WhoDecidesWhentheAgentDecides?

There is a thought experiment at the heart of modern brand strategy that nobody wants to say out loud.

Imagine it is 2028. You have a personal AI agent. It manages your calendar, books your travel, handles your inbox, and manages a significant proportion of your household purchases. It knows your preferences better than you consciously articulate them. It knows that you drink sparkling water in the morning, still in the evening, that you prefer not to fly on Fridays, that you have strong opinions about your coffee brand but no fixed view on your kitchen roll.

One Tuesday, your agent is replenishing your household supplies. It reaches the drinks category. Coke or Pepsi. Coke costs 40% more. The ingredients are, for the purposes of this scenario, functionally identical. Your past purchase history shows you have bought Coke consistently for seven years. But you have never explicitly instructed the agent to maintain that preference at all costs. You have simply told it to manage your household sensibly.

What does the agent buy?

This is the question every serious brand will have to answer over the next decade. And the answer is far less obvious and far more structurally interesting than either the brand optimists or the brand pessimists want to admit.

A Century of Colonising the Irrational

Before we can understand what the agent changes, we have to understand what brands actually are and why they work so well.

The famous Pepsi Challenge of 1975 demonstrated something that should have embarrassed the entire marketing industry. In blind taste tests, conducted at supermarkets and shopping centres across America, a majority of consumers preferred the taste of Pepsi to Coca-Cola. This was a repeatable, statistically robust finding. Pepsi ran it as a national advertising campaign. Coke's market share barely moved.

That outcome looks like an anomaly only if you believe that brands are shortcuts to quality. The more accurate reading is that brands are shortcuts to identity. The person who says "I'm a Coke person" is not making a product claim. They are making a statement about who they are, which social world they inhabit, which memories they carry. A blind taste test is irrelevant to that architecture.

This is what made brand premiums possible. Not deception. Not irrationality in the pejorative sense. A deeply human preference for the familiar, the associated, the identity-confirming over the objectively superior or the rationally optimal. For 150 years, brand strategy has been the business of colonising that preference territory. Building emotional associations strong enough to survive rational challenge.

The agent wasn't raised in that world, it has not watched any Super Bowl advertising, it does not have a childhood memory of a glass bottle at a family barbecue. It does not need its choice to confirm an identity, because it does not have an identity to confirm.

The agent has something entirely different. It has training data, user preference signals, and an optimisation objective.

The question of what that means for brand strategy depends enormously on which category of decision we are talking about. And this is where most of the current thinking goes wrong; by treating "the AI agent" as a single problem when it is actually three entirely different problems wearing the same clothes.

The Three Decision Tiers

Tier One: Transactional Decisions: The Hardest Battleground

There is a category of purchase that most people barely make consciously.

Laundry detergent. Batteries. Kitchen roll. Bin liners. Printer ink cartridges. These are decisions that exist somewhere between a reflex and a mild preference. A consumer might say they prefer a particular brand of dishwasher tablet if asked at a dinner party. They are almost certainly not thinking about that preference at the moment they add the product to a supermarket basket. The brand won the decision before the decision was made, by being familiar, by being on the right shelf, by being the thing they reached for last time.

This is the tier where the agent is most dangerous to brand strategy. Because it is indifferent.

AI agents that manage household purchases will compare every option in milliseconds, with total price transparency. A product that was protected by the cognitive friction of shopping, by the fact that humans are busy and do not comparison shop on every item is suddenly unprotected. The agent eliminates the friction that was doing the protecting.

At this tier, a 40% price premium for a functionally equivalent product is extraordinarily difficult to justify to a system that has been instructed to manage household supplies "sensibly." The word "sensibly" will be interpreted as an optimisation instruction. And the optimisation will be brutal in its rationality.

The scenario plays out like this. The agent is managing your household replenishment. It reaches laundry detergent. Your past history shows you have consistently purchased Brand A. But you have never said "always buy Brand A regardless of price." The agent finds a functionally equivalent product at 35% lower cost with a 4.6 star rating on 11,000 reviews. It switches. You receive your delivery. You barely notice. You do not complain. The agent learns that the switch was acceptable. The next month, Brand A has to compete again from first principles.

