningfully better. That is why I increasingly think retail does not have a loyalty problem. It has a relevance problem.

We have become very good at recognising customers

Recognition is relatively easy now. A retailer can identify that I bought running shoes three months ago, opened an email about sportswear last week and have enough points to qualify for a reward.

Relevance requires something harder: understanding why any of that matters now.

The difference sounds subtle until you experience it as a customer. Buy a sofa and you may spend the next month being shown more sofas. Purchase a luxury handbag and an algorithm occasionally concludes that your principal remaining ambition in life is another luxury handbag. Buy something for a child and suddenly an entirely new demographic identity is assigned to you.

None of these systems is technically wrong. That is almost the problem. They recognise the transaction perfectly. They understand the customer rather less well.

A purchase is not always a preference. It may be a gift, an event, a replacement, an experiment or something bought because the item the customer really wanted was unavailable. Behaviour without context is useful data, but it is not understanding.

Retail has spent a lot of money closing the first gap and rather less attention on the second.

Loyalty became a marketing programme when it should have become an operating philosophy

Somewhere along the way, loyalty became something the CRM or marketing team largely owned. That made organisational sense because programmes needed campaigns, rewards, communications and increasingly complex customer segmentation. Customers never received that organisational chart.

Their willingness to return is shaped by considerably more than the loyalty programme. It is influenced by whether the item was available, whether delivery happened when promised, whether a return was painless, whether service remembered the problem they had yesterday, whether the store associate could see what they bought online and whether the brand treated them like the same person from one interaction to the next. A perfectly targeted offer cannot compensate for a badly fulfilled promise.

This is one of the things that becomes obvious when you have operated across both the commercial and technology sides of retail. The customer experience is horizontal, while most enterprises are organised vertically. Marketing owns one piece, digital another, stores another, service another, technology another, supply chain another.

Then we wonder why the customer relationship occasionally feels as if several companies are taking turns. Real loyalty sits in the gaps between those functions.

More personalisation will not solve this by itself

Personalisation has become another word that means so many things it risks meaning very little.

At the shallow end, it is still a first name in an email or a recommendation based on the previous transaction. At the sophisticated end, it can involve propensity models, behavioural signals, next-best-action engines and increasingly AI interpreting enormous amounts of customer context in real time.

The technology is getting dramatically better.

That does not guarantee the experience will.

A retailer can have an extraordinarily sophisticated model predicting what a customer might do next and still send the wrong message because the commercial calendar says it is promotion week. A system can identify a high-value customer and still subject them to the same returns process as everyone else because policy lives somewhere else. AI can recommend precisely the right product while the inventory system confidently promises stock that does not exist.

Intelligence without operational authority becomes an expensive observer.

This is why I think the next phase of personalisation will be less about generating more messages and more about improving decisions.

Who actually needs an incentive?

Who would have purchased without one?

Which customer is quietly drifting away?

Who has just experienced a service failure and should probably not receive an enthusiastic “We miss you!” email 12 hours later?

When should the business intervene?

And, just as importantly, when should it leave the customer alone?

That last question deserves more attention. Relevance sometimes means doing less.

The economics of loyalty are changing

Traditional loyalty mechanics were largely designed around frequency and spend. Spend more, earn more. Return more often, progress further.

Those mechanics still work, particularly in categories where purchase frequency is naturally high. But the economics become more interesting when the organisation stops treating loyalty as a discount mechanism and starts treating it as a customer-value system.

A valuable customer is not necessarily the customer with the highest recent spend. Someone purchasing heavily during a promotion, returning frequently and requiring expensive service may look very different once contribution economics are considered. Another customer may purchase less frequently but buy full price, shop across categories, engage through multiple channels and remain with the brand for years.

The distinction matters because loyalty investment is still investment.

Every reward has a cost. Every discount transfers margin. Every communication consumes attention. Every service intervention uses capacity.

The commercial question is not simply, “How do we make the programme more generous?”

It is, “Where does the next unit of loyalty investment actually create incremental customer value?”

That pushes loyalty much closer to customer economics. It also makes the conversation more interesting for the CFO.

AI could finally make customer data useful — but only if the organisation is ready

This is where AI has genuine potential.

Retailers have accumulated years of transactional, behavioural, service and engagement data. Historically, much of it has been difficult to connect quickly enough to influence a live customer decision. Different systems held different fragments of the relationship, and the operational cost of turning those fragments into action was high.

AI can change some of that. It can interpret more signals, identify patterns humans would miss, recognise context faster and potentially help decide what the business should do next. That opens the door to something far more valuable than “personalised marketing”.

The chain I care about is:

Data → Intelligence → Relevance → Retention → Lifetime Value

Notice that data sits at the beginning, not the end.

This matters because retailers have sometimes behaved as if assembling the customer data platform was itself the achievement. It is not. A beautifully integrated customer profile that changes no customer decision is an expensive database with excellent internal publicity.

The value appears when intelligence changes action. That requires more than algorithms. It requires clear ownership, trusted data, connected systems and enough organisational flexibility for the recommendation to influence pricing, service, communication, inventory, experience or whatever else the customer actually needs. This is where the loyalty discussion quickly becomes an operating-model discussion.

It usually does.

Relevance is not the same as knowing everything

There is another tension here that retail will need to navigate carefully.

The more data businesses collect, and the more capable AI becomes at interpreting it, the easier it becomes to cross the line between helpful and unsettling.

Customers generally appreciate being understood when the benefit is obvious. Remember my preferences. Make the return easier. Show me the right size. Tell me something is back in stock. Do not ask me to repeat information I already gave you.

That feels useful. The same customer may react very differently when personalisation reveals just how much the company knows about them without offering any corresponding value. The standard should therefore not be maximum personalisation.

It should be appropriate relevance. That requires judgement, which is inconvenient because judgement is considerably harder to put into a transformation roadmap than “deploy personalisation engine Q3”.

Still, it matters. Trust is part of loyalty too.

The best loyalty programme may eventually become invisible

I suspect this is where the category is heading.

The future of loyalty may look less like a programme customers consciously participate in and more like a relationship that simply becomes better over time.

The retailer knows enough to make useful decisions. Store colleagues recognise the relationship appropriately. Service understands the history. Marketing becomes more selective. Offers reflect genuine intent rather than blanket promotion. Customers are rewarded for the behaviours that actually matter to the business, while the business becomes progressively more useful to them.

Points may still exist. Tiers may still exist. Benefits certainly will.

But they become mechanics underneath the relationship rather than the relationship itself. That is a subtle but important shift.

For years, retail has asked how to make customers more loyal to the programme. The better question may be how to make the business more worthy of the customer’s loyalty. Because once the price is broadly competitive and the product is broadly available, the advantage increasingly comes from something harder to copy:

knowing the customer well enough to matter, without making them feel like they are being managed by a database.

That is not a loyalty mechanic; It is relevance. And relevance may turn out to be the most valuable loyalty currency of all.