Enterprises are moving quickly to put AI agents in front of customers. These agents are no longer limited to answering frequently asked questions or routing support requests. They are increasingly being trusted to authenticate users, access sensitive information, resolve complex issues, and make decisions on behalf of the businesses deploying them.
But there is a fundamental disconnect between what enterprises are asking AI agents to do and how much customers actually trust them to do it.
New research from Cyara reveals just how significant that disconnect has become. In a survey of 1,000 U.S. consumers, 47% said they have abandoned an interaction with an AI tool or AI-powered customer service agent because they did not trust the information or guidance it provided. Another 46% have changed or simplified their questions out of concern that AI would misunderstand them or deliver an incorrect response.
Think about what that means. Nearly half of consumers have either walked away from an AI interaction or adjusted their own behavior to compensate for the technology’s perceived limitations.
That’s not the seamless customer experience enterprises are investing in. It’s a warning that AI capability is advancing faster than customer confidence.
And as organizations give AI agents greater autonomy, the consequences of that trust gap will only become more significant.
Customers don’t just want faster answers. They want the right ones.
For years, customer experience technology has been measured largely by efficiency: how quickly an interaction can be resolved, how many requests can be automated, and how effectively businesses can reduce the burden on human agents.
Those metrics still matter, but they don’t tell the whole story.
Cyara’s research found that 42% of consumers are concerned AI agents will provide incorrect information, while 43% worry the technology won’t understand their requests. By comparison, only 21% cite slow resolution as a top concern.
In other words, consumers are more than twice as likely to worry about getting the wrong answer than waiting too long for one.
This distinction matters because an AI agent can appear to perform successfully while fundamentally failing the customer. It can respond immediately, maintain a natural conversation and bring an interaction to a close, all while providing inaccurate information or making a commitment it was never authorized to make.
From a traditional operational dashboard, that interaction might look like a success. From the customer’s perspective, it’s anything but.
If enterprises want AI agents to take on more responsibility, they must prove more than whether an interaction was completed. They need to validate that agents understand different customer intents, navigate complex journeys and deliver the right outcomes, even when conversations don’t follow a predictable script.
When AI gets it wrong, the brand pays the price
One of the biggest misconceptions about enterprise AI is that responsibility for its performance rests primarily with the technology provider.
Consumers don’t see it that way.
According to Cyara’s research, 52% of consumers believe the companies developing AI models and the brands deploying them share equal responsibility when an AI agent causes harm. Another 17% place responsibility solely on the brand using the technology.
That means 69% hold the deploying brand at least partly accountable.
And the reputational consequences of failure are substantial. An overwhelming 87% of consumers say they would lose trust in a company if they caught its AI agent providing an incorrect answer, including 50% who say they would lose a significant amount of trust.
Customers don’t care which model powers an interaction, which vendor developed the agent or how many systems were involved behind the scenes. They know which company they contacted, and they expect that company to stand behind the experience.
The implications become