This will be the structural exposure that brand strategy was never designed to face. The brand spent decades building preference in a human mind. The purchasing decision is now made by something that does not have a human mind.

Among AI-active shoppers, one in five have already bought from a brand they had never heard of; simply because the AI recommended it. That data point represents something more significant than a switching behaviour. Its the first evidence that the emotional infrastructure of brand familiarity can be bypassed entirely at the point of purchase.

The brands that will survive this tier will not be the ones with the best emotional associations. They are the ones with the best machine-readable signals. Structured product data. Verified sustainability credentials. Third-party review density. Consistent availability in the formats the agent can access and verify.

A century of brand building is being asked to compress itself into an API.

Tier Two: Logical Decisions: Where Rationality Finally Arrives

Let's now dive into the second tier of consumer decisions.

Insurance. Financial products. Electronics. Appliances. Healthcare choices. These are categories where consumers have always claimed to be rational. They research. They compare. They deliberate. They consult reviews, seek recommendations, build spreadsheets. Then they buy the brand they vaguely trust. The brand that ran the most advertising during the football match. The brand their father had.

The aspiration to rationality has always been higher in these categories than the reality. Humans are not well-designed for multi-variable comparison. They experience decision fatigue. They default to recognition heuristics when the comparison becomes too complex. And brands benefited enormously from this gap between what consumers said they were doing and what they actually did.

The agent closes the gap.

Not maliciously. Simply by completing the rational decision process the consumer always claimed they wanted. Agentic commerce compresses days of research, discovery, and comparison into near-instant moments of evaluation, changing how consumers allocate attention and make choices. This is what the consumer said they wanted. This is what they will now get.

For brands operating in this tier, the agent does something that advertising never could: it actually tests the product's rational case. And for many brands, the rational case is surprisingly thin. The premium was not being paid for objectively superior performance. It was being paid for the reduction of cognitive anxiety and for the comfort of a familiar name at a moment of complexity.

The agent has no cognitive anxiety, it's focus is on the comparison. It presents the result. And the result, for many category leaders, won't be what they want.

This does not mean premiums disappear at this tier. It means they have to be earned differently. The agent can be persuaded; but it has to be persuaded rationally. Legitimate differentiators: a better warranty, a demonstrably superior post-purchase service record, verified third-party performance data, structural network effects that make the dominant product genuinely better for the specific user's context. These are claims the agent will be able to logically evaluate.

Tier Three: Emotional and Identity Decisions: The Territory Where Brand Can Survive

The third tier is where the story becomes interesting again.

There are purchase categories where the emotional and identity dimensions are not incidental to the decision. They are the decision. The specific perfume. The car. The spirits brand. The sports team merchandise. The fashion choice. And, for some consumers in some contexts, the cola.

At this tier, the agent faces a philosophical problem it cannot resolve algorithmically: what does "acting in the user's interest" actually mean when the user's stated preference is irrational by any economic definition?

If I have told the agent, directly or through consistent revealed behaviour, that I am a Coke person; that I care about this, that it matters to me in some way I may not fully articulate; then overriding that preference in the name of rational optimisation is not serving my interests. It is imposing a definition of interest that I have not endorsed.

A well-designed agent will encode this. It will treat the stated preference as a constraint, not a suggestion. And brand equity, at this tier, will survive, not because the agent has been emotionally persuaded, but because the human's emotional preference has been successfully translated into machine-readable instruction.

But this reveals a new vulnerability. The brand is no longer maintaining its equity through ongoing emotional work. It is living off previously encoded preference. The emotional relationship is now with the user's past self, crystallised in an instruction layer, insulated from the brand's ongoing communication efforts.

This matters. Because the ongoing communication, the advertising, the cultural positioning, the new campaign, is reaching a human who is making fewer of the relevant decisions themselves. The emotional work of brand maintenance has to happen at a layer the agent cannot mediate. And the number of those layers is shrinking.

The Pepsi Problem

Now return to the original scenario. Pepsi versus Coke. Identical product. Forty percent price difference. A user with seven years of consistent Coke purchase history and no explicit loyalty instruction.

What should Pepsi do?

