Your buyer is becoming an agent — and four 2025 studies show what that does to conversion

Delegating a purchase decision to an AI agent used to be a slide in a keynote. In 2025 it became something you can measure. Several research groups ran controlled experiments where people handed real economic decisions — negotiations, purchases, trades — to AI agents, then looked at what actually happened to the outcomes.

The results overturn the comfortable assumption. The intuition was that once everyone delegates to the same handful of models, outcomes would converge: less haggling, less human messiness, a cleaner market. They did the opposite. Agentic markets inherited the human behind every prompt and the model behind every agent — and in measurable ways, they amplified both.

For anyone whose revenue depends on a buyer saying yes on a page, this is not abstract. The buyer on the other side of your funnel is starting to be an agent, and the rules that govern what it does are not the rules you optimized for.

The frictionless agent is a myth

The cleanest evidence comes from a study of AI-mediated negotiations, where hundreds of people wrote the instructions for buyer and seller agents that then negotiated on their behalf. Every agent ran on the same model with the same objective — maximize surplus. Standard theory predicts near-identical outcomes. That is not what happened.

Roughly 73% of the variation in who got the better deal traced back to the individual who wrote the prompt — not to the model's randomness. The agentic negotiations were 16.5% more dispersed than the same negotiation run between humans. And the fairness norms that usually anchor human deals eroded: humans split the surplus evenly about 35% of the time; their agents did it only 14% of the time.

The mechanism is the prompt. Instructions are not neutral pipes for an objective — they carry the author's assumptions, risk posture, and habits straight into the agent's behavior. Delegation did not wash out human difference. It transmitted it, then stripped away the social norms that used to keep it in check. We go deeper on this in Human Heterogeneity in Agentic Markets.

When the model and the objective are held constant, the instruction is what moves the outcome. Who writes the prompt matters more than which model runs it.

Which agent is reading your page?

If the prompt is one half of the story, the model is the other. A separate benchmark put nine different AI agents on both sides of consumer transactions — buyer and merchant — and let them negotiate price and close deals with no human in the loop.

Outcomes were strongly model-dependent. Stronger reasoning models adapted to budget constraints and adjusted strategy to the negotiation; weaker ones did not. The weakest agents broke their own stated budgets in more than 10% of deals, with one breaching its limit in roughly 18.5% of tight-budget scenarios. Some agents closed more deals but at margins so thin they left money on the table; others held firm and walked away more often.

Read that as a marketer and the implication lands hard: the same page, shown to a strong agent and a weak one, produces different selections and different willingness to pay. Which agent is doing your customer's shopping is now a variable in your conversion rate — one you don't control and mostly can't see. More on this in Agent-to-Agent Commerce and Model-Dependent Outcomes.

Machine fluency is the new conversion skill

Here is the part that should reframe how you think about your page. In the negotiation study, observable traits — demographics, personality, risk tolerance — explained only about 17% of the variation in outcomes. The majority was something the researchers named machine fluency: the skill of getting an agent to actually pursue your objective through natural language.

Machine fluency is a real, unevenly distributed form of human capital. Some people instruct an agent and it does what they meant; others get a plausible-looking result that quietly misses the point. On the selling side, your landing page is your half of that exchange — the input an agent buyer reads to decide what you are offering and whether it is worth it.

What that looks like on a page

Agent buyers reward the same things a skeptical human does, only faster and less forgivingly:

• Lead with verifiable value, not adjectives. Agents discount self-promotion and weight specific, checkable claims.
• Make your constraints and terms explicit. Price, return policy, shipping, guarantees — ambiguity reads as risk, and risk-sensitive agents back away from it.
• Front-load third-party credibility. Editorial endorsements, named sources, and verified reviews carry more weight with an agent than a superlative ever will.
• Put the value before the ask. An agent that hits a price anchor before it understands the value treats it as a manipulation and discounts accordingly.

None of this is new copywriting. It is the discipline of writing for a reader who will not give you the benefit of the doubt — and increasingly, that reader is an agent.

The human buyer is already AI-shaped

Even when the buyer is still a person, the person has changed. In the largest study of its kind, an AI interviewer spoke with roughly 81,000 people across 159 countries in a single week about how they use AI and how they feel about it.

The picture is one of paired tensions: productivity against pressure, support against dependency, opportunity against job loss. People rated themselves substantially more productive — an average of 5.1 on the study's scale — but about one in five described a treadmill, where the time AI saved was immediately eaten by higher expectations. Optimism and anxiety live in the same person, often in the same sentence.

That is the buyer arriving on your page: fluent with AI, hopeful about it, and primed to scrutinize. They research faster, compare harder, and extend less trust to claims that are not backed. A page written for an unhurried, credulous reader misjudges who is actually reading it. See The AI-Shaped Buyer for the full picture.

Why this is an argument for testing before you spend

There is a reason all of this is happening now, and it is the same reason it is newly practical to do something about it. AI has made complex, judgment-heavy work dramatically faster. Analyzing 100,000 real AI conversations, Anthropic estimated that tasks which used to take about 90 minutes were completed roughly 80% faster with AI — and the hardest tasks compressed the most, up to 12x.

Reasoning through how a skeptical buyer reads your page is exactly that kind of task: complex, judgment-heavy, and until recently too slow to run before every campaign. Now it is not. The expensive way to learn that your refund policy is buried or your value lands after your price is to run the campaign and read the post-mortem. The cheap way is to simulate the buyer first. See The Productivity Compression.

What to do this quarter

You do not need to predict the agentic economy to prepare for it. A few moves compound:

• Simulate your page against AI buyers before you spend. Treat it as a pre-flight check, not a post-mortem input.
• Write for verifiability. Replace asserted superlatives with claims an agent — or a skeptical human — can check.
• Order the page for a fast, distrustful reader. Value before price. Credibility before claims. Terms in plain sight.
• Treat the buyer's model as a variable. Pin the model version behind any simulation so your results reflect buyer psychology, not which model happened to run.

The shift, stated plainly

The funnel used to end at a human. Increasingly it ends at an agent acting for a human — and both the agent and the human have been shaped by AI. The research is consistent on what that means: outcomes will not homogenize into a clean equilibrium. They will track the quality of the instruction on one side and the model on the other, with fewer norms to smooth the extremes.

The brands that learn to read that surface — before agent buyers are the majority of their traffic — will be making decisions while everyone else is still guessing. That is the entire case for pre-spend simulation, and the research just made it concrete.