Fifty-three per cent of customers who experienced marketing personalisation reported a negative experience. Those same customers were 3.2 times more likely to regret the purchase and 44 per cent less likely to buy from that brand again. The numbers are Gartner’s, from its 2025 personalisation survey, and they invert the founding assumption of the entire discipline: that more personalisation is better.

Sit with the arithmetic for a moment. A majority of the people your personalisation engine reaches are being damaged by it. Continuously, across millions of interactions, with your brand on the invoice.

That is the personalisation paradox.

Agentic AI has made one-to-one personalisation trivial to execute. It has done nothing whatever to make that personalisation welcome. The machine knows what your customer did yesterday, what their browsing history implies, which life event is quietly approaching. The question worth asking is not how to use all of that. It is which parts the customer should feel, and which parts the customer should never realise you saw.

Most organisations have no answer. They deploy the capability they bought, measure the revenue uplift, and miss the trust loss entirely, because trust loss does not surface in a weekly dashboard. It surfaces in year-three retention curves that nobody has thought to attribute.

The Uncanny Valley, Relocated

Masahiro Mori named the uncanny valley in 1970, describing prosthetic hands and humanoid robots that unsettle us precisely because they come close to human likeness. Marketing has now built its own version, and the mechanism is identical. A recommendation that is roughly right feels helpful. A recommendation that is exactly right, about something the customer never disclosed, feels like being watched.

Traditional personalisation could not reach the valley. Historical data, rule-based logic: bought running shoes, therefore show running socks. Primitive, and transparent to anyone paying attention.

Agentic personalisation reaches it without effort. Autonomous agents synthesise browsing patterns, social signals, contextual factors, local weather, and calendar proximity to predict not only what a customer might want but when and how they will want it. The agent does not wait for a trigger; it anticipates.

The commercial case is genuinely strong, and I want to state it at full strength before dismantling anything. McKinsey puts the revenue effect of effective personalisation at ten to fifteen per cent, with outliers near twenty-five depending on sector and execution. Lifetime value rises where a brand demonstrates understanding beyond crude demographics. That evidence is sound.

The quarrel is with what the same capability does at the margin. Gartner has measured this paradox in every cycle since 2023, and the gap widens each time, because capability compounds faster than comfort does. Personalisation can now be accurate, timely, and unwelcome: all three at once, in the same message, from the same model that hit its conversion target.

A system that infers what a customer never disclosed is running surveillance, whatever the vendor’s category page calls it. Accuracy was never the constraint. Permission is.

Progressive Disclosure and the Consent Ledger

The remedy is not technical, which is why so few organisations have found it.

Start with disclosure. Rather than switching on the full capability at launch, let the customer experience increasing levels of customisation as trust accrues — the way any relationship discloses itself, at the pace the other party sets. The engine holds capability in reserve. That reserve is an asset, not waste.

Then the harder work: a governance framework that states, in writing, which data sources agents may access, which inferences they are permitted to draw, and at what point a human must sign. An agent may adjust product recommendations from observed browsing behaviour on its own authority; that is a low-stakes act with a visible basis. Inferring a health condition, a pregnancy, a redundancy, or a deterioration in financial circumstances belongs to an entirely different class, and requires consent obtained ex ante — before the inference is drawn, not after the complaint arrives.

Most consent registers cannot survive that test, because they were built for data capture rather than data reasoning. The customer ticked a box permitting you to hold their transaction history. Nobody asked whether you could deduce a divorce from it.

Castiglione called it sprezzatura: the studied effortlessness that conceals its own labour. It remains the right standard for a personalisation engine. The customer should experience ease and never machinery — and the machinery, in this case, includes the visible evidence of how much you know.

So the target is optimal personalisation rather than maximum. The two are not adjacent settings on the same dial; they are different governance decisions, taken by different people, with different accountability. Which raises the question your roadmap has been avoiding. What proportion of your current personalisation would a customer describe as anticipated rather than surveilled?

AI-Native Journeys

The linear journey map, running from awareness to advocacy, now looks quaint. Traditional mapping assumed predictable paths and predetermined sequences. Agentic AI dispenses with both.

Live orchestration replaces the map. Agents monitor time of day, device, engagement pattern, competitive activity, inventory position, and local events to construct the next best experience as it happens. A customer researching laptops on Tuesday morning gets a materially different experience from the same search on Friday evening, not through a rule but because the system reads context as intent.

This obliges marketers to stop designing paths and start designing components: modular elements that agents assemble and sequence on the fly. Four principles govern the components.