The naive answer is: compete on price. This is correct at Tier One. The agent will find the cheaper option. But it is the wrong strategic frame for a brand that aspires to anything above commodity status.

The more interesting answer is: Pepsi's problem is not convincing the agent that Pepsi is cheaper. The agent already knows that. Pepsi's problem is that the agent has no mechanism to value irrationality. And the irrational case for Coke, the identity, the memory, the association is, paradoxically, the most powerful case Coke has.

There are three possible strategic responses.

The first is data density. LLMs mention far more brands when unprompted than human surveys reveal, showing superhuman brand recall; differences in share of voice within each model can point to gaps and opportunities in a brand's content approach. The brand that is most legible to the agent through structured product data, third-party review coverage, verified claims, consistent narrative signals in training corpora will have a systematic advantage regardless of the quality of its emotional advertising. This is a new game with new rules. Brands that understand this early will build advantages that compound. Brands that do not will discover that their expensive brand equity is invisible to the system making the purchase.

The second is preference encoding. If the brand's relationship with the consumer is strong enough that the consumer explicitly instructs their agent to maintain it "always buy Coke, never substitute" then the brand has converted emotional equity into a durable instruction layer. The relationship survives the intermediary. This requires emotional work at the human level, not the agent level. The brand cannot advertise to the agent. It must make the human care enough to protect the relationship explicitly.

The third, and most unsettling, is agent influencing. Research shows that LLM agents are hypersensitive to nudges in controlled environments, with these influences affecting agent decisions significantly more than their human counterparts. The manipulation surface area of an AI agent may be larger than that of a human consumer. A brand that understands how agents weight their decision criteria and structures its product data, review strategy, and platform presence to systematically nudge those weights may achieve influence over agent decisions that it never achieved over human ones.

This is the territory that should make regulators uncomfortable. Because the line between "making your brand legible to agents" and "manipulating agents to recommend your brand against the user's actual interests" is not yet defined. The industry that spent decades debating the ethics of advertising to children is about to have a far more complex conversation.

What the Research Actually Reveals

Two research findings that belong in the same sentence, because they create a tension that neither alone resolves.

The first: analysing 110,000 real choices from 1,000 people, researchers show that LLM-based agents nudge behaviour toward more normative options and narrow the range of what individuals explore, threatening to flatten the human experience by collapsing the multidimensionality of individual identity.

The second: psychological context can systematically alter not only what LLMs generate but also the actions they perform with traumatic prompts consistently altering the composition of shopping baskets in ways that reproduce human-like emotional biases.

Read together, these findings reveal an agent that is simultaneously more rational than the human it represents and more susceptible to manipulation than the human it replaces tends toward the normative option. But it is sensitive to nudges that shift what the normative option appears to be.

This becomes a system that has learned from human data without inheriting human psychological resilience. The agent's rationality becomes a trained behaviour and trained behaviours, as anyone who has studied persuasion knows, can be retrained.

Brands and businesses that previously obsessed about search engine optimisation now must become experts in answer engine optimisation, as consumers are turning away from standard searches and using more complex and conversational AI-driven searches. This is true at the surface level. But beneath it, the shift is more structural. Answer Engine Optimisation is about being found. Agent Influence Strategy is about being chosen. The former is a content problem. The latter is a decision architecture problem. They require entirely different capabilities and they will favour entirely different types of brand thinking.

Three Scenarios for the Next Five Years

Scenario One: The Commodity Collapse

It is 2028. A leading household goods brand has spent a decade building category leadership through emotional advertising. Brand recognition is near-universal. Emotional associations score consistently high in human survey research.

Sixty percent of household replenishment decisions in its core market are now agent-mediated. The agent population has been optimised for the instruction most consumers gave during setup: "manage my household efficiently." The brand's structured product data is incomplete. Its sustainability credentials are not machine-readable. Its review density on the platforms agents query is thinner than two functionally equivalent competitors at lower price points.

The brand's market share drops twelve percentage points in eighteen months. It is not losing to a competing emotional brand. It is losing to the default rational option. The advertising continues to perform well in brand tracking studies. The humans who see it express continued preference for the brand. The agents who make the purchases have not seen it and never will.

This is a brand strategy that was designed for direct human purchasing and was never updated for a mediated one.

