Stripe’s chief executive has made the case that buying could soon be settled by research rather than attention. In a post on X this month, Patrick Collison argued that personal AI agents, working on behalf of individual buyers, could act as “a structural subsidy for product quality”, weakening the commercial tactics that depend on customers not looking closely.

Collison framed the post as speculation, and much of it concerns consumer markets. The part that bears on B2B SaaS marketing is the part that can already be measured. Software buyers have been handing the research stage of a purchase to AI tools for two years, and G2’s 2026 Buyer Behavior Report shows how far they have gone, and where they stop. If agents increasingly decide which vendors make a shortlist, the evidence an agent can find about a product starts to compete with the budget spent promoting it.

What did Patrick Collison actually argue?

Collison’s starting point is an imbalance of attention. Companies are well resourced, patient and organised, while individual buyers are busy and poorly informed, and a good deal of pricing is built on that gap. He gives the example of an insurance add-on offered late in a booking. Personal agents, he suggests, could make people noticeably more rational in their dealings with companies.

From there he works through the consequences. Coupon-based price discrimination may work less well if every buyer’s agent hunts for the discount. Cross-subsidies may shift too: subscribers who forget to cancel currently help pay for those who remember, and an agent does not forget. Collison says he is unsure how the overall burden would fall.

The central claim concerns quality. He asks how markets would look if every buyer spent at least ten hours researching each purchase, and argues that improving the product would become a more effective strategy than it is now. Information about quality would gain value, because agents will want such signals constantly in a way most people never do.

He leaves two questions open. Agents could strengthen superstar effects, with rational agents converging on the same few products, or they could spread demand more widely as tastes diverge. And the whole argument depends on agents acting on the buyer’s side, which Collison says he expects to win out. His examples are consumer ones; the post does not address enterprise software procurement.

The economist Tyler Cowen republished the post on Marginal Revolution on 11 October and added a less comfortable possibility. Sellers’ agents, he wrote, could collude on price in ways the law struggles to reach, possibly without the merchants intending it, leaving prices more volatile until the collusion breaks down or new entrants arrive.

How far have AI agents got in B2B software buying?

G2’s evidence suggests B2B buyers have adopted agents for research and stopped short of delegating the decision. The report, published on 22 July, is based on a survey of more than 1,000 B2B software buyers and decision-makers and interviews with more than 50 sales and marketing leaders.

According to G2, 82% of buyers sourced software recommendations from an AI chatbot in the previous 24 months, and half of them said AI made the biggest difference when narrowing and comparing options. Some 61% currently use or plan to use AI agents in the buying process, with a further 19% open to it for selected tasks. The most common uses map closely onto the research Collison describes: assessing total cost of ownership (51%), building shortlists (51%), researching solutions (49%) and evaluating shortlisted vendors (46%).

The limit is just as clear. Fewer than half of buyers, 47%, would let an agent research and recommend while humans keep final authority. Only 9% are comfortable with an agent executing a purchase within approved guardrails, and 2% would allow one to buy without pre-approval. The Marketing Helm reported the same ceiling last week, as Meta and xAI launched consumer agents built to shop.

Two caveats apply. Several of G2’s agent figures combine current use with planned use, and the survey measures what buyers say rather than what they do. Nor do Collison’s coupon and forgotten-subscription mechanics carry over neatly to software bought by committees under procurement rules. His quality argument transfers more cleanly. An agent asked to compare total cost of ownership across four vendors rewards whichever vendor makes that comparison easy and verifiable.

G2 also found review sites narrowly ahead of AI chatbots as the source shaping shortlists, at 38% against 37%, the first time review sites have come out on top. Third-party evidence about a product still weighs heavily alongside the agent doing the reading.

What could stop agents rewarding the better product?

The risks start with Collison’s own superstar question. If buyers’ agents converge on the products with the deepest trail of reviews, documentation and comparisons, demand could concentrate on established category leaders, and a better product from a newer vendor with a thinner record could lose out. Research efficiency and market concentration can rise together.

Legibility is a second risk, and a nearer one. A Siteline benchmark of 100 leading B2B websites, published in June, found that when an AI agent could not read a vendor’s pricing page, 58% of its final answer came from third-party sources, against 12% on runs without errors. Pricing loaded through JavaScript or held behind a “contact sales” button pushed the agent towards directories and competitors’ comparison pages. On that evidence, the quality an agent rewards is the quality it can verify. A strong product with hidden pricing can read worse to an agent than a weaker one with a clear price list.

The third risk is influence. Collison’s case assumes agents serve buyers, but the market is already testing ways to reach them directly. Time has sold advertising that only AI agents see, and Perplexity has moved to block it. Cowen’s warning about collusion points the same way: agents working for sellers can learn the market as quickly as agents working for buyers.

For marketing teams, the practical consequence is a change in what promotion has to carry. Claims an agent cannot check, such as “best-in-class” or “enterprise-grade”, count for little in a comparison an agent assembles from specifications, prices and reviews. Published pricing, documented integrations, stated limitations and current review profiles become the material an agent works from.

Collison described his view as conditional on agents working for the buyer. In G2’s survey, 2% of B2B software buyers would currently let one make the purchase without asking first.