Context-aware adaptability. Each component must work regardless of when or how it is deployed. A product education module should adjust its depth to the customer’s demonstrated expertise, not to a segment label applied nine months ago.

Fluid handoffs. Context must transfer intact between agent and human. Nothing corrodes goodwill faster than repeating what you already told the machine.

Graceful degradation. Where full personalisation is not available, whether through thin data or an explicit privacy preference, the experience should fall back to something meaningful rather than something broken. This principle is quietly the most important of the four, and the one product teams cut first under deadline. Not every customer wants to be known. A journey that punishes them for it — slower, clumsier, visibly worse — converts a privacy preference into a grievance.

Emotional coherence. One agent should not manufacture urgency while another promotes relaxed browsing.

Above all, AI-native journeys need agency moments: points at which human choice explicitly overrides the recommendation. Treat them as pressure valves rather than defects in the routing logic. Without them, customers experience the journey as a tunnel.

The Hierarchy of Interaction Value

The more capable autonomous agents become, the more valuable human interaction becomes. The relationship is complementary, and most operating models have it backwards.

Transactional exchanges sit at the base. Order status, basic product information, routine troubleshooting. Agentic AI handles these at volume with perfect consistency, and there is no differentiation to be won here.

Consultative interactions sit above them: complex decisions, specific needs, expert guidance. The optimal configuration is hybrid, with agents equipping human representatives with customer history, relevant precedent, and live recommendations.

Moments of meaning sit at the apex, and they are where the operating model usually fails. A complaint handled with genuine care. A milestone acknowledged. A difficult situation met with judgement rather than policy. These remain human, and the challenge is not classifying them but detecting them in flight.

Consider a return. An agent processing it might notice patterns that suggest something larger — the customer’s third return this quarter, sentiment that runs hotter than the immediate issue warrants. Handled mechanically, the return closes and the relationship quietly ends. Handled properly, the system routes to a person with full context and the authority to address the cause. Pattern recognition finds the moment; a human resolves it. That escalation logic is worth more than another increment of model accuracy, and almost nobody is funding it.

Purposeful Friction

Efficiency is not always the right optimisation target.

A luxury brand may route its highest-value customers to human stylists, not because the agent could not produce good recommendations but because the human interaction is the positioning. The friction is the product.

Some argue that as agents improve, customers will come to prefer their consistency to the unpredictability of people. That view underrates how much of a commercial relationship is reassurance rather than information. In my experience the most data-fluent clients we serve at S&P Global — institutions that run quantitative models for a living, who could interrogate the data themselves — still want a person in the room when the decision is large. Nobody wants an algorithm validating a two million pound commitment. They want someone who understands what is at stake if it goes wrong.

Channel choice follows the same logic. Text suits quick private queries, voice carries reassurance, video matters where visual cues do. AR and VR will extend the canvas rather than replace the need.

One design rule holds across all of them: the agent should not pretend to be human. The most effective ones are authentically artificial while demonstrating personality, warmth, and reliability. Clarity is what builds the connection. Impersonation spends it.

Before the Next Roadmap Review

Two audits and one reallocation, each of which returns a number your team would rather not measure.

First, sample one hundred personalised interactions across your customer base and classify each as anticipated or surveilled. The result is rarely flattering on first measurement, which is the point; you cannot remediate a distribution you have never seen.

Second, map every personalisation lever your agents pull to the specific consent the customer gave. Where consent is implied rather than explicit, you are running a regulatory exposure. Where consent is explicit but the inference travels well beyond it, you are running a trust exposure, and that one settles more slowly and costs more.

Then the reallocation. Identify the single moment of meaning in your journey that your operating model currently treats as transactional — the renewal, the complaint resolution, the first purchase after a long lapse. Give one human team ownership of that moment with agents in support, rather than the reverse. The economics of that single change typically beat a quarter of automation work, and you will know within two renewal cycles.

By the time the year-three retention curves resolve, the brands that chose optimal over maximum will hold relationships the maximum cohort has quietly spent. Shorten the list of things your agents are allowed to know, this quarter, while the shortening is still voluntary.

Keep Reading

That’s all for this week book chapter summary, come back next Monday for the next chapter summary.

The Agentic CMO - Second Edition is available today in hardcover, paperback and ebook.

Disclaimer: The views and opinions expressed in The Agentic CMO, Chronicles of Change and on my social media accounts are my own and do not necessarily reflect the official policy or position of S&P Global.