Saturday, November 21, 2026
AI

The Permission Gap: Why the Smartest AI Agents Still Can’t Buy a Plane Ticket

Nearly every company uses AI agents, but almost none let them act alone. Why the next bottleneck in AI is not intelligence. It is permission.

The Permission Gap: Why the Smartest AI Agents Still Can’t Buy a Plane Ticket

Nearly every large company now uses AI agents. Very few let them act on their own. The next bottleneck in artificial intelligence is not intelligence at all. It is permission.

On November 11, United Airlines drew a line that says more about the state of AI than most product launches this year. The carrier announced it would restrict unauthorized, third-party personal AI agents from booking or changing flights, whether they arrive through United’s own app and website or through travel agencies and fare aggregators. The assistants can still look up United flights and show fares. They just cannot complete the purchase. Reservations made by agents before the cutoff are honored.

The airline’s stated reason is security. In its words, sharing account credentials with third-party AI assistants “can expose sensitive information.” That is a reasonable worry. It is also a striking moment: the technology that was supposed to book your dinner, your dentist and your flights has hit a locked door at the last step, and the door belongs to a company that has every reason to keep the key.

The numbers say yes. The permissions say no.

If you only read adoption surveys, you would conclude that the agent era has already arrived. Zapier’s 2026 survey, run by Centiment among 513 managers and above at US companies with more than 1,000 employees, found that 88% of enterprise leaders say their organizations are using or testing AI agents, up from 72% a year earlier. Seventy-one percent report agents in production, up from 40%.

Now look at what those agents are allowed to do. In the same survey, only 30% of leaders say their agents complete more than 60% of assigned tasks without a human stepping in. A separate November study from PYMNTS Intelligence tells a sharper version of the story: 89% of enterprises use agentic AI, yet 78% would limit agents to reading company directories. Just 20% would let an agent complete tasks with approval, or act on its own.

Read those figures together and a pattern appears. Companies have deployed agents widely, and then fenced them in. The agent can look, summarize, draft and suggest. It is rarely trusted to send, buy, change or delete. We have built a very large workforce of interns and, so far, mostly decided to keep them away from the cash register.

Autonomy is not a model capability. It is a legal, commercial and organizational decision, and most of the people making it are not engineers.

Intelligence was never the only bottleneck

For two years the public conversation about agents has been a conversation about capability: can the model plan, can it use tools, can it recover from mistakes. Those questions matter, and the answers keep improving. But the United decision and the survey data point to a different set of questions that capability alone cannot answer. We see three locks on the door.

The first lock is identity. When you log in to an airline, a bank or a hospital portal, the system believes it is talking to you. An agent acting for you is a new kind of party: not you, not a stranger, but something in between. Today the only way to bridge that gap is often to hand the agent your credentials, which is exactly what United says it does not want. Until agents can carry their own narrowly scoped, revocable, clearly labeled identity, many institutions will treat them as intruders by default.

The second lock is liability. If an agent books the wrong flight, empties a gift-card balance or approves the wrong invoice, who pays? The user who delegated, the company that built the agent, the model provider, or the business that accepted the order? Consumer law and card-network rules were written for humans pressing buttons. They have no settled answer yet for software that decides on its own. Every unanswered question here pushes a cautious risk officer toward the same word: no.

The third lock is incentives, and it is the one that gets the least attention. Travel industry coverage has noted that an assistant trusted to choose an itinerary and buy the ticket could gain real influence over the sale. Whoever controls the point of purchase controls the shelf. Airlines, retailers and marketplaces spent decades building direct customer relationships and paid placements. An agent that quietly compares everything and picks the best option for the buyer is wonderful for the buyer and uncomfortable for the seller. Seen this way, a block on agent purchasing is not only about safety. It is also about who keeps the customer.

Autonomy follows architecture

One finding in the Zapier data deserves more attention than it received. Among organizations where more than half of their agents are embedded in broader automated workflows, 77% say agents complete most tasks without intervention. Where half or fewer are embedded, that figure falls to 8%. A caveat is needed here: Zapier sells workflow automation, so the survey naturally favors that conclusion, and the correlation does not prove cause. Still, the direction makes sense.

An agent dropped into a chat window has to be watched, because nothing around it limits what it can do. An agent inside a well-defined workflow, with fixed inputs, fixed outputs, logging and a clear owner, can safely be given more room. The surprise is that trust is built less by making the model smarter and more by making the surroundings stricter. The most autonomous agents in business may turn out to be the most constrained ones.

That also explains why the best deployments we see are unglamorous: document analysis, data entry, ticket triage, report drafting. These are bounded jobs with measurable results and easy review. If you want the practical version of this, our earlier piece on five questions to ask before you deploy an AI agent at work walks through how to scope one.

The case for the skeptics

It would be easy to read all this as institutions being slow and fearful. That is only half right. Companies that limit agents to read-only access are making a rational bet while error rates, liability rules and security practices are still moving. The PYMNTS report also notes that about half of executives see internal barriers, not technology, as the larger obstacle to adoption, which suggests the hesitation is organizational as much as technical. And no survey here reports failure statistics, so we do not actually know how often agents go wrong when they are let loose. Caution without data is understandable. It is also a gap that someone will eventually fill with evidence.

What we expect to happen next

These are BytesWall’s views, not reported facts, and we hold them loosely.

A permission layer will become its own product category. Expect tools that issue agents scoped, time-limited, auditable credentials: this agent may read invoices but not pay them, may book up to a set amount, may act only between these hours, and every action is logged and reversible. Whoever makes that layer trustworthy will matter as much as whoever makes the model.

The web will split into agent-friendly and agent-hostile zones. Some businesses will publish official, authenticated ways for agents to act and will price and police them. Others will block agents altogether, as United now does. Shoppers and companies will quietly learn which is which.

The first big disputes will be about blame, not capability. The headline will not be that an agent was too dumb. It will be that an agent did something plausible that nobody had clearly authorized, and nobody could say who was responsible.

A test you can run this week

If your organization uses AI agents, ask three questions. First: what is the most consequential action any of our agents can take today, and who approved that permission by name? Second: if that action goes wrong, who is accountable, and is that written down anywhere? Third: does the agent act through a scoped identity of its own, or through a person’s borrowed login? If the third answer is a borrowed login, you have found your biggest risk, and you have probably found the same thing United found.

The takeaway

The story of AI agents in 2026 is not that they are failing. Most of the people using them report real benefits. The story is that capability has outrun the systems of trust around it: identity, liability and who owns the customer. Those systems will not be solved by a better model. They will be solved by contracts, standards, new infrastructure and a few expensive mistakes. The companies that treat permission as a design problem, rather than a compliance afterthought, are the ones most likely to give their agents real work, and to keep their customers’ trust while doing it.

Sources:

Bytes Brief

Get the stories that matter in AI, in your inbox.

Written by

BytesWall

BytesWall is an independent newsroom covering AI, automation and emerging technology for developers, founders and business leaders. About BytesWall