AI Agents for Ecommerce Ops: Where They Save Time and Where They Don't

Every ecommerce platform and app now promises an AI agent. Some save a team hours each week. Others create a new job, checking what the agent did and putting it right. The difference has less to do with the tool than with which tasks you hand over.

At User Actually, we ask two questions about any operational task before automating it. How easy is it to check the agent's work, and how much does a mistake cost?

Automate the work that's easy to check. Keep people on the work that's expensive to get wrong.

1. Listings and Catalogue Upkeep

Catalogue work is the clearest win. Agents handle the repetitive jobs around a listing well, such as filling missing attributes from supplier spreadsheets, standardising titles and tags, writing alt text and flagging products with no size chart or mismatched variants. Each output is small and quick to check. Writing the descriptions themselves needs more care, which we covered in scaling product descriptions with AI.

Keep a person on anything that touches price, stock or product claims. An agent that adds "waterproof" to a jacket that isn't creates returns and complaints.

2. Support Triage

Triage is a different job from answering customers. An agent can read each incoming ticket, tag it by type, pull up the order and draft a reply for your team to check and send. Your team spends less time sorting and looking things up, and keeps control of what goes out.

Hand over complete replies only for narrow questions you can check against your own data, such as order status. Refunds, complaints and anything involving a vulnerable customer stay with people. The customer-facing chat agent is a separate project, covered in the AI support agent that increases conversion.

3. Reporting and Monitoring

Agents are useful for the weekly chore of pulling numbers together. They can collect figures from the store, ad platforms and email tool into one summary, and flag anything outside its normal range, such as a product with a jump in returns or a campaign whose spend doubled overnight.

Treat the summary as a prompt to look, not a conclusion. An agent can misread a data source or explain a change with a confident story that isn't true. Every number in the summary should link back to where it came from.

4. Merchandising Tasks

Routine merchandising rules suit agents too. They can re-sort collections by stock and sell-through, hide sold-out variants, suggest bundles from products often bought together and draft copy for seasonal collections. Set these up as suggestions that a merchandiser approves in a batch, at least until the suggestions are reliably good.

Decisions about which products to push, how deep to discount and what the brand stands for stay with your team.

5. Guardrails and Human Review

• Read before write. Start every agent with read-only access and add write permissions one task at a time.
• Limits per run. Cap how many products, prices or tickets an agent can change in one go.
• A log of every action. You should be able to see what changed, when and why, and undo it.
• Review by risk. Spot-check low-risk output and approve high-risk output item by item.
• A named owner. Someone on the team is responsible for each agent and what it does.

6. What Not to Automate

• Reviews and testimonials. Never let an agent write, edit or post them. Fake reviews are a banned practice under the Digital Markets, Competition and Consumers Act 2024, and the CMA's fake reviews guidance expects businesses that publish reviews to take steps to prevent them.
• Prices and promotions. Changes here hit margin directly and can mislead customers if they're wrong, so a person approves them.
• Refunds and complaints. These need judgement, and a bad call damages trust fast.
• Commitments on your behalf. Anything that promises delivery dates, stock or terms to a customer or supplier.

The Bigger Point

AI agents save the most time on repetitive tasks where the output is easy to check. Start there, give each agent narrow permissions and a log, and keep people on decisions about money, trust and the law. Add the next task only once the last one has run cleanly for a few weeks.