United’s restriction on AI booking agents exposes a practical question for every online seller: does a customer’s permission to use an assistant also give that assistant permission to transact with your business? Those are separate decisions. A customer can want help, an agent can be technically capable of completing checkout, and the seller can still decline that route.
United’s November 11, 2026 announcement, United Restricts Access for Third-Party AI Agents, describes restrictions on unauthorised personal agents across its digital platforms and indirectly through travel agencies and aggregators. The commercial lesson is broader than travel: define permitted actions, connect those rules to actual access controls and preserve a usable customer journey. Policy and documentation in this analysis were checked on November 17, 2026.
- 01Separate three permissionsIdentify the agent, establish what the customer approved and decide which actions the seller allows. None substitutes for the other two.
- 02Read the operational boundaryA rule about booking does not explain every search or servicing case. Put each action and channel into the policy.
- 03Verification is one inputAuthenticated agent traffic can still encounter stock, payment, account and approval checks. Recognition does not require automatic acceptance.
- 04Measure completed journeysTrack legitimate customers reaching confirmed orders, alongside rejected actions and support problems, before expanding an agent pilot.
01 — Policy textWhat United actually changed
United’s Contract of Carriage, revised November 11, defines AI Agents broadly, including autonomous or semi-autonomous software acting with or without human supervision or approval. Its definition includes discovery and search as well as purchasing and servicing. The ticket rule’s prohibited-practices clause states: “The use of AI Agents to make, modify or service a Ticket, and/or to purchase any service or product.” A separate cancellation provision includes reservations made, modified or serviced by AI Agents.
The distinction between a definition and an operative rule matters. A definition explains which tools the document means; the particular clause explains the conduct covered there. Reading only the ticket sentence can miss other parts of the policy. Conversely, treating every activity named in a definition as an identical contractual remedy oversimplifies the document. Sellers drafting their own terms should make the action and consequence easy to find together.
United’s announcement identifies account security and unintended actions as concerns. It does not publish a detection method or establish how effectively any restriction is enforced. A policy announcement is evidence of a business decision. It is not evidence that every automated session can be distinguished from a person, nor that every agent-generated reservation will receive the same outcome. Keep those claims separate when discussing this example internally.
Write an explicit rule for each automated action and its consequence. A broad statement about AI is difficult for customers, partners and support staff to apply consistently.
02 — Research boundaryDistinguish advice from access to the seller
United’s Digital Trust Center guidance draws an additional boundary. It permits AI-assisted trip research when the tool does not access United’s website or app. It describes agents searching, collecting information or transacting on those surfaces without United’s permission as unauthorised. It also says assistive technology for customers with disabilities is unaffected. These details come from the current guidance checked on November 17, rather than an assumption that a final human click makes all earlier agent activity acceptable.
For another merchant, the comparable exercise is to map the journey. Reading a public buying guide, retrieving live availability, opening a customer account, reserving inventory and completing a payment have different consequences. The business may choose different rules for each. A furniture retailer might welcome product comparison while requiring a person to approve a delivery address change. A subscription service might allow plan discovery but reserve cancellation or renewal changes for a clearly authenticated customer flow.
Begin with the actual customer need, then identify where the assistant touches your systems. Our agentic commerce readiness checklist covers the product and checkout foundations that make this mapping useful. Do not confuse permission to receive a recommendation with permission to inspect private account information. Each additional capability should have a defined purpose and an accountable owner.
| Journey stage | Example action | Decision to document |
|---|---|---|
| Discovery | Read a product guide | Which information may agents retrieve? |
| Account access | View an existing order | How is the customer’s authority established? |
| Transaction | Reserve or purchase an item | What approval and commercial checks apply? |
| Aftercare | Change delivery or request a return | Which actions require renewed confirmation? |
03 — Three permissionsKeep identity, consent and merchant policy distinct
Agent identity answers who is making a request. Customer consent answers what that person authorised the assistant to do. Merchant policy answers which requests the business is prepared to accept. A useful implementation carries all three through the journey instead of reducing them to one “trusted” flag. The same recognised assistant can request a permitted catalogue lookup and an impermissible account change in consecutive steps.
Consider a hypothetical customer who asks an assistant to find a replacement appliance under a stated budget. That instruction might authorise comparison, but it leaves important purchase details unresolved: the exact model, installation, removal of the old unit and the delivery date. A seller that treats the initial instruction as approval of any later basket creates ambiguity. Equally, a verified agent should not inherit account privileges simply because its traffic comes from a known provider.
The policy should produce an answer at the level of an action: permit, require a customer decision, or decline. Record enough context to explain that answer later, while avoiding unnecessary collection of personal information. A support team should be able to tell whether an order failed because the agent was unrecognised, the customer had not approved the purchase, or the merchant could not fulfil it. Those are different problems with different remedies.
Who sent the request
Verify the claimed origin through an appropriate supported mechanism. A familiar name in a request is insufficient evidence by itself.
What the customer approved
Make the item, amount and material conditions understandable at the point the person authorises a purchase.
What the seller accepts
Keep eligibility, inventory, fraud checks and fulfilment constraints attached to the transaction.
04 — Operating choiceChoose where to block, allow or verify
A blanket restriction is administratively attractive when a business cannot yet support delegated actions. It gives staff a clear starting position and avoids implying that an untested path is supported. Its commercial cost must still be investigated: legitimate customers may encounter friction, and partners may need a different route. Do not assume that the only customers using assistants are attempting abuse, or that declining automation eliminates every source of abuse.
