Agent Readiness Audit: 7 Checks Before AI Shoppers Reach Your Checkout

In July 2026, an AI agent browsed a real merchant's website, picked a product, and paid with a real customer's card. Visa ran that transaction live across 31 European issuing banks, at merchants including lastminute.com, Frasers, Cleverbridge and BrickDepot. This is not a pilot in a sandbox anymore. Agent traffic is arriving at ordinary checkouts, and most merchants have no idea whether their store welcomes it, blocks it, or silently breaks in front of it.

Checkout.com's June 2026 merchant survey puts AI agents at roughly 3% of transactions today, with 89% of merchants saying they are actively preparing for agentic commerce. Both numbers matter. The volume is small, and the readiness gap is enormous. This post is the audit we run before we touch anything: seven checks that tell you where your store actually stands.

1. Can a trusted agent even reach your checkout?

Visa's Trusted Agent Protocol attaches signed HTTP headers to agent requests, verified against the Agent Directory. Cloudflare and Akamai, the same edge networks that spent a decade blocking bots, now help merchants admit verified agents. If your bot mitigation rules predate October 2025, they are probably blocking TAP-signed requests as hostile traffic. You are refusing orders from customers who verified themselves at the network level.

Check your edge logs for requests carrying signature-agent or related TAP headers, and what status code they got. If every one of them was a 403, you just found revenue you are declining.

2. Can an agent read your catalog?

Agents do not browse, they parse. Visa now ships an Agent Score tool that tells merchants whether their own site is navigable by an AI agent. Product data trapped in JavaScript, prices behind infinite-scroll, sizes that only exist after four clicks: all of it is invisible to an agent, which means your products do not exist in the agent's consideration set.

Structured data, clean product feeds, and machine-readable inventory are the entry ticket. Ask yourself the simpler question first: can an agent answer "does this store have my size in stock" without a human?

3. Do you log agent identity on every order?

When an agent-initiated order comes in, the TAP signature tells you which agent placed it and on whose behalf. Most merchants store none of this. The order lands in the OMS looking identical to a card-not-present order, and 45 days later the cardholder calls their bank with "I never authorized this."

If you cannot reconstruct, per order, that a verified agent acted on a verified authorization, you will defend that chargeback with nothing but your word. We wrote a full playbook on capturing this evidence for TAP and Mastercard Verifiable Intent transactions; the short version is that the proof exists at checkout time and it decays fast.

4. Are your policies agent-aware?

Refund, return and cancellation policies were written for humans. Agents enforce them literally, at scale, across hundreds of stores in one session. A return window stated as "30 days" means something different to an agent scheduling a return on day 29 at 11pm. Ambiguity that a human shopper absorbs becomes dispute volume when an agent acts on the strict reading.

While you are in there: decide whether agents get the same prices, the same promotions, and the same inventory as humans. If your promo stack has a loophole, agents will find it the way coupon hunters do, just faster and without a forum post.

5. Is your fraud stack ready for legitimate agents?

Here is the uncomfortable pattern from the first live agent transactions: the payment authorizes cleanly, the fraud score looks normal, and then the cardholder disputes it anyway, because agentic commerce runs on delegated authority and people forget what they delegated. First-party fraud by confusion, not by malice.

Your risk rules were tuned for humans clicking. An agent order has a different signature: perfect form fills, no browsing history, a tokenized credential, sometimes a passkey authentication from the network itself. If your rules cannot tell a delegated Visa Intelligent Commerce token from a stolen card, you will decline good orders and eat bad ones.

6. Do you know your exposure number?

You cannot manage what you have not measured. The audit question is blunt: of your last 1,000 orders, how many came through an agent, and what is the chargeback rate on that cohort versus the rest?

Today most merchants answer "unknown," because agent orders are not tagged. Visa's data suggests agents are already near 3% of transactions industry-wide, and our own read of dispute trends says agent-linked disputes run meaningfully hotter than human ones. If 3% of your volume carries 3x your average dispute rate, that is not a future problem. That is this quarter's ratio problem.

7. Do you have a written agent policy at all?

The networks moved first. Visa's Intelligent Commerce Connect (April 2026) accepts agent payments over TAP and three other protocols through one integration. Mastercard's Agent Pay binds tokens to consent policies. The rails decided that agents are customers now. Your terms of service, which still say "you must be a human," have not caught up.

An agent commerce policy is one page: which agents you accept, what they can buy, how disputes are handled when an agent acted within its mandate, and what evidence you retain. Merchants who write this down will handle agent disputes in hours. Everyone else will argue each one from scratch.

The scorecard

Seven checks, each pass or fail. In our experience auditing merchant sites this fall, the typical DTC store passes one or two. The most common failure is the second one: a store that technically accepts agent traffic but is unreadable to the agents doing the shopping, followed closely by the third: no evidence trail on orders that were placed by an agent.

The good news is that none of this requires rebuilding your stack. TAP verification is an edge configuration. Evidence capture is a checkout-time logging change. The policy page is an afternoon. The merchants who do this work now will be the ones agents shop with by default, and the ones who find agent chargebacks in six months wondering why their dispute rate tripled.

If you want the audit run against your store with concrete findings instead of a checklist, that is a service we now offer. The checklist above is the syllabus. The audit is the exam.