AI agents in ecommerce: what's actually working right now

I talk to merchants every week who are trying to figure out where AI agents fit into their operations. The question used to be “should we be looking at this?” Now it’s “we know we need to do something, but what actually works?”

That’s the right question. Because the gap between what AI agents can do in a demo and what they can do in your actual commerce stack is still pretty wide. But it’s getting narrower in specific, useful ways.

Here’s what I’m seeing work right now.

Inventory and supply chain

This is the area where AI agents are delivering the clearest returns, and it’s not close. Gartner has been tracking this, and their data shows supply chain and inventory management as the leading AI use cases in retail this year.

What does that look like in practice? A mid-market distributor I know deployed agents that monitor SKU velocity across warehouses and automatically adjust reorder points based on seasonal patterns, supplier lead times, and current pipeline status. They cut stockouts on their top 200 SKUs by roughly 30% in the first quarter.

That’s not a chatbot answering customer questions. That’s an agent making operational decisions faster than a human planner could, based on data the human planner would eventually get to but not at 2 AM on a Saturday.

Walmart has been vocal about using AI to optimize product assortments and supply chain operations. They’re ensuring high-demand items stay in stock while reducing excess inventory on slower movers. When the biggest retailer in the world bets this hard on something, it’s worth paying attention to how they’re doing it, not just that they’re doing it.

Fraud detection

Here’s one that surprised me with how concrete the numbers are. AI-based return fraud detection is reducing fraudulent returns by about 38%, according to Gartner’s analysis. For a retailer processing $50 million or more in returns annually, that translates to savings between $800,000 and $2.4 million per year.

Return fraud has been a persistent problem that merchants have mostly just accepted as a cost of doing business. AI agents that can flag suspicious patterns across thousands of transactions in real time are changing the math on that. The agent doesn’t catch every case, but it catches enough to make a real difference on the P&L.

Search and product discovery

This is the area where I think we’ll see the most movement in the next 12 months. Natural language search powered by AI agents is getting genuinely good. A buyer can type “I need mounting hardware for a 200-pound industrial shelf that ships to our Dallas warehouse by Friday” and get relevant results.

The old approach was filtering by category, then attribute, then checking availability one product page at a time. An agent that can parse intent, check inventory across locations, calculate shipping times, and return ranked results changes how buyers interact with B2B catalogs.

I’ve seen early implementations of this on Magento and BigCommerce. The results are mixed but promising. Where it works well, conversion rates on first-time buyers improve noticeably because the friction of finding the right product drops significantly.

What’s not ready yet

Fully autonomous purchasing agents, where an AI decides what to buy and places the order without human review, are still more hype than reality for most merchants. I’ve seen pilots, but the error rates in complex B2B scenarios, things like contract-specific pricing, volume tiers, and custom configurations, are still too high for most teams to feel comfortable removing the human from the loop.

And that’s fine. The practical approach is agents that do the research, prepare the recommendation, and hand it to a human buyer for approval. That’s where the real value is right now. Not replacing people, but making them faster and more accurate.

Where I’d start

If you’re running commerce operations and haven’t moved on AI agents yet, I’d start in two places: inventory optimization and search. Both have clear metrics, both have mature tooling available, and both can show results in weeks rather than months.

The mistake I see most often is trying to boil the ocean. Don’t try to agent-ify your entire operation at once. Pick one painful, measurable process and solve it. Use that win to fund the next one.

What’s your team experimenting with? I’m curious where you’re seeing the most traction, or where you’re hitting walls.