Your Customer Didn't Visit Your Website. Their AI Agent Did.

Imagine you have your eye on a jacket, but not at £220. You prompt your AI agent: check it every morning and tell me when it falls below £150.

For the next two weeks, the agent does exactly that.

From your point of view, very little has happened. You expressed an interest once. You have not returned to the retailer's website or reconsidered the jacket fourteen times.

But the retailer may see fourteen product checks: repeated engagement, increasing recency and frequency, potentially attached to the same customer identity that feeds affinity, propensity, segmentation and next-best-action models.

A system designed around human browsing could reasonably read that pattern as rising purchase intent.

The reality is different: one customer prompt, zero human revisits, fourteen automated checks.

Now take the thought experiment one step further. If propensity helps determine who needs an incentive, repeated product engagement might suggest that this customer is increasingly likely to buy without one. The agent searching for a better price could therefore contribute signals that make the offer it is looking for less likely to appear.

There is no evidence here that this specific failure is happening in production today. The point is that the pieces of the architecture already exist — and their assumptions are starting to collide.

The behavioural contract has changed

Digital personalisation is exceptionally good at learning from proxies.

A product view suggests interest. A return visit strengthens it. Search reveals a need. Basket activity suggests consideration. Recency and frequency change propensity. Hundreds of small signals accumulate until a platform has enough evidence to choose the next product, message, offer or experience.

In simplified form:

customer → behaviour → inferred intent → treatment

There have always been complications: shared devices, cookie loss, identity resolution and bots. But the semantic assumption has remained surprisingly durable. When an interaction is attributed to a customer, we generally treat it as something the customer did.

AI agents create a different relationship:

customer → authorised AI agent → interaction

That interaction is automated, but it is not necessarily noise. It may represent extremely strong customer intent.

So the familiar human-versus-bot distinction is no longer enough. Discard every agent interaction and we may throw away valuable intent. Treat every agent interaction as ordinary human behaviour and we may distort the customer profile.

Spotify Research's 2026 RecSys paper Who Are We Recommending To? makes the adjacent point from the recommendation side: when AI agents mediate between people and recommender systems, assumptions about preferences, engagement signals and evaluation need to be reconsidered. ¹

There is another assumption worth testing too: that behaviour attributed to a customer was actually performed by that customer.

Identity is being solved faster than meaning

This is not waiting for some distant generation of autonomous shopping agents.

Visa's Trusted Agent Protocol is designed to help merchants distinguish approved commerce agents from malicious bots, identify the purpose of an interaction and, with consumer consent, recognise the customer an agent represents. ²

Cloudflare's Web Bot Auth work uses cryptographic HTTP signatures so agent traffic can be verified rather than guessed from a user-agent string or IP address. Its work with payments networks explicitly considers trusted agents browsing and transacting on behalf of consumers. ³

Mastercard's Agent Pay and Verifiable Intent work tackles the problem from the authorisation side: proving that an agent's action reflects authenticated user intent and consent. ⁴

The infrastructure is therefore getting better at answering:

Is this a legitimate agent? Who does it represent? What has it been authorised to do?

The personalisation question is different:

What should the agent's behaviour mean?

Dynamic Yield illustrates why that matters. Its affinity capabilities learn from customer preferences, purchases and real-time interactions, with recent and high-intent engagement carrying greater significance. ⁵ That is exactly what a good personalisation engine should do.

But an agent executing the same price-checking prompt every morning has created repeated, recent engagement without repeated human consideration.

This is not evidence that Dynamic Yield — or any other named platform — mishandles agent traffic. Public product material does not expose enough implementation detail to make that claim. It shows why actor provenance becomes important to the meaning of an otherwise legitimate behavioural signal.

Adobe is simultaneously moving customer profiles towards richer conversational intent and agentic experiences, while Bloomreach describes AI agents acting across the commerce journey and conversational interactions becoming a new source of ecommerce insight. ^{6 7}

Microsoft Research goes further conceptually. From Hidden Profiles to Governable Personalization argues that LLM assistants acting between users and platforms change where user representations are created and used, raising questions around intent translation, ownership and accountability. ⁸

The direction is clear: richer intent, agent-mediated journeys and increasingly verifiable delegation.

What is less visible in current public material is a mature answer to how external consumer-agent activity should alter the human customer's behavioural profile.

The missing dimension is provenance

A simplified customer event might contain:

customer_id + event + timestamp + channel/context

Agent-mediated interaction needs more context. A useful conceptual model could distinguish:

• principal — the customer being represented
• actor — human, authorised agent, crawler or other automation
• agent identity — the verified agent or provider
• delegation — what the customer has authorised
• purpose — browse, monitor, compare, purchase or service
• declared intent — the customer's explicit objective, where appropriate and consented
• profile policy — whether and how the event should influence affinity, propensity, experimentation or marketing.

This is not a proposed standard. It is a way of describing the semantic gap.

With provenance, fourteen product checks can remain useful without pretending they represent fourteen moments of human consideration. The system can understand that a customer gave one instruction and an authorised agent executed it repeatedly.

That matters because the prompt and its behavioural exhaust can point in opposite directions.

The agent knows: the customer is unwilling to buy at the current price.

A behavioural model may infer: the customer is increasingly interested in buying.

The same activity carries two very different meanings depending on who acted, under what instruction and why.

Intent may become better data than behaviour

This is where the opportunity becomes more interesting than the risk.

Personalisation has spent years getting better at inference because customers rarely tell brands exactly what they want:

observe behaviour → infer intent → choose treatment

An authorised agent may already know the answer.

The prompt "buy this when the price falls below £150" contains information that fourteen product views cannot reliably provide. It expresses the product, the customer's objective, a price condition and potentially permission to act.

The model can invert:

receive delegated intent → verify authority and provenance → choose treatment → observe outcome

Recommendation systems could work with explicit constraints rather than guessing them. Brands could reduce irrelevant retargeting. Customers could gain more control over the preferences represented about them. And merchants and agents could avoid repeatedly loading pages just to discover whether a condition has changed.

Why should an agent visit the same product page every morning if the merchant can support a machine-readable instruction such as: notify this authorised agent when the price is below £150?

Explicit intent creates new questions. Does the customer want the merchant to know their reservation price? Should that information influence an incentive? How much intent should an agent reveal to complete a task?

Those questions are important. But they do not make today's behavioural model safer. They show that agentic commerce changes the data contract as well as the interface.

From inferred intent to governed intent

The first response to AI-agent traffic will understandably be defensive: identify it, authenticate it, stop malicious automation and keep analytics clean.

That is necessary, but it is not the destination.

The larger opportunity is to distinguish the customer, the actor, the instruction and the intent behind an interaction — and decide explicitly what each signal is allowed to change.

For personalisation leaders, CDP teams and digital architects, the next evolution may not simply be a better propensity model. It may be a better representation of why an interaction happened at all.