74% of companies roll back their AI agents. Why?
Almost three quarters of companies switch their AI agents off again. The model is rarely to blame.

In shortKurz gesagt
According to Sinch’s report “The AI Production Paradox” (May 2026, 2,527 decision-makers surveyed), 74% of companies have rolled back or switched off a customer-facing AI agent after go-live. Among companies with the most mature governance structures, the rate is as high as 81%. So AI projects rarely fail because of the technology. They fail because of missing trust, a poor experience and systems that run technically but don’t work for people.
What exactly does the number say?
First some context, so the 74% doesn’t sound like clickbait.
For the Sinch report “The AI Production Paradox”, 2,527 senior decision-makers from ten countries and six industries were surveyed. The largest groups came from financial services and healthcare, along with telecommunications, technology, retail and professional services. 74% of these companies have rolled back or switched off a customer-facing AI agent after deployment.
Two more numbers belong in the picture:
- This isn’t about projects stuck in the pilot phase. 62% already had their agents in production. They failed after go-live.
- Investment keeps rising anyway. 98% of companies are increasing their budget for AI communication in 2026. Nobody is getting out. Everyone wants to make it work.
So the report doesn’t say “AI doesn’t work”. It says that keeping AI live is harder than getting it live. That’s exactly what makes it interesting.
Why do companies with better governance roll back more agents?
This is where the study gets surprising.
You’d expect that the more mature a company’s control and security structures, the less often it would have to switch off its bot. It’s the other way around. Among companies with the most mature governance frameworks, the rollback rate is 81%, above the average.
Sinch’s head of product, Daniel Morris, explains it, in essence, like this: the most advanced companies don’t fail more often, they notice their failures sooner. Better monitoring finds the errors that go unnoticed elsewhere.
For the industry, that’s uncomfortable. It means the 74% is more of a floor. At companies without good monitoring, the problems probably just keep running, and nobody notices. Governance doesn’t prevent failure, it makes it visible. The real work lies elsewhere.
Why do AI agents really fail?
Hardly ever because of the model. According to the report, the main reasons for rollbacks revolve around trust and control. At the top are concerns about exposing customer data (around 31%) and hallucinations and brand risks (around 22%).
That matches what we see in projects. Agents get rolled back for three reasons.
Trust is missing. Customers don’t trust the system, and often the companies don’t trust it themselves. A bot that has once said something wrong with full conviction has lost that trust. And hallucinations can’t be switched off entirely, because they arise from how the models work. We’ve written in detail about why language models will always hallucinate and what helps.
The experience is poor. Technically, the agent is impressive. In everyday use, it’s frustrating. It doesn’t understand users, leads them in circles and offers no way out. People want to get their issue resolved, not wrestle with a machine. We show how users actually interact with bots, and why many bots are designed past that, in our article Nobody wants to talk to your chatbot.
Control is missing. Nobody has clearly defined what the agent may and may not do. Without clear limits and working escalation, every mistake becomes a risk, all the way to liability. In May 2026, the Higher Regional Court of Hamm made clear that companies have to answer for what their bots say. Read more in our article on the Hamm ruling and chatbot liability.
None of these three reasons is a model problem. They’re problems of design, trust and governance. A better language model solves none of them.
Why isn’t good technology enough?
“Works technically, but not for humans.” We’re hearing this sentence from a lot of companies right now, and it describes the problem more precisely than any statistic.
The technical side of AI is becoming a commodity. Models are getting better, cheaper and more interchangeable. What one model can do today, every model can do tomorrow. So whether an agent gets rolled back or stays is almost never decided by the technology. It’s decided by things you can’t simply buy:
- Does the system understand how real people behave, or only how they’re supposed to behave?
- Does it build trust because it’s transparent, deals honestly with uncertainty and shows its limits?
- Is it embedded in real processes, or is it a pretty chat window in front of a black box?
- Does it know when to hand over to a human, and does it do that cleanly?
These are questions for UX, design and trust. An agent that runs technically but has no answer to them will eventually be switched off. At the latest when monitoring is good enough to show that it does more harm than good.
How can companies keep their AI project from failing?
The obvious reaction to these numbers would be to invest even more in guardrails. But the study shows that isn’t enough on its own. More control makes errors visible. It doesn’t fix their cause.
You have to start earlier, at the concept stage rather than when securing a finished system:
- Start from the user’s need. The first question isn’t “Which model do we use?” It’s “Which need are we solving for whom, and how does that feel?”
- Design for trust from the start. Make uncertainty visible, show sources, be open about limits. An agent that honestly passes gets used again.
- Build limits and escalation into the experience. The best AI agent knows what it can’t do and then hands over cleanly to a human.
- Embed it in real processes. An agent that isn’t connected to the real systems and workflows remains a demo. Rollbacks happen where the integration was only skin-deep.
So the 74% don’t fail because AI doesn’t work. They fail because too many projects start with the technology and stop at the human. It should be the other way around.
Conclusion: The problem isn’t the AI. It’s everything around it.
The 74% isn’t a verdict against conversational AI. It’s a verdict against one particular way of building it: projects that start with the technology and only think about people at the end. The companies whose agents stay live don’t have the better model. They’ve understood that an AI agent only works once it wins people over too.
We’ve been building chat and voice bots in Berlin since 2007. We start with the conversation, not the model, connect the bots to your real systems and design the path to a human along the way. Do you have an agent that runs technically but isn’t being adopted? Or are you planning one that should stay from day one?
FAQ: AI Agents and Rollbacks
How many companies are rolling back their AI agents?
According to the Sinch report “The AI Production Paradox” from May 2026, 74% of the companies surveyed rolled back or shut down an AI agent used in customer interactions after it went live. Among companies with mature governance structures, the figure is as high as 81%.
Why do AI projects fail after going live?
Rarely is it a technical issue. The main reasons are concerns about customer data, hallucinations, and risks to the brand, a poor user experience, and a lack of clear boundaries. These are issues related to design and governance, not problems with the model itself.
Why do companies with better governance roll back more agents?
Because they catch their mistakes sooner. Better monitoring uncovers problems that go unnoticed elsewhere. The higher rate, therefore, points to better control—not to inferior systems. Across the industry, the actual error rates are likely to be higher.
So is AI even worth it in customer service?
Yes. In fact, 98% of the companies surveyed plan to increase their investments by 2026. The study shows that it’s harder to keep AI running than it is to launch it. Success depends on design, trust, and integration.
What distinguishes an AI agent that sticks around from one that gets phased out?
An agent that sticks around understands real user behavior and builds trust because it is transparent and honest about its limitations. It is integrated into real-world processes and seamlessly hands off to a human when necessary. Technical performance alone is not enough to achieve this.
QuoteZitat