American Express releases a merchant playbook for agentic commerce

The verified development

American Express released its Business Playbook for Agentic Commerce on 6 October 2026. The company says 62 percent of surveyed merchants expect AI agents shopping for customers to have a moderate or significant business impact in the next year. Its research covered 502 United States business leaders responsible for online sales or payment decisions and 2,005 United States adults who regularly buy online. American Express recommends that businesses review how AI tools describe them, keep product information current and examine the purchase journey for operational gaps.

The survey also found different comfort levels by purchase type: American Express reports greater stated comfort with agents buying everyday low-cost items than high-value goods. Survey responses describe expectations and attitudes, not observed merchant conversion or fraud performance. The playbook is guidance from a card network and issuer, not a regulation or universal technical standard. These are attributed facts from the cited source. The analysis below is independent and does not imply FixPays support for American Express agentic-commerce products or protection programmes.

Independent analysis: operational scope

For merchants, agent readiness begins with machine-readable facts that also remain clear to people. Product names, variants, price, stock, delivery, returns, subscription status and seller identity should agree across the website, feed, application programming interface and checkout. An agent that receives stale stock or an ambiguous cancellation term can create customer harm at greater speed. Assign owners to high-impact fields and record effective dates so corrections propagate consistently instead of being patched in one channel.

Purchase intent needs more precision than a general instruction to shop. The merchant should receive enough context to identify the item, quantity, price ceiling, delivery destination and any recurring commitment, while avoiding unnecessary collection of a customer’s private reasoning. Authentication, tokenisation and agent registration may provide useful signals, but none should be treated as proof that every order matches the customer’s intent. High-value, regulated or unusual purchases may require additional confirmation or human review.

Independent analysis: customer and risk controls

Checkout should reveal material changes rather than silently letting an agent accept them. Substituted products, increased prices, changed delivery dates, add-ons and trial-to-paid transitions deserve explicit handling. Receipts should show what the merchant observed, which credential was used and how the customer can seek help. A merchant should not design an agent-only cancellation route that is harder for the customer to use. Refund and complaint teams need the same order context that supported the purchase.

Fraud controls should distinguish authorised agent activity from account takeover, manipulated instructions and ordinary customer mistakes. Build rules around risk and evidence, not a blanket assumption that agent traffic is either trusted or hostile. Rate limits, device and account signals, delivery changes, merchant-category risk and product value can inform proportionate review. Keep a manual path for contested intent. Monitor false positives because aggressive controls can exclude legitimate customers or push them toward less secure workarounds.

Independent analysis: rollout and measurement

The commercial test should compare agent-assisted and conventional journeys using completion, cancellation, returns, disputes, support contacts and margin after refunds. Attribution will be difficult when agents influence discovery but a person completes payment. Merchants should define the cohort and observation method before making performance claims. They should also review terms with commerce, payment and AI providers to understand data use, liability, change notices and service continuity. A playbook can guide questions but cannot substitute for contract evidence.

Independent analysis: wider merchant implications

Search and recommendation visibility will depend on structured, truthful information rather than hidden keyword repetition. Merchants should maintain canonical product pages, accurate availability and clear policies that both people and authorised tools can interpret. They should also monitor incorrect third-party summaries and provide a correction route. Do not create false reviews, unsupported comparisons or special agent-only claims. A merchant earns reliable recommendation by making verifiable facts accessible and keeping fulfilment consistent with those facts.

Agent traffic can alter capacity planning. Automated comparison may create bursts of requests, inventory holds or repeated pricing checks without purchases. Rate limits should protect systems while allowing legitimate tools to operate. Separate browsing, quotation, reservation and payment actions so an agent cannot accidentally commit inventory too early. Monitor whether automated activity disadvantages ordinary shoppers during scarce releases. Commercial teams should account for infrastructure and support cost when judging the value of new traffic.

Competitors may announce agent partnerships quickly, but merchants should evaluate interoperability and switching. Ask whether product data, customer permission and payment tokens are portable or locked to one ecosystem. Understand which party authenticates the agent and how liability is allocated. A proprietary path may deliver useful reach yet create concentration risk. Maintain ordinary web and accessible human journeys unless a business case justifies removal, and never force customers to delegate purchasing to obtain standard terms.

Users should be able to see, stop and challenge agent activity connected to their accounts. Account settings can show authorised agents, recent actions and revocation where the merchant controls those functions. Notifications should be timely without becoming noise. A disputed purchase review should use the recorded instruction, authentication, order changes and fulfilment, not presume that registered software acted correctly. Design for household accounts and caregivers carefully because delegated authority can be legitimate but limited.

Merchants also need a policy for content generated by agents during support or ordering. Free-text instructions may contain sensitive data, abusive content or impossible promises. Validate structured fields before they reach fulfilment and avoid inserting untrusted text into receipts, shipping labels or internal commands. Train staff to recognise when an agent summary differs from the underlying order. The authoritative record should remain the confirmed transaction and contract, not a conversational paraphrase that may omit conditions.