Back to insightsThe knowledge journal

The Agent Access Wars: Why Businesses Are Blocking Their Customers’ AI

Customers want to delegate online tasks to AI agents; businesses want control over who acts on their systems. What United's restrictions reveal and how to design controlled agent access.

20 min read

United’s restrictions expose a growing conflict: customers want to delegate online tasks, while businesses want control over who accesses their systems and how. The next competitive advantage may be a dependable way to negotiate that access.

TL;DR

  • United announced restrictions on unauthorized third-party personal AI agents on 11 November 2026, covering direct access and indirect access through travel intermediaries.

  • Customer approval, agent identity, and a business’s permission to perform an action are separate questions.

  • The conflict combines genuine operational risks with commercial competition over distribution, customer attention, and accountability.

  • Search indexing, product discovery, account access, purchase execution, and AI training need distinct policies.

  • Payment capabilities such as ChatGPT Wallet do not establish a universal right to transact on every website.

  • Product teams should define permitted actions, provide clear human handoffs, measure legitimate customer friction, and build controlled delegation where it makes business sense.

You ask an AI assistant to compare flights and book the one that fits your budget and schedule.

You approve the itinerary. You are the traveler, the account holder, and the person paying.

The airline still says the agent cannot act.

That tension sits behind the United AI agent ban, the shorthand being used for the airline’s restrictions on unauthorized third-party agents.

The customer sees a tool carrying out their instructions. The business sees software operated by another company entering an account, consuming system capacity, and potentially making consequential changes.

Both perspectives describe part of the same interaction.

As AI assistants move from answering questions to completing tasks, this disagreement becomes a product problem for every company with an online customer journey.

Which parts of the journey can be delegated? How can a company verify the delegation? What happens when the customer approves an action but the business does not permit that agent to perform it?

The agent access wars are becoming a contest over the conditions under which software can represent a customer.

What United actually restricted

On 11 November 2026, United said it was taking steps to restrict unauthorized third-party personal AI agents and assistants. The announcement covers direct access to its digital platforms and indirect access through travel agencies and aggregators, with changes to booking, ticketing, and distribution policies. Source: United.

The word unauthorized matters. “United rejects all AI” would describe a much broader position than the announcement establishes.

The policy also reaches beyond the final payment step.

Search Engine Journal’s 14 November examination of United’s updated terms and Digital Trust Center reports a distinction between search-engine indexing and agent access. It says agents need permission to search, book, buy, manage, or collect information through United’s site or app. Its account of the FAQ says AI-assisted trip planning remains allowed when the tool does not access those services. Source: Search Engine Journal.

That makes “the agent can research but cannot buy” an incomplete description. Where the research takes place and how the information is accessed matter too.

A policy decision is different from technical enforcement

A published restriction establishes the company’s stated position. It does not tell us how accurately every agent will be detected or how consistently the policy will be applied.

Nor does an accessible page establish that every form of automation is permitted.

For product leaders, the practical lesson is to align the rules users read, the permissions partners receive, and the behavior the system enforces.

If those layers disagree, the result is confusion for customers and support teams.

A restrictive policy is one possible industry response

Southwest provides a useful contrast. Its 6 October 2026 announcement introduced a ChatGPT plugin for discovering and shopping flights, described as the first phase of a broader approach with OpenAI and Amazon Web Services. That announcement should not be read as permission for every third-party agent or as proof of unrestricted autonomous booking. Source: Southwest announcement.

The choice is therefore more specific than whether a business supports AI.

A company can welcome one defined integration, restrict another, and require the customer to complete sensitive steps directly.

Why businesses are blocking AI agents

There are legitimate reasons to be cautious about software acting inside a customer account.

There are also commercial reasons to resist another intermediary.

A useful analysis considers both without assuming that either explains every decision.

Account access can expose more than the current task requires

United cites risks from customers sharing account credentials with third-party assistants, including exposure of sensitive information, unintended purchases, account changes, and use of MileagePlus miles. It also raises potential aviation-regulation concerns, using baggage acknowledgment at check-in as an example. Those are United’s stated concerns, rather than a finding here that every agent-assisted action violates a regulation. Source: United.

An instruction such as “find a suitable flight” does not inherently require authority to alter account details or spend a loyalty balance.

