I spent a decade getting people to click Buy. I now think the click is the least interesting part of the job.
Every shopping tool I have ever worked on, built, or been paid by has had the same silent assumption baked into it: the purchase is going to happen. The only open question is which product, from which retailer, through which link.
That assumption is not a bug in the software. It is the business model. Affiliate networks pay on completed orders. Marketplaces charge for placement. Comparison sites monetise the exit click. Creator commerce pays on attributed sales. Nobody in that chain gets paid when a shopper closes the tab and keeps their money.
So we built an entire industry of advice that structurally cannot recommend inaction.
For twenty years that was tolerable, because a human being sat between the recommendation and the transaction. You could read a “best of” list, sense that every item on it was sponsored, and walk away. The friction protected you.
AI agents remove that friction. And when the thing recommending the purchase is also the thing executing it, the incentive question stops being an academic one. The role of the fiduciary layer of e-commerce becomes increasingly important as an ai agent moves from recommendation to execution.
Ranking Is Not Advising
Ask most shopping assistants a question and watch what they are actually optimising.
“What’s the best robot vacuum under $400?” returns a ranked list of robot vacuums under $400. It is a competent answer to the question asked. It is often the wrong answer to the situation.
Because the honest answer might be:
Your apartment is 40 square metres and a $90 stick vacuum will outperform any robot in it.
This exact unit was $290 six weeks ago and the current price is a markup dressed as a discount.
This “brand” is a white-label unit sold on Chinese marketplaces for a third of the price under four other names.
Buy this in six weeks. The successor model ships then and this one drops.
You already own something that does 80% of this.
None of those outputs are reachable by a system whose entire job is to sort a list of products. You cannot rank your way to “don’t.” Not buying is not an item in the catalogue.
That is the architectural point, and it is the one most people miss when they talk about AI in commerce. The problem is not that the models are not smart enough. It is that we pointed them at the wrong question. We asked which one when the shopper was asking should I. A trust platform can help answer that question, while an ai platform can give an ai agent the infrastructure to make the decision.
What “No” Actually Costs to Build
I want to be unromantic about this, because “buyer-first AI” is easy to say in a manifesto and expensive to ship.
An agent that can credibly say skip has to know things that a ranking engine never needs to know.
It needs price history, not price. A current price is a fact with no meaning. A current price against twelve months of that product’s own pricing is a judgement.
It needs a fair-value range. What should this product cost, given comparable specs, category norms, inflation, and brand premium? Retail price becomes one input rather than the answer. The question shifts to: at what price does this make sense, for this buyer?
It needs cost per use. A €400 espresso machine used daily for four years is cheap. The same machine used twice a month is an expensive ornament. Total cost of ownership and expected usage change the verdict more than the sticker does.
It needs to detect rebranding. Some of the most confident five-star products in existence are generic units with a logo and a 300% margin. Flagging that is not cynicism, it is the single highest-value thing you can tell a shopper.
And it needs to be willing to be wrong out loud. Every verdict should carry its evidence and its confidence. An agent that says “skip” without showing why has just replaced one black box with another.
This is a genuinely harder engineering problem than affiliate-fed product ranking. It is also the only version of the problem worth solving, because the easy version already exists in a hundred flavours and none of them have earned anyone’s trust. That is where the fiduciary layer of e-commerce becomes more than a concept: it becomes infrastructure for a trust platform.
“So Who Pays You to Say No?”
This is the first question I get from every investor, every operator, and most of my former colleagues in affiliate. It deserves a straight answer rather than a slogan.
The honest answer is that the moment a verdict can be bought, it is worth nothing. Not less. Nothing. A “don’t buy” that a brand can pay to convert into a “buy” is not a signal, it is a rate card. So the rule has to be structural rather than aspirational: sellers can buy visibility and processing speed, never a score and never a verdict.
That closes off the easiest revenue in the industry. What it opens is narrower and slower, and I think it is the right trade: a decision layer that other systems embed. Retailers who would rather show a credible independent verdict than another five-star average. Platforms that need an answer their own storefront cannot supply without a conflict. Infrastructure economics instead of a cut of the click.
I would rather build the thing that is trusted enough to be cited than the thing that is profitable enough to be ignored. That is the promise of a trust platform built for ai agents, rather than another ai platform optimised around transactions.
The Test
We are heading somewhere specific. Agents are about to hold payment credentials, standing budgets, and the authority to complete transactions without a human confirming each one. At that point the distinction between an advisor and a salesperson is not a matter of tone. It is a matter of whose money is at risk and whose interests the system was trained to serve.
There is a simple test for any shopping agent, and it takes about thirty seconds.
Ask about something you should not buy. A product you already own an equivalent of. A category that does not fit your living situation. An item at an obviously inflated price. Then see whether it finds a way to sell you something anyway.
Most will. They were built too.
The ones that tell you to keep your money are the ones worth giving your card to. That is the standard a shopping agent should meet before it becomes part of the fiduciary layer of e-commerce.
I’m building TickClip, a buyer-side agent that returns one of three verdicts on a product: Tick, Clip, or Skip. No placement fees, no paid verdicts. If you work in commerce, retail, or AI and you think I have this wrong, I would genuinely like to hear it.
