Three in four AI support agents get pulled back. The off switch is the problem.

Sinch surveyed more than 2,500 enterprise decision makers across ten countries this year and asked a question most vendors avoid: after you shipped the AI agent, did you have to pull it back? Seventy-four percent said yes. At least once.

The number that should stop you is the next one. Among the organizations that described their own guardrails as fully mature, the rollback rate was 81 percent. Higher. Sinch’s chief product officer, Daniel Morris, told IT Pro that “the most advanced organizations aren’t failing less; they’re seeing failures sooner.”

He’s right, and it’s still worth asking why seeing a failure costs you the whole agent.

The causes aren’t about model quality

Look at what triggered the rollbacks. Customer data exposure, 31 percent. Hallucination or brand risk, 22 percent. No way to audit what happened, 16 percent.

Not one of those says the model wasn’t smart enough. They’re scope problems. The agent touched data it shouldn’t have touched, or asserted something it had no source for, or did something nobody could reconstruct afterward. Those are questions about what the agent was permitted to do and what it had to prove before doing it. A larger model does not answer them.

One switch, one lever

Here is the part that goes unsaid. In most deployments, the unit of deployment is the agent. You configure it, you test it, you flip it on, and from that moment it is answering everything that arrives.

If the unit of deployment is the agent, then the unit of rollback is also the agent. There is exactly one lever and it is labeled OFF.

So a team finds that refund questions are producing confident nonsense. Refunds might be four percent of the volume. The other ninety-six percent, order status, shipping windows, password resets, hours, was working. The lever doesn’t know that. Pull it and you lose all of it. Then the postmortem gets written up as “we rolled back our AI customer service agent,” which reads like a verdict on the technology and is really a verdict on the granularity.

Gartner’s prediction that over 40 percent of agentic AI projects will be canceled by the end of 2027 lands differently once you see it this way. Cancellation is what happens when the only response to a small failure is a large one, twice.

Make the unit smaller

Give an agent one job at a time and the economics of a rollback change completely.

If shipping questions are live and refund questions are not, then refunds going sideways is a Tuesday afternoon, not an incident. You demote the one topic. Everything else keeps working. Nobody writes a memo. And because the topic is a real object in the system rather than a paragraph buried in a prompt, you can actually answer the auditability question that sank 16 percent of these deployments: which topic, which conversation, which source, what did it decide and on what basis.

That also changes what “not ready yet” means. A topic the agent has not graduated isn’t off. It’s drafting. The agent writes the answer on live traffic, a person reads it, edits it if it’s wrong, and sends. You accumulate a real record on real customers before a single unreviewed word reaches one. By the time you graduate a topic you already know its failure rate, because you watched it for a few hundred conversations.

The other half is architectural, and it’s the part people skip. The model cannot be the thing that enforces the scope. If the same system that writes the answer also decides whether it’s allowed to send the answer, there’s no control loop, just a very articulate intern grading its own homework. The decision to send has to sit in ordinary deterministic code that checks a permission, a topic, and a source, and refuses when any of the three is missing. The AI proposes. Hard code decides.

What this looks like in practice

Celeste learns from a business’s own support history and starts with zero topics live. Every answer is a draft the owner approves. You graduate one topic when you’ve seen enough of it to want to, and you can put it back in drafting with one click when it disappoints you. No topic sends without a source it can point at. The send decision is code, not judgment.

None of that is a claim that nothing will ever go wrong. Things go wrong. The design goal is that going wrong costs you one topic for one afternoon instead of the whole program for a quarter.

Read the 81 percent number again with that in mind. The mature teams aren’t worse. They’re the ones actually looking. The thing to fix isn’t how often they roll back. It’s how much a rollback costs them.