Six months. That is roughly the distance between a feature announced as the future of commerce and a feature quietly withdrawn.
It is worth holding onto while reading everything else written about AI agents doing your customers' shopping, because a great deal of it is being written with considerably more confidence than the track record supports.
What is genuinely happening
The infrastructure is real and it is being built quickly.
Google unveiled a Universal Commerce Protocol at the retail federation's January conference, co-developed with Shopify, Etsy, Wayfair, and Target, with endorsement from twenty or more payment and retail partners. Mastercard completed live agent transactions in several Asia Pacific markets. Visa moved its Trusted Agent Protocol to commercial release after piloting with over a hundred partners. American Express released a developer kit and announced purchase protection for registered agent purchases.
All three major card networks now support payments initiated by software acting for a person. That is a substantial change in plumbing and it happened in about eight months.
What is not happening
Your customers are not, in any meaningful volume, delegating purchases to software today.
Salesforce reports that 75 percent of retailers expect AI agents to be essential. That is a statement about what retailers anticipate, not about what shoppers do, and the two are frequently confused in coverage of this topic.
The direction of travel from the buy-in-chat reversal is also informative. Moving from completing the purchase in the assistant to handing off to the merchant site suggests the hard problems are not technical. They are about trust, returns, disputes, and who is responsible when software buys the wrong thing.
The finding worth more than the forecasts
Columbia Business School researchers built a simulated marketplace in 2025 to watch what happens when agents do the buying without a person making the final choice.
The interesting result was not that agents buy efficiently. It was that they carry biases of their own, that those biases differ between models, and that they are exploitable.
This deserves more attention than it gets. The assumption underneath most agentic commerce writing is that software will evaluate options rationally and pick the best one, which would be excellent news for a small business offering genuine value against a larger competitor with a bigger marketing budget.
The early evidence suggests something less tidy. These systems have preferences shaped by what they were trained on and what they can parse, which means the winner is frequently whoever is easiest to process rather than whoever is best.
The part that applies to you
Here is the one thing in this subject that transfers to an ordinary business, including a service business that will never sell anything through an agent.
When an agent encounters unclear or inconsistent information, it does not ask. It skips.
If your delivery window is vague, your pricing is absent, your service area is stated differently in two places, or your terms require a phone call to understand, a system comparing you against alternatives moves on. No human being ever sees that you were considered and dropped.
This is a different failure from ranking badly in a search result. A poor ranking means somebody might still scroll. Being skipped for ambiguity means the comparison happened and concluded without you, silently, and nothing in your analytics will record it.
What legibility actually means
Not schema markup, though that helps. Something more basic.
State what you sell in words a person would use rather than words your industry uses. State where you work, specifically, rather than referring to the local area. State what things cost, or the range, or at minimum what determines the price. State how long things take. State what is included and what is not.
Then make those statements identical everywhere they appear. A business listing hours one way on its profile and another way on its site has not made an error that matters to a human visitor, who will simply believe whichever one they saw last. It has created an inconsistency that a machine reading both will resolve by discarding the source it trusts less, which may be yours.
None of this is new advice. It is the same advice that made websites usable for people, arriving again with a new reason attached.
What to do this quarter, honestly
Nothing dramatic. For a first year business the agentic commerce infrastructure is not a channel yet and treating it as one would be a mistake.
What is worth doing is the legibility work, because it pays immediately in ways that have nothing to do with agents. Clear prices reduce unqualified enquiries. Specific service areas reduce wasted calls. Consistent information reduces the small friction that makes people choose somebody else without quite knowing why.
If that work also makes you legible to whatever arrives in two years, that is a genuine benefit and it costs nothing extra.
The pattern to notice
An infrastructure buildout is happening fast and the consumer behavior it assumes is arriving slowly, if at all in the form predicted. That gap is where a lot of money gets spent by businesses reacting to the buildout rather than to the behavior.
The buy button in the chat lasted six months. The protocols may last longer or they may consolidate into one, and nobody writing about this knows which.
What will still be true either way is that a business stating plainly what it does, where, and for how much does better than one that does not. That has survived every previous change in how people find things, and it is a safer thing to spend an afternoon on than any protocol.