Agentic ecommerce: what actually changes

The promise sold on stage is the store that runs itself. What exists today, in real stores, is narrower — and more useful than that: a few pieces of the operation are now executed by software that talks, queries a system and writes back into it. First line support, cart recovery, post purchase and triage are the four that already hold up. The rest is still a demo.

It is worth separating the two, because the investment decision is different for each.

What agentic means inside a store

The word became an umbrella and lost its edge. In practice, what separates an agent from a well written chatbot is the ability to make something happen in the systems: look up an order, read the carrier tracking, check stock on a variant, open a return ticket, register the lead in the CRM.

That distinction is not academic. When a customer asks where their order is, a chatbot produces a correct paragraph about delivery times. A vertical agent opens the order, reads the last scan and says where the package is. The first one postpones the problem — and postponing costs: the customer comes back, now annoyed, now to a human.

The second trait is the gate. An agent in production needs limits on irreversible actions, escalation on uncertainty and an audit trail of every tool call. It is the least glamorous part of the project and the one that decides whether it survives month one.

What already works today

First line support. This is the most mature case, for a simple reason: a handful of questions concentrate most of the volume in any store. Where is my order, what is the delivery window, does this fit my case, how do I return it, my coupon did not apply. All of them have answers derivable from a system. The channel varies by market — WhatsApp dominates in Brazil and most of Latin America, while web chat, SMS and email split the volume in the US and Europe — but the mechanics are identical, and so is the gap between answering in ten seconds and answering in two hours.

Cart recovery. The fixed three email sequence has aged badly. An agent can ask why the person stopped, check whether the item is still in stock in the chosen size, answer the specific objection about shipping cost or delivery date, and only then offer the next step. What changes is less the copy and more the fact that a conversation exists with the person who abandoned.

Post purchase and tracking. This is where most tickets are born, and it is the most automatable of all, because almost all the information already exists somewhere. Flagging a delay before the customer asks changes the nature of the contact: instead of a complaint, it becomes a status update. In ticket volume, it is usually the single highest impact item.

Triage and qualification. In high ticket, consultative or B2B stores, the agent does the first pass — understands what the person is looking for, checks availability, records it in the CRM and hands the salesperson a contact with context instead of a bare name. The full boundary between what the agent does and where the human takes over is laid out in the sales pipeline run by agents.

What is still a promise

Broad autonomy. An agent that grants discounts, approves refunds and cancels orders on its own is not a mature product, it is a risk in production. Where those actions exist, they exist with a clear ceiling and human confirmation. Anyone promising otherwise has either never run volume or is about to find out.

Buying agents at scale. The idea that consumers will delegate purchases to an assistant that browses, compares and pays is real as a direction and still small as volume. The infrastructure work — machine readable checkout and payment protocols — is being built right now, and it is the subject of agent commerce. Getting the store readable is cheap; reorganising the whole operation around it today is betting too early.

The magic conversion number. Every time someone presents a fixed percentage lift from adding AI, with no control group, what is being measured is seasonality and media spend. Without incremental measurement, the number says more about the campaign than about the agent.

What changes in the operation, not on the site

The visible change is small: a chat window. The real change happens behind it.

The queue changes composition. When the repetitive work leaves the human queue, what remains is the hard part. The team stops typing tracking numbers and starts handling exceptions, negotiations and upset customers. That is good for the result and it requires recalibrating expectations about average handle time: time per human ticket goes up, because the easy ones are gone.

Opening hours stop being a limit. Most purchases happen at night and on weekends, when the team is not there. An agent resolving order status at 11pm is not replacing anyone — it is covering a window that used to be empty.

Handoff becomes a process, not an improvisation. Passing to a human is an operation with rules: when it happens, what goes with it, who receives it, what the customer sees. A well designed handoff is the difference between saving time and creating a second conversation on top of the first.

Writing rules becomes operations work. Return policy, delivery windows by region, what can be promised, what needs approval: all of it has to be written down to become a tool precondition. Most companies discover at this moment that those rules were never written anywhere — they lived in two people’s heads.

What breaks first

In the order it usually breaks:

1. An unreadable catalogue. Products with no structured attributes, variants with no measurements, descriptions pasted as HTML from the supplier. The agent answers badly because it cannot find things, and “I could not find it” reads as incompetence. It is the number one cause of weak results, and it is worth fixing before blaming the model.
2. Partial integration. An agent that reads orders but not tracking solves half of the most common question in the store. Half an integration produces half an answer.
3. Unwritten rules. When the return policy depends on who is answering, it cannot be automated — and the agent will invent one.
4. No gates. Without a volume circuit breaker, one loop in automation becomes dozens of messages to the same customer in minutes. That is not an AI bug, it is an operations bug — and it is the failure that destroys trust fastest.

How to measure without fooling yourself

Three numbers are enough to know whether it is working, and none of them is answer quality:

• Resolution rate without a human, by subject. The overall average hides everything; split by subject, it shows exactly where a tool is missing.
• Time to first response, measured on the real channel, including nights and weekends.
• Cost per resolved conversation, compared with the human hour it replaced.

For revenue — cart recovery, repeat purchase, assisted conversion — the only honest measurement uses a control group. Without one, you are crediting the agent with sales that would have happened anyway. The same discipline applies to any process a vertical agent takes over.

Where to start

Start with volume and repetition, not with what is most visible. The hard case, the angry customer, the complex negotiation: that is precisely where AI has the least advantage and where mistakes cost the most.

The sequence that works is boring and predictable. One channel. A set of five to ten questions that dominate the queue. Catalogue, order and tracking integrations actually working. Gates and escalation defined before going live. Measurement from day one.

With that running and measured, scope grows on its own — and it grows on evidence, not opinion. Which is the opposite of how most AI projects start, which is from the slide.