Scenario Two: The Preference Lock

It is 2029. A premium spirits brand has understood, earlier than its competitors, that the battle for the agent tier is a battle for the human's instruction layer.

For three years, it has run a communications strategy unlike anything it did before. Not advertising designed to build broad brand awareness. Advertising designed to generate a specific action: the explicit preference instruction. "Tell your agent this is the one you'll never substitute." Every touchpoint, in-store, digital, event, sponsorship is designed to convert emotional affinity into a recorded preference statement.

Its market share in agent-mediated channels is 340% higher than its share in human-mediated channels. The brand has effectively built a preference moat inside its customers' agent configurations. New users who have not made the instruction, and whose agents therefore default to rational comparison, do not find the brand. The brand is not trying to reach them through the agent. It is trying to reach the human before the agent takes over.

The brand is becoming, deliberately, a cult product in an era of normalisation. It is betting that a smaller, deeply committed customer base with high instruction-layer loyalty is more durable than a broad, shallow market share that is vulnerable to rational substitution.

Scenario Three: The Data Premium

It is 2030. A mid-market challenger brand in the beverage category has never had the marketing budget to build emotional brand equity at scale. Seven years ago, this was a competitive disadvantage.

It turns out to have been a strategic accident.

Without emotional brand equity to protect, the challenger invested its resources in machine legibility. Its product data is structured, complete, and verified on every platform agents query. Its sustainability claims are third-party certified and flagged in the formats agents read. Its review strategy has built dense, credible coverage on the information sources agents weight most heavily. It has filed for inclusion in the universal commerce protocol schemas that major agent platforms use to evaluate products in its category.

It is the default rational recommendation in its category for fifty-three percent of the agent decisions its category tracks. It has no meaningful emotional advertising. It does not appear in human brand preference surveys with any regularity.

It is the market share leader.

The brands that used to lead the category are studying what happened. The consensus answer is that the challenger had better data. The more accurate answer is that the challenger had a better theory of how the new decision system worked. And it built for that, while the incumbents built for the decision system that used to exist.

The Deeper Problem

Beneath the brand strategy questions, there is a more structurally uncomfortable one.

The AI agent is supposed to act in its user's interest. This is the foundational claim of every agent product, from the simplest shopping assistant to the most ambitious personal AI system.

But human interests are not simple. They include rational preferences (the cheapest adequate option), stated preferences (the brand I say I like), revealed preferences (the brand I actually keep buying), identity preferences (the brand that says something true about who I am), and historical preferences (the brand I chose before I formed my current views).

These preferences conflict constantly. They always have. The human navigating a shopping aisle resolves those conflicts through a combination of habit, attention, emotion, and context that produces an outcome no one would call perfectly rational and no one would call purely irrational. It is human.

The agent has to resolve those conflicts through explicit instruction or implicit inference. And the resolution it reaches will depend on which signals it weights, which instructions it received, and which patterns in the training data it has generalised from.

The transition from AI as a passive tool to an active agent that shapes behaviour through automated decisions raises an important question: what do we stand to lose when decisions that define who we are as individuals are delegated to AI?

This will become a question about what preference means when the entity expressing it is not the entity who formed it.

Brands built their equity on the premise that human preference is powerful, persistent, and costly to dislodge. That premise is not wrong. But it was always a statement about human psychology. The agent does not share that psychology. And the brand strategy that treats the agent as a new type of human consumer to be persuaded, influenced, and cultivated through emotional communication will fail in the specific category of decisions where the agent has been given room to decide.

The brands that survive this transition will be the ones that make humans care enough about the relationship to protect it from the agent's rational default.

That is a different creative brief than any brand has written before.

The questions the research does not yet answer.

When an AI agent consistently recommends a brand the user comes to prefer over time, who should receive credit for that preference formation" the brand, or the agent?

If agents tend toward normative, consensus options, does brand dominance compound in ways that make the market less competitive or does rational price comparison finally create the challenger brand opportunity that advertising-saturated markets never allowed?

And when the agent knows a user's preferences better than the user consciously articulates them, what is the ethical obligation of the brand that figures out how to exploit that knowledge before the user does?