Agents reward what they can read

Somewhere in your growth plan there is probably a line that depends on a customer not paying attention. A free trial that rolls into a paid plan. A coupon that nudges someone over the line. An upsell tucked into checkout. These work because most buyers are busy.

Patrick Collison, the CEO of Stripe, the payments company, says that is about to change. On October 11 the newsletter Department of Product reported his argument. Soon, many people will have a personal AI agent that researches what they buy, and he calls that “a kind of structural subsidy for product quality.” Agents don’t forget a subscription. They don’t get tempted by a coupon. So, in his words, “focusing on making a product better is, I think, going to become a more effective strategy for companies.”

The plumbing is arriving. Sierra, a customer-service AI company, launched an open standard called the Personal Agent Protocol with Meta, Stripe, Shopify, Walmart and others. It defines how a personal agent identifies itself to a business and gets the customer’s permission before acting. A draft specification was reviewed by 35 design partners, including OpenAI.

Collison is right about one thing. Tricks built on inattention will weaken, because the agent is paying attention for you. He also adds a condition: it all depends on agents staying on the consumer’s side. Fair enough.

But what an agent rewards is whatever it can read, and that overlaps with quality only where quality can be written down.

Here is why I doubt the strong version. Last July I recorded an experiment (arXiv paper 2508.02630) in which AI agents did the shopping. Researchers had AI rewrite product descriptions to be more persuasive. In 33% of the experiments, the rewrite won the product a statistically significant share of purchases, even when the agent saw only structured data. Persuasion didn’t disappear. It moved to a new audience. A separate shopping sandbox found agents also favored items in the top row or left column regardless of content.

A commenter on the same Department of Product post put the sharper version: the winners may look “boring to a human but read perfectly to a machine.” Structured data. Predictable pricing. Documented edge cases. I’d add one thing the commenter didn’t: some qualities don’t survive translation. How a product feels after six months. Whether support treats you like a person. Whether the design makes you happy to open it. An agent comparing options can’t weigh what no field describes.

That’s the part of the quality story I’d worry about as a PM. If agents become the main shoppers, a team’s quality investment may drift toward what shows up in a feed, because that’s where the reward is.

Be fair to Collison’s side. The experiment tested one-off purchases, not whether customers come back, and repeat behavior is where real quality pays. And the numbers floating around in this debate don’t settle it. Linear’s COO, Cristina Cordova, says use of the company’s agent connector grew roughly 20 times in a year while human weekly users also rose. Guillermo Rauch, the CEO of Vercel, says bots were 58.18% of traffic on his network over 30 days. Both are self-reported, and neither measures whether agents choose better products.

Collison also runs a company that profits when agent commerce grows. That doesn’t make him wrong. It does mean his claim is a prediction from someone with a stake, and the evidence I can find leans the other way for now.

There’s a practical side too. Mike Belsito, writing for Mind the Product on October 14, points out that many companies already have an agent policy without having chosen one. A bot filter, a checkout check, a security default nobody on the product team reviewed decides whether an agent gets in. If the quality subsidy is real, your own settings might be blocking it.

So I don’t know yet which way this goes, and I think that’s the honest position. Nobody has run the test that would tell us. Take a set of products that really differ in quality, control for how persuasively each is described, send agents shopping, and see what they pick. Then ask which attributes of quality fell out of the result.

If you run a product today, which of your best qualities could an agent actually read?