A survey of more than 2,500 senior decision-makers this year found that three quarters of companies have already rolled back or shut down customer-facing artificial intelligence agents after deploying them. That number is getting quoted everywhere, usually as proof that agents are not ready for customers.
The more useful number is further down the same study. Among organizations with mature governance frameworks, the rollback rate was higher, not lower. It was 81 percent.
That is worth sitting with. The companies with the most discipline pulled back the most often. I do not read that as governance failing. I read it as governance working, just late. Those companies could see what their agents were actually doing, and they acted on it. The rest are still live and quieter, which is not the same thing as safer.
So the question for customer experience leaders is not whether agents will do something you did not intend. They will. The question is whether you find out from your own systems or from a customer.
Companies have spent 20 years building operational muscle for deterministic software. A transaction completes or it throws an error. Tests pass or fail. Monitoring alerts you when something breaks.
Customer-facing AI agents break that model in the following three ways:
- They are non-deterministic. The same customer question can produce a different answer on Tuesday than it did on Monday, with nothing having changed on your side.
- They are goal-driven rather than scripted. You give an agent an objective, and it finds a path to it, which means the full set of possible conversations cannot beenumerated in advance.
- They make commitments, not just answers. Refunds, exceptions, policy interpretations, promises about what your company will do. Those are liabilities, and they are made in your brand's name.
The hardest part is that none of this announces itself. When an agent gives a customer a confidently wrong answer, nothing crashes. There is no alert and no error message. In your logs, the interaction looks like a success. The customer simply did not get what was needed.
Most agents perform well in evaluation. That is what evaluation is for, and I want to be clear that pre-launch testing matters more now, not less. But an evaluation score is a per-trial success rate, and per-trial success rates are unforgiving at volume. A 99 percent score against a golden set sounds like a finished product. At a million conversations a month, it is 10,000 failed conversations.
The bigger issue is the conditions. Agents get graded against clean inputs: the expected question, complete data, a cooperative customer, a quiet line. Real customers change topic mid-sentence, talk over the agent, give half an account number, call from a moving car, and arrive angry about something that happened last week. An agent that is 99 percent accurate in the lab is a different agent under accents, background noise, interruption, and emotion.
The work is not to test less before launch. It is to test under the conditions and volume your customers will actually create and to keep testing after launch, because the model, the prompts, the knowledge base, and the customers all keep moving.
Most Hallucinations Start Before the Model Responds
When an agent gives a wrong answer, the model takes the blame. We test against more than 55 AI agent vendors and every major contact center-as-a-service platform, and the pattern holds no matter which model is underneath. It usually is not the model. It is the context the model was handed: an outdated knowledge article, a CRM field that never got populated, two policies that contradict each other, one of them retired last quarter but never removed. Each system passes its own tests. Every component is green. The journey still fails in the seams between components, which is exactly where nobody is looking. That layer between the systems is the one we have spent years instrumenting, across 450 enterprises and 350 million customer journeys a year. Agentic or not, it breaks the same way.
That is why a component-level view is not enough. You have to validate the whole path: what the agent was told, what it was allowed to access, what it committed to, whether it escalated when it should have, and what the human on the other end of the handoff actually received.
It is also why the grading cannot come from the agent vendor. No vendor should judge its own AI. Validation has to run independently against your policies and regulatory obligations, not against a vendor's definition of success.
The companies getting real return from agents did not deploy fastest. They built an operating model around evidence. The following three things show up consistently:
- They define the box before launch. Which sources are authoritative, which customer data the agent can touch, which decisions it can make on its own, and when it has to hand off to a person. That is a business decision, not a technical one, and it belongs with the people who own the brand and the compliance exposure.
- They measure in production, continuously, against real customer behavior rather than the scenarios someone imagined during design.
- They close the loop, which most programs are missing. Every failure production becomes a test the agent has to pass before it ships again. Without that, you have built an expensive way to rediscover the same problem. The agent you run in month six should be measurably better than the one you launched, and you should be able to show the work.
None of this is a reason to slow down. It lets you move. Most stalled AI programs I see are not stuck on technology. They are stuck waiting on a sign-off nobody wants to give, because nobody can produce evidence the thing is safe. Produce the evidence and the sign-off gets easy.
Boards are already moving past "Did we launch it?" to "How do we know it is working?" That second question needs a real answer with evidence behind it. "The agent did it" will not survive a regulator, a reporter, or a board meeting.
Sushil Kumar is CEO of Cyara.