An artificial intelligence agent connected to an online store reads a shopper's order, queries the carrier's tracking and answers inside the conversation. It updates a delivery address as long as the parcel has not shipped. It writes a product page in five languages from a single source entry. It checks your competitors' prices on your own references every morning.
Everything else goes back to a person. A refund commits money, a carrier dispute gets negotiated, a trade order gets quoted. Tandem works with more than 50 companies on these topics, including ecommerce brands selling both direct and on marketplaces. Every deployment that lasted wrote that boundary down before the first line of code.
What does an ecommerce AI agent actually do?
Four workstreams carry most of the value. They share one trait. Their raw material already exists somewhere in your systems, in the order, in the product record or on a public page.
- Customer service. The agent answers order tracking, returns and product questions, across chat, email and messaging.
- Product content. It writes the page, adapts it per market and translates it, from the attributes and one reference shot.
- Marketplace monitoring. It tracks prices, stockouts and new entrants on the marketplaces where you sell.
- Visibility inside assistants. It cleans catalogue attributes so answer engines quote your products correctly.
The need grows with volume. French ecommerce federation Fevad reported, in its September 2026 release, more than 52 billion euros of online sales in the second quarter. Transactions rose 13.5 % while the average basket fell to close to 60 euros. The market therefore places more orders worth less each. Every euro collected now generates more contacts than it did two years ago.
| Workstream | What the agent produces | Typical lead time |
|---|---|---|
| Customer service | An answer sent inside the conversation, ticket closed | 1 to 6 weeks |
| Product content | A page written and translated, ready to review | 2 to 4 weeks |
| Marketplace monitoring | A daily reading and a threshold alert | 1 to 2 weeks |
| Visibility inside assistants | Normalised attributes pushed to the assistants | 3 to 8 weeks |
What the agent handles alone, what it hands back to a person
The boundary gets drawn before deployment. Redrawing it mid-project always costs more. On one side, requests whose answer already sits in the order or in the product record. On the other, anything that commits money or requires negotiation.
Refunds deserve a note. Many teams want to hand them to the agent by week two, because that is the contact reason that irritates most. Keep control. An agent that announces a refund creates an expectation nobody will honour once the business rule changes. Support platforms handle that handover very unevenly, and we compared them in our roundup of AI ticketing tools.
How a “where is my order” request gets handled end to end
It is the number one contact reason for almost every store, and the easiest to automate. The answer lives in two systems, the order and the carrier's tracking. No sensitive data enters the loop.
Quality comes from how the answer is phrased, not from the plumbing. A shopper reading “your parcel is in transit” stays unhappy. The same shopper reading an estimated date and the name of the pickup point closes the conversation. Write the replies with the service team, not with the engineering team.
Product pages, the workstream with the fastest payback
A catalogue ages badly. Attributes go missing, descriptions date, translations trail six months behind. That is exactly the material language models handle well, because the source exists and checking the output is easy.
Cdiscount put a figure on it. Its technology arm Peaksys reported in February 2025, in Journal du Net, 30 % more conversion on product pages correctly recategorised by a language model. The rollout targeted 50 million products. Recategorising touches neither price, nor imagery, nor stock. It simply puts the product back on the right branch of the tree.
We measured the same mechanic on imagery. During our audit at Moose Knuckles, an outerwear brand selling online, we costed generating colourway variants from a single studio shot. It made roughly half the catalogue's visuals addressable. The unit cost dropped from around a hundred euros in studio to a few cents, and a three-month cycle compressed into weeks.
Marketplace monitoring, the workstream nobody plugs in
Selling on a marketplace means living with your neighbour's price. No store platform ships that monitoring natively. It has nonetheless become simple to build, now that models read unstructured content.
I built a team of agents exactly for this and walked through the setup in my newsletter on scraping. The principle has three beats. One component fetches the page and reformats it, a model extracts the variables into a fixed shape, a last agent writes the report. The full set of reports cost a few cents.
A manufacturer of technical products that Tandem works with sells direct and on several marketplaces, in France and three neighbouring countries. Its price watch runs every morning and raises an alert as soon as a reference drops below a threshold. Its call-handling agents also exist in Spanish, Italian and German, because expansion arrived before the hires did.
Why your agent invents a process that does not exist
This is the most underrated risk in automated customer service, and it does not look like a failure. The agent answers fast, with confidence, and the answer is wrong. Above all it is plausible, which makes it hard to spot.
He sums the problem up with an image I reuse often. A language model behaves like a pupil in an exam who does not know the answer. It has a go anyway, just in case. On a belt return request, his agent invented an in-store return that the merchant did not offer. A customer would have travelled for nothing.
The fix fits in one rule. The agent runs off the knowledge base, and it simply has no right to answer outside that scope. A second reflex completes the first. Every action the machine cannot perform, such as triggering a refund, gets removed from that base. A procedure absent from the knowledge can no longer be promised.
Automation, workflow or agent: what actually runs in production
The word agent covers three very different realities, and the confusion costs real budget. Plain automation executes fixed rules. An AI workflow frames a process and calls a model at one or two steps. An agent pursues a goal and picks its own actions.
Most of what runs in production across ecommerce sits in that middle column, and I own that position in this post on agents. A company wants control, reliability and a return on investment. The workflow delivers all three. We unpack the distinction in our workflows versus agents comparison.
What ships natively on the store, what has to be built
The platforms have moved a long way. Shopify presented in June 2026, in its Spring 26 edition, four useful building blocks. A merchant assistant, a structured product catalogue switched on by default for eligible products, an open protocol co-built with Google, and a sales assistant sitting on the storefront. That base covers the standard case. It stops where your catalogue stops being standard.
La Redoute shows the order of magnitude within reach. Its conversational agent has been live since autumn 2024. It handles 60 % of the requests received through the mobile app messaging without human intervention, according to a Microsoft case study published in July 2025. The retailer reported around 150,000 conversations since go-live.
What the law requires of a storefront agent
One simple obligation has applied since August 2026. Article 50 of the EU AI regulation requires telling a person that they are interacting with an artificial intelligence system, unless it is obvious from the context. An explicit greeting line is enough, and it raises the resolution rate rather than lowering it.
The rest is data hygiene. The agent handles names, addresses and purchase histories, so personal data. The French Directorate General for Enterprise published a guide for retailers, updated in April 2026, listing the use cases and the legal precautions. Ask your provider where the conversations are hosted, and have the answer written into the contract.
Should you plug an AI agent into your online store?
Yes for repetitive volume, no for commercial decisions. An agent absorbing order tracking, returns and product questions gives the team back the hours it spent copying tracking numbers. The adviser keeps the job. They get back the requests where their presence changes the outcome.
The decision rests on three criteria. The volume of repetitive contacts, how clean the catalogue is, and how clearly escalation is defined. Tandem has run around 100 engagements of this kind, from field audit through to custom automation. An online bookseller we work with started with customer support alone, and a fashion brand with the product page.
To go further, our AI retail use cases cover the store and omnichannel, and our guide to the conversational agent deals with channel choice. Our retail engagements give the wider frame.