When the available connection exposes a full account session, the service may struggle to distinguish the intended task from everything else the session can do.

A customer’s confidence in the assistant does not eliminate that mismatch.

Consequential actions require dependable interpretation

Travel illustrates the problem particularly well.

Changing a flight may affect a connection, a companion’s itinerary, a fare condition, or an important arrival deadline. Canceling a booking can be harder to reverse than retrieving a schedule.

The product needs a dependable way to establish what will change, who approved it, and what was actually completed.

A generic confirmation that “the customer wants help” is weak evidence for a specific irreversible action.

Automation changes the cost of exploration

An assistant can compare many options and repeat searches while pursuing a customer’s requirements.

That can be useful. It can also create workload without producing a transaction.

Consider an agent that checks several date ranges across multiple routes every few minutes. The customer’s goal may be reasonable, but the aggregate activity can become expensive when many agents behave similarly.

This is an operating consideration for any search-heavy service. It should be measured in actual workload, capacity, and completed customer tasks, rather than assumed from the presence of automation.

A company may lack a workable accountability arrangement

When an agent makes a mistake, the customer expects help from someone.

The business may need to establish whether the action reflected the customer’s instructions, whether the agent changed the plan, whether a third party supplied stale information, or whether the business itself mishandled the request.

A blanket restriction can be an interim response when the service has no dependable way to answer those questions.

That does not make it a complete long-term product strategy.

Customer delegation and business permission

The central disagreement comes from treating several permissions as though they were one.

A customer says, “I authorize my assistant to help me.”

A platform says, “This request came from our assistant.”

A business asks, “What action may this software perform here, on whose account, under which conditions?”

Those statements answer different questions.

Three checks belong in the same journey

CheckWhat it establishesWhat remains unresolved
Customer delegationThe customer has authorized a defined taskWhether the business supports that agent and action
Agent authenticationThe request comes from an identified operatorWhich customer it represents and what that customer permitted
Service authorizationThe requested action is permitted under the service’s controls and arrangementsWhether execution succeeds and the customer’s goal is met

This is an operating model for evaluating agent access, rather than a description of a universal legal standard.

For a particular service, contracts, applicable rules, and implementation details can impose additional conditions.

Payment approval adds another bounded capability

OpenAI launched ChatGPT Wallet on 13 November, two days after United’s announcement. The launch uses user approval and network payment tokens, initially through dots for eligible US Pro subscribers. Source: OpenAI.

Wallet helps address how an agent pays.

It does not establish that every business permits the agent to access an account, retrieve information, place an order, or complete a regulated acknowledgment.

A successful payment authorization cannot resolve a separate disagreement over access.

That distinction will become increasingly important as payment products improve faster than the surrounding delegation arrangements.

Customers still need an understandable answer

Most customers will not think in terms of authentication layers or distribution agreements.

They need to know what the assistant can do, where it must stop, and how to finish the task.

An experience that researches a trip for several minutes and reveals the access restriction only at the final step creates avoidable frustration.

A better experience identifies the supported boundary early and carries the customer to the appropriate direct channel with their work preserved.

The commercial battle behind agent access

Security concerns explain important parts of the conflict. They do not remove the commercial stakes.

An agent can influence the shortlist, explain trade-offs, and become the place where a customer expresses intent.

For a company that previously received that interaction directly, this changes distribution.

Search Engine Journal, citing The Wall Street Journal’s reporting, quotes United chief executive Scott Kirby saying, “I don’t want to create another middleman.” Source: Search Engine Journal.

The statement makes the strategic tension unusually clear. It does not establish that security concerns are a pretext.

Businesses can lose context while retaining responsibility

A company may still fulfill the order and handle complaints while seeing less of the discussion that led to the purchase.

The agent platform may know the customer’s budget, alternatives, rejected options, and willingness to compromise.

The business may receive only the final request.

That information difference can influence future recommendations, repeat purchasing, and negotiating power. Its commercial effect will vary by category and by what information platforms and customers choose to share.

Customer convenience and business control can diverge

The customer may want one assistant to compare every provider.

Each provider may prefer a supported channel that preserves its own presentation, rules, service model, and commercial terms.

Neither side can solve that disagreement through a better checkout button alone.

The likely negotiation concerns access conditions: supported actions, identity evidence, request volumes, data use, commercial arrangements, and responsibility when something goes wrong.

Some companies will find the incremental demand attractive. Others may decide that particular forms of access impose more cost or risk than value.

Broad restrictions have a customer cost too

If an assistant becomes a normal way for a customer to organize tasks, blocking it can mean asking that customer to change their workflow.

The cost could be abandonment, additional support demand, or preference for a provider whose permitted journey is easier.

These are outcomes to test, rather than reasons to assume every restriction destroys demand.

A company with scarce inventory or a highly differentiated service may face a different trade-off from one competing against several close substitutes.

The question is whether a restriction protects the business in a way that justifies the customer friction it creates.

Why identifying an agent is only the beginning

Businesses need to know who is making a request. They also need enough context to decide what that requester may do.

Cloudflare’s 5 November analysis describes how hosted agents can obscure the relationship between a site and the underlying user. It frames the challenge around identity, reputation, and accountability, and introduces the Trusted Agent Delegation coalition as work toward a more interoperable approach. Source: Cloudflare.

This is a useful explanation of the coordination problem. It is not evidence that a complete delegation standard has already solved it.

A signed request does not carry every permission

OpenAI’s browser-access documentation describes Web Bot Auth, which uses signed requests to identify certain cloud browser traffic. It also distinguishes these visits from search crawling. Source: OpenAI developer documentation.

Recognizing the operator is useful, but it does not by itself establish that the customer authorized a particular account change.

The service still needs to evaluate the account, action, scope, and relevant conditions.

A familiar agent operator may represent both legitimate customers and someone attempting abuse. A policy should avoid granting every request identical privileges merely because it comes through the same platform.

An allowlist is a coarse decision

An allowlist identifies traffic that receives a particular treatment. It can be a practical control, but it does not describe the full customer task.

The service may want the same agent to read public information while requiring stronger verification for an account action.

Cloudflare’s 4 November release adds personal-agent visibility and policies for supported agents, including allowing or blocking them separately from search and training choices. The release also acknowledges that recognizing the agent does not necessarily supply the user context a site needs. Source: Cloudflare.

That is the direction product teams should examine: more precise decisions based on the action and context, with appropriate privacy protections.

Machine-readable rules need matching enforcement

Written policies help people understand a business’s position. Machine-readable capabilities can help an agent discover the supported path before attempting a task.

Neither replaces access checks inside the service.

A product page saying “booking is unsupported” should align with what the transaction system actually permits. A documented booking interface should enforce the conditions it describes.

Clear rules and reliable enforcement reduce unnecessary trial and error for both sides.

A practical access model for product teams

Start with a list of customer actions.

“Allow AI” is too broad to guide implementation. “Permit an approved agent to retrieve this customer’s itinerary for this session” gives a team something concrete to design and evaluate.

Separate actions by consequence

The following is a proposed policy framework, not a description of United’s implementation.

ActionExamplePossible product treatment
Public informationRead a published baggage policyPermit supported retrieval with suitable usage limits
Public comparisonSearch available products or servicesDefine permitted channels, freshness, and request budgets
Account readingRetrieve an existing reservationRequire customer authentication and limited delegated access
Reversible preparationBuild a cart or prepare a changeKeep it separate from final commitment
Purchase or modificationBuy a ticket or change a reservationCheck service permission and specific customer approval
Sensitive acknowledgmentComplete a safety-related declarationPreserve the required direct human step where applicable
High-impact account changeChange recovery details or transfer valueApply stronger controls or require direct completion

The framework helps a team explain why two actions receive different treatment.

Make permission narrow and revocable

A proposed delegated connection should define the account, permitted actions, duration, and meaningful transaction limits.

“Read this reservation until tonight” is easier to reason about than “do whatever I can do.”

Where supported, use a delegated connection rather than asking customers to hand over broad account credentials. Make it possible to withdraw access and explain what revocation stops.

Revocation also needs a clear boundary: stopping future actions does not automatically reverse an order already completed.

The service should enforce these limits. Putting them only in an agent’s prompt leaves the protection dependent on the model following the instruction correctly.

Keep preparation separate from commitment

A useful assistant can do substantial work before a sensitive action.

It can collect requirements, compare permitted options, explain conditions, and prepare a proposed transaction. The service can then require the customer to review the exact commitment through a supported channel.

This allows useful assistance even when complete delegation is inappropriate.

A handoff should preserve what can safely be carried forward: the selected option, current conditions, and the next required action. Sensitive data should travel through an appropriate protected mechanism.

Build a clear decision path

The following illustrates how a service could evaluate a request.

Flow diagram: Agent requests an action → Supported agent and action → Explain limit and offer direct path → Required customer authority verified → Request supported verification → Direct human step required → Hand off for human completion → Execute within enforced limits → Confirm outcome and record evidence Flow diagram: Agent requests an action → Supported agent and action → Explain limit and offer direct path → Required customer authority verified → Request supported verification → Direct human step required → Hand off for human completion → Execute within enforced limits → Confirm outcome and record evidence

The policy is evaluated before the action. The result must also be checked afterward.

A timeout should lead to reconciliation of the transaction state before retrying. A refusal should explain the supported next step. A changed offer should return to the appropriate approval point.

Design denials as part of the product

A generic “access denied” page gives the customer little help.

Where safe and appropriate, distinguish an unsupported agent, an unsupported action, expired permission, excessive request volume, and a requirement for direct human completion.

Provide enough information to recover without exposing sensitive account or security details.

Clear denials reduce the temptation for agents to guess, retry repeatedly, or promise a result they cannot deliver.

What changes for platforms and intermediaries

Agent access will affect more than the company operating the website.

It changes what assistant platforms can promise and what intermediaries must verify.

Agent platforms need dependable capability discovery

An assistant should establish whether a service supports the requested action before presenting it as something it can complete.

When access is unavailable, it should explain the boundary and preserve a useful handoff.

An assistant that can navigate many pages but routinely fails at the final action offers a weaker product than its demonstrations suggest.

Platforms will need to evaluate supported task completion, recovery, and policy compliance alongside model capability.

Travel agencies and aggregators inherit another constraint

United’s announcement explicitly includes indirect access through agencies and aggregators. Source: United.

The implication is practical: an intermediary cannot assume that its existing access arrangements automatically cover every downstream agent experience.

Before promising a capability, it should establish which uses are supported, who operates the agent, and what conditions apply.

That verification belongs in partner and product review, rather than being left until launch.

Integration providers gain a different kind of opportunity

There may be demand for services that help businesses identify agents, communicate supported capabilities, verify delegated access, and reconcile transactions.

The opportunity depends on solving those operating problems. Adding another intermediary that obscures responsibility would reproduce the conflict.

Businesses should evaluate whether an integration makes the authority chain clearer, whether it supports the channels they need, and how easily it can be changed or removed.

Customers need continuity across boundaries

A customer should not have to reconstruct the entire task because one provider requires a direct step.

Preserving continuity can become a competitive advantage: the assistant helps with preparation, the business handles the sensitive action, and both provide a clear outcome.

This requires coordination over the handoff rather than agreement that every step must happen inside one interface.

How to measure the cost of allowing or blocking agents

A blocked request is an event. It is not automatically a blocked attack or a lost sale.

An allowed request is equally ambiguous. It may represent useful customer work, repetitive exploration, or abuse.

Measure the customer task and the operating outcome.

QuestionUseful evidence
Are agents attempting useful tasks?Observed journeys, supported research, and verified customer feedback
Where does access fail?Denials and errors by action, policy reason, and identifiable agent
Are restrictions stopping legitimate customers?Sampled support cases and controlled journey tests
Does supported access create value?Completed tasks and contribution compared with an appropriate baseline
Does it create operating cost?Request workload, exception handling, and support effort
Do approvals lead to correct actions?Confirmed outcomes, reversals, duplicate actions, and complaints
Can customers recover?Handoff completion and time to resolve failed or uncertain actions

Estimate false positives carefully

A service cannot usually infer the number of legitimate customers behind every denied automated request.

Use controlled tests and sampled customer cases to estimate friction. State the limits of those estimates.

Separately inspect whether a broad policy is blocking a low-risk action that the business intends to support.

That gives the team a precise change to evaluate.

Compare policies at the same level of risk

It would be misleading to compare unrestricted public-page access with a tightly controlled purchase flow and conclude that one policy “converts better.”

Compare similar tasks, customers, and conditions. Account for inventory, pricing, promotion, and seasonality where those affect outcomes.

A staged rollout can establish whether a supported delegation path improves completion without increasing errors or support burden beyond an acceptable level.

Track the cost of recovery

The first action may be fast while the exception is expensive.

If an agent creates an incorrect booking, a duplicate order, or an uncertain modification, measure the time needed to establish what happened and restore a workable state.

A dependable access arrangement makes both success and failure understandable.

The business case should include that recovery cost.

Five decisions leaders should make now

Define the customer tasks worth supporting

Choose tasks with clear value and understandable consequences.

Reading a public policy, retrieving an order, preparing a comparison, and committing a payment can justify different access models.

Begin with a scope the company can operate and support.

Publish one coherent position

Align public guidance, partner terms, product behavior, and support instructions.

Explain what is supported, what requires approval, and where the customer must act directly.

A customer should receive the same practical answer from the assistant, the website, and the support team.

Establish an accountable cross-functional owner

Security can define threats and controls. Product can define the journey. Commercial teams can evaluate distribution economics. Legal and compliance can evaluate relevant obligations and arrangements.

Someone needs accountability for the combined decision.

Otherwise a marketing team may promote an agent experience that an access policy blocks, or a security change may disable a journey without a recovery plan.

Test the boundary cases before expanding

Include expired permissions, changed offers, unsupported actions, canceled requests, and uncertain transaction outcomes.

Test whether revocation actually stops later actions and whether a customer can complete the direct handoff.

Give support teams realistic examples of the failures they may see.

Set a review date and evidence threshold

An interim restriction can be reasonable when the operating model is incomplete.

Make the conditions for reconsideration explicit: supported identity, adequate customer authority, acceptable error rates, clear responsibilities, and demonstrated customer value.

Likewise, an access pilot should have conditions for narrowing or stopping it.

The objective is to make the policy responsive to evidence rather than let an initial allow-or-block choice become permanent by default.

Frequently asked questions

What is the United AI agent ban?

United’s 11 November restrictions cover unauthorized third-party agents, including indirect access through travel intermediaries. Source: United.

Does United prohibit all use of AI for trip planning?

Search Engine Journal’s examination of United’s FAQ distinguishes AI-assisted planning that does not access United’s site or app from unauthorized agents accessing those services. The restriction should not be summarized as a ban on every use of AI to think about travel. Source: Search Engine Journal.

Why are websites blocking AI agents?

Reasons can include account risk, unwanted automation, transaction errors, system cost, uncertainty about delegated authority, and commercial concerns about intermediation. The relevant combination differs by business.

Is customer approval enough for an agent to complete a purchase?

Customer approval is one part of the journey. The service must also support the agent and action, and the transaction must satisfy its applicable controls and conditions. Payment approval alone does not establish that access is permitted or that an order was completed.

Does Web Bot Auth prove what the customer approved?

Web Bot Auth helps verify the source of supported signed agent requests. A service still needs the relevant customer and action permissions. Operator identification and customer delegation are distinct checks. Source: Browser-access documentation.

Is blocking AI agents the same as blocking AI training?

No. A personal agent performing a customer task, a search crawler indexing pages, and a system collecting material for model training represent different uses. Businesses should state which uses their policies address.

Can a business allow browsing but require a human to complete the transaction?

A product can be designed around that boundary. The access policy, supported interfaces, and handoff need to make the distinction clear and enforce it reliably.

What should product teams build first?

Start with a task-level policy and a tested recovery path. Establish what the assistant can do, what evidence the service needs, and how the customer finishes an unsupported or sensitive step.

Controlled delegation becomes a product capability

The United announcement makes an emerging conflict visible.

Customers increasingly want assistants to carry out tasks. Businesses need to decide what those assistants may access, what they can change, and who answers when something goes wrong.

The response will vary across services. Some actions may remain direct. Others may work through supported partners or carefully bounded delegation.

A stronger model makes those boundaries understandable, gives customers a useful path through them, and lets companies adjust access as the evidence improves.

The competitive advantage is the ability to support useful delegation with clear authority and dependable outcomes.

Sources and further reading

Thanks for reading.
More on enterprise ai & ai agents, technology strategy & operations and ai security & governance

Back to top ↑
Find a new perspective

Explore the site

↑↓ to move ↵ to openSearch the entire collection