Broad access can make sense for low-impact public information, provided the service remains reliable and the information is accurate. Purchase and account actions deserve a more specific decision. Verification offers an intermediate approach: recognise supported agents, permit defined actions and preserve transaction controls. It requires ongoing ownership of integration behaviour and exceptions, rather than a one-time switch that labels all future requests acceptable.
Write the first policy around a manageable set of actions. For example, permit public product discovery, pilot basket preparation through a supported route and require a clear customer decision for purchase completion. This is an illustrative design choice for a retailer, not a recommendation to disregard another company’s restrictions. The right configuration depends on the seller’s products, customer relationships and ability to recover when an automated journey fails.
| Approach | Useful starting condition | Trade-off to measure |
|---|---|---|
| Restrict | Delegated actions are unsupported | Lost legitimate journeys and support friction |
| Allow selected actions | Public information is dependable | Service load and accuracy of responses |
| Verify and permit | A supported agent path exists | Integration upkeep and exception handling |
| Require a person | An action needs a concrete decision | Whether the handoff is understandable |
05 — Agent recognitionUse authentication without outsourcing judgement
OpenAI’s agent-access documentation describes Web Bot Auth signatures for recognising its cloud-browser requests. It advises applying a targeted policy to verified traffic while retaining normal controls for other requests. That is a mechanism for identifying a supported source of traffic. It does not establish that a particular customer has authorised a particular order or that the seller should accept every action.
A merchant evaluating such a route should ask what the verification result changes. It might prevent an inappropriate bot challenge during a permitted journey, while payment and account checks continue normally. Avoid granting broad exemptions that are unrelated to the original compatibility problem. If the failure occurs when a legitimate customer reaches a challenge, first establish that point of failure and the scope of the supported fix.
This differs from making product information available through a commerce integration. The guide to agentic commerce standards explains why discovery, checkout and payment capabilities need separate treatment. A business can support one route without having implemented every other route. Document exactly which integration is being tested, and avoid describing its existence as proof of universal compatibility with AI shopping tools.
What specific decision changes after the request is authenticated? If the answer is “everything is allowed,” the rule is broader than the evidence supports.
06 — Purchase approvalKeep the customer’s decision attached to the order
Payment capability makes the permission question more urgent. Our ChatGPT Wallet guide explains the announced customer-approved payment flow and its separation from merchant acceptance. For sellers, the important work is to preserve the relationship between the basket the customer saw, the amount approved and the order actually created. An assistant’s fluent explanation cannot compensate for a mismatch between those records.
Use a hypothetical multi-item purchase to examine the boundary. One item becomes unavailable after the customer approves the original basket. The system should not silently substitute a materially different product or delivery condition simply to complete the journey. Decide which changes invalidate the approval and how the customer returns to a meaningful decision. The same principle applies when taxes, installation fees or delivery charges change the final amount.
An understandable handoff includes the reason a person is needed and enough context to continue. Preserve the selected product and relevant choices where appropriate; do not make the customer reconstruct the whole conversation. After approval, provide an unambiguous order outcome. If the payment attempt succeeds but order creation fails, the recovery process needs an owner. These are ordinary commerce responsibilities that become more visible when an assistant spans multiple steps.
07 — Distribution channelsApply the rule beyond the main website
United’s announcement explicitly includes travel agencies and aggregators, making the distribution channel part of the story. For an online seller, a direct-site policy can leave gaps if the same products are offered through partners, embedded checkout or account-management tools. Inventory those paths before announcing that agents are either supported or prohibited. A customer should not receive contradictory explanations depending on which part of the business answers.
Work through one journey with commerce, support and integration owners together. Identify who accepts the order, who authenticates the customer and who handles a disputed change. If a partner prepares a basket but your system finalises it, establish where approval evidence is preserved. If a partner owns the whole purchase, establish how your support team obtains the information needed to help. The goal is clear responsibility at the handoff, not duplication of every partner’s records.
The same exercise should include accessibility and ordinary customer assistance. A policy aimed at autonomous purchasing should not accidentally classify every assistive browser feature as an unsupported shopper. Describe the relevant behaviour and provide an escalation route for unexpected blocks. Review actual failure reports before expanding a rule. A precise exception process is more useful than asking support staff to guess whether a customer used “AI” somewhere in the journey.
08 — Controlled rolloutRun a pilot that can answer the commercial question
Start with one product category, one supported access route and a clearly defined set of permitted actions. Establish the current customer journey first, including the point at which an order becomes confirmed. Then examine where the agent-assisted version succeeds or fails. Useful observations include abandoned handoffs, repeated approval prompts, rejected account changes, duplicate attempts and the time support needs to resolve an incomplete purchase.
Evaluate customer outcomes alongside control outcomes. A rule that rejects every request can look effective if the only metric is blocked automation. A permissive route can look successful if the only metric is traffic reaching checkout. Neither measure tells you whether legitimate customers obtained the right products under the right conditions. Compare confirmed, correct orders and the effort required to recover mistakes, while keeping the pilot’s sample size and limitations visible.
Assign an owner who can narrow or suspend the pilot when a specific failure appears, and define what evidence would justify expansion. The next step may be a better approval screen, a partner clarification or more accurate product availability, rather than a different AI policy. Our ecommerce services focus on those connected purchase journeys: dependable information, workable checkout and clear order handling. Expand supported automation when the whole path is ready to support it.
A recognised agent, an understood customer decision and a correct confirmed order should all be visible in the same journey. Measure the gaps between them.
Make the permission boundary usable
United’s move makes agent access an explicit commercial decision. Other sellers can use the same moment to define their own permitted actions, recognise supported traffic and preserve customer authority throughout checkout. The useful outcome is a policy that customers can understand and operating teams can apply, with evidence about where the journey works and where it needs a person.