AI agents in ecommerce: where they actually earn their place

Every ecommerce team has now been pitched an AI agent. Most of the pitches describe the same thing: a chat box that answers questions. That is the least interesting version, and it is the one most likely to end up switched off within a quarter.

The useful question is narrower. Which decisions in your store are made repeatedly, from information you already have, where being wrong is cheap to correct? Those are the ones an agent should own. Everything else is either not worth automating or too expensive to get wrong.

What passes the test

• Support triage. Reading a ticket, summarising it, tagging it and routing it. Made hundreds of times a week, from text that is right there, and a misrouted ticket costs a minute.
• Catalogue hygiene. Finding product pages with missing attributes, inconsistent sizing language or absent care instructions. Boring, endless, and exactly the shape of work that decays quietly.
• Stock and demand flags. Noticing that a size is about to run out or that a SKU is moving unusually. The agent raises it; a person decides.
• Guided discovery. Turning “something for short business trips” into five candidates. The shopper is the safety net, because they can simply disagree and keep browsing.

What fails it

• Pricing. Repetitive and well-evidenced, but a wrong price is not cheaply reversible once orders land against it.
• Refunds and goodwill. Cheap in money, expensive in trust, and the evidence is almost never in the system.
• Anything the customer will quote back at you. Sizing promises, delivery dates, compatibility claims. An agent that guesses confidently here creates a liability, not a saving.

The pattern is that agents are good at reading and bad at promising. Most failed deployments come from putting one on the promising side of that line.

Recommendations are a data problem before they are an AI problem

The reason product recommendations disappoint is rarely the model. It is that one product page lists fabric weight, care, fit and origin, and the next one says “great everyday tee”. No system reasons its way out of that.

Before adding a recommendation layer, the boring work pays more:

• Consistent attributes across a category, not per product
• A taxonomy where a category means one thing
• Stock state the recommender can actually see
• Policies written somewhere machine-readable, not only in a PDF

A recommendation is also more useful when it says why. “Pairs with the jacket in your cart” and “best of these for a small space” are answers. Four similar photos are a grid.

A recommendation engine is a mirror. Point it at a messy catalogue and it will show you the mess, faster.

Where the agent sits changes what it can do

There are two separate conversations happening under the phrase “AI in ecommerce”, and they pull in opposite directions.

The first is agents you run: support triage, merchandising, ops. You choose the data, the guardrails and the escalation path.

The second is agents you do not control, arriving from outside to read your storefront on a shopper’s behalf. You cannot give those guardrails. You can only make the store legible enough that they represent it accurately, which is a structural problem rather than a feature, and one we have written about separately.

Teams that conflate the two build a chat widget and believe they have handled both. They have handled neither.

Rolling one out without regretting it

1. Pick one repetitive decision your team already makes by hand.
2. Write down what a good outcome looks like, before building anything.
3. Give the agent read access to the data that decision needs, and nothing more.
4. Let it propose rather than act, until the proposals are boring.
5. Define the escalation path first, not after the first bad answer.
6. Expand to the neighbouring workflow only once the first is dull.

Step four is where most of the value is and where most teams skip ahead. An agent whose suggestions a human approves for a fortnight tells you exactly where its judgement is thin, at no cost. An agent that acts on day one tells you the same thing through customers.

The short version

Automate reading, not promising. Fix the catalogue before you add a recommender. Keep pricing, refunds and anything a customer can quote back to you on the human side of the line. And treat the agents you run and the agents that visit you as two separate problems, because they are.