Blog

AI shopping agents are coming to your checkout. Can you tell a real one from a bot?

Agentic commerce is the fastest-moving trend in payments: AI agents that browse and buy for real people. Why they break old bot defences, and how to let the genuine ones through.

If you run an online store, the single biggest shift heading into 2027 is not a new scam — it is a new kind of customer. AI shopping agents that browse, compare and check out on a person's behalf have gone from demo to daily traffic in under a year, and some forecasts now put the share of online transactions involving an AI agent at close to 40% by 2027. Whatever the exact number, the direction is not in doubt, and it quietly breaks an assumption your fraud stack was built on.

Why agents break old bot defences

For twenty years, the working rule was simple: automation is bad, humans are good. Catch the script, let the person through. Agentic commerce ends that. A legitimate agent acting for a real, paying customer is a script — it has no human moving a mouse, no human filling a form, no human behaviour to measure. Block it on reflex and you have turned away a genuine sale. Wave all automation through and you have opened the door to every bot that claims to be an agent. The old binary no longer sorts your traffic.

The real question moves from behaviour to provenance

This is the shift the whole industry is making, and I think it is the right one: stop asking "is this automated?" and start asking "who is this acting for, and where is it really coming from?" Behavioural signals — mouse paths, typing cadence, time on page — are losing value fast, because software now produces a convincing session with no person present. What does not disappear is the infrastructure. Every agent, honest or fake, has to connect from somewhere, and that network is a signal you can read.

What you can actually act on today

  • Read the network behind every agent request. A reputable commerce agent operates from known, stable cloud infrastructure you can recognise. A bot pretending to be one often routes through residential proxies, VPNs or scattered hosting to hide — and that mismatch is the tell.
  • Verify the declared identity against the network. Traffic that calls itself a well-known agent but arrives from an address that does not belong to that operator is impersonation, not automation. (More on that in verifying AI agents by their network.)
  • Gate the risky steps, not the browsing. Let agents browse and compare freely; ask for human confirmation only on limited-stock checkout, payment changes and high-value orders.
  • Corroborate with the payment. The card's issuing country, read from the BIN, is a cheap second check that an agent-placed order lines up.

Prepare now, not when it is 40%

The merchants who will win agentic commerce are the ones who can welcome a good agent and refuse a fake one, at the network level, without a human in the loop. Fraudex is built for exactly that decision: a live verdict on any IP's network and type, with a bot flag, plus BIN intelligence, in a single call. Start with the proxy detection API and the VPN detection API, and keep the IP checker handy to inspect an agent's address by hand.