Specoria
Menu
TR Free test

InsightsAgentic commerce

Agent orders are reaching your fraud filter: what new order data and Mastercard's agent score mean for stores

Forter says agent orders on its network grew more than fivefold in two months, and Mastercard is testing a score that flags AI-initiated payments. Why fraud rules built for human shoppers can misread an agent, and what a store can check now.

By 8 min read
A torn paper browser window with a keyhole cutout sits on a stone arch, with plaster hand fragments and folded parcels beside it and a small violet key above.
Every order walks through the same old gate, and the gatekeeper has never met a helper like this. Illustration: Specoria · generated with AI

Most of the conversation about shopping agents happens before the order: whether an agent can find your product, read your price and get through your checkout. There is a step after that which gets less attention. Once an agent submits an order, it passes through the same fraud screening as every other order, and that screening was built around how people shop.

Two pieces of news from the past three weeks make this step worth a closer look. Forter, a fraud prevention company, published order data showing that agent orders on its network grew more than fivefold between early August and early October. And on 30 September, Mastercard announced a score that estimates whether a transaction was started by an AI agent, so that banks can approve legitimate agent purchases instead of treating them as suspicious.

Neither development tells a store exactly what to do. Together they show that agent orders are no longer hypothetical for fraud teams, and that the industry expects existing rules to misread some of them.

What the order data shows

Forter's analysis covers agent orders across its merchant network from 1 August to 7 October 2026. The main points:

  • Volume is rising quickly. Agent orders reached 5.3 times their early August level by the end of the period. Agent orders grew 24% from August to September, or 48% when two agents first detected on 1 September are included.
  • The mix changed in weeks. Through August, OpenAI accounted for 82% to 93% of agent orders each week. After Forter began detecting orders from Meta's Muse agent on 21 September, Muse placed 46% of agent orders in the last week of September.
  • Agents that use the website directly now dominate. Forter calls agents that read and fill in ordinary web pages "screen-scraping agents". Their share rose from 1% of agent orders in early August to 65% in the most recent week. Muse bought from 74% of the stores that received any agent order; Forter's explanation is that such agents can shop anywhere, while protocol-based checkouts only reach stores that have integrated them.
  • The baskets are not small. The average agent order was $223, and only 27% of agent orders were $50 or less.

Forter is careful about the limits. The figures come from its own network, detection dates mark when it could reliably recognise each agent rather than when the agent started buying, and orders are concentrated in North America. An Opus Research analysis of the same data adds that the numbers do not establish agent commerce's share of all online shopping or prove that the growth will continue.

What the data does show is that agent orders are arriving at stores that never integrated with any agent, through the same checkout pages people use. That means they are being scored by the same fraud rules.

Why a fraud filter can misread an agent

Two plaster hands hold a sealed paper envelope between stone pillars, with receipts and price tags on a table and a violet dot on the seal.
A helper can carry the order, but someone must still check the seal before the goods leave the shelf. Illustration: Specoria · generated with AI

Fraud tools rely heavily on signals about the person behind an order: the device they use, how they move through the site, how long they spend on each step, whether their behaviour matches their history. When an agent places the order, many of those signals change or disappear.

Riskified and HUMAN Security put it directly when they announced a partnership in August 2025: rules-based fraud management "can fail when an AI agent transacts, removing key behavioral signals", which can lead to "more false declines or undetected fraud". The risk runs in both directions. A rule tuned to catch bots may block a real customer's agent, and a rule that waves agents through may let automated abuse in.

Mastercard's announcement gives an example of the first problem. A person asks an assistant to plan a trip, and the assistant books a flight, a hotel and transport with a single approval. Several purchases at different merchants in quick succession "may appear unusual on the surface", in Mastercard's words, even though the activity is legitimate.

For a store, the practical consequence is that some agent orders may be declined or sent to manual review without anyone noticing a pattern. A declined agent order rarely produces a complaint the way a declined human order does; the agent may simply try another store.

What the networks and vendors are adding

Several efforts aim to give fraud systems better information about agent orders instead of asking them to guess.

  • Mastercard's probability score. The first of Mastercard's new Agent Pay services, rolling out for testing in the United States, is a score indicating how likely it is that a transaction was initiated by an AI agent. Mastercard says the aim is "giving issuers the confidence to approve legitimate agent-led purchases", and that it will add signals about behaviour, merchant risk, transaction patterns and credential risk over time. Mastercard is also working with Cloudflare on combining web and payment signals, and with Skyfire on "Know Your Agent" identity checks.
  • Signed approvals. Google's Agent Payments Protocol uses signed mandates that record what the person asked for and approved. Opus Research points to these as the kind of evidence a store can keep with an order.
  • Agent identity proposals. Forter has proposed an open Trusted Agentic Commerce Protocol that verifies an agent and its link to the consumer behind it using cryptography rather than easily faked signals such as user agent strings and shared IP addresses. It remains a proposal that Forter is inviting merchants and developers to shape.

These are early. The Mastercard score is in testing in one country, and none of the sources above publish approval rates, decline rates or fraud rates for agent orders. Anyone claiming to know how often agent orders are wrongly declined today is going beyond the published evidence.

What a store can check now

None of these steps depends on a new standard, and most can be done with the tools a store already has.

  1. Look at what your fraud tool did with recent unusual orders. Review declined and manually reviewed orders from the past few weeks. Orders that completed checkout very quickly, came from cloud or data centre IP ranges, or arrived with a referrer from an AI assistant are worth a second look. You are not trying to prove they were agents, only to see whether a rule is blocking a pattern you would want to accept.
  2. Ask your payment or fraud provider a direct question. Does it recognise orders from AI agents, and does it treat signed or registered agents differently from anonymous automation? Does it support the card networks' agent programmes, such as Mastercard Agent Pay or Visa's Trusted Agent Protocol? If you are on Shopify, WooCommerce, ikas, Ticimax, IdeaSoft or T-Soft, the answer often depends on the payment provider connected to your store rather than the platform itself.
  3. Avoid blanket rules. A rule that declines every order without mouse movement, or every order from a data centre, will increasingly catch real customers. A rule that trusts every self-declared agent is just as risky, since anyone can claim to be one. We covered how agents identify themselves in Bot or buyer?.
  4. Keep the evidence with the order. Opus Research recommends asking agent providers what proof of authorisation they can supply and storing any verifiable approval alongside the order. If a dispute follows, that record matters; we explained why in Who pays when a shopping agent gets it wrong?.
  5. Separate buying from changing. The same analysis notes that authority to buy should not automatically extend to changing the delivery address, accepting substitutes or negotiating refunds. For high risk purchases, such as non-refundable bookings or high value items, ask the customer to confirm directly when the agent cannot show comparable evidence.
  6. Make the honest path easy. Clear prices, shipping costs shown early and a guest checkout reduce the number of odd-looking retries an agent makes on the way to an order, and fewer retries mean fewer patterns that resemble abuse. Our field notes on where agents drop the cart list the most common sticking points.

The bottom line

Agent orders are now a measurable stream on at least one large fraud network, most of them come through ordinary checkout pages, and the card networks are building signals because they expect existing fraud rules to misjudge some of them. The published data does not yet say how many good agent orders are being declined, so the sensible move is not to rewrite your fraud rules but to check what they do with orders that do not look human, and to ask your provider how it plans to tell a customer's agent from a bot.

Specoria looks at the steps before the fraud check: whether a real browser agent can reach your checkout, where it stops, and which friction makes its path look erratic. Start with a free readiness test.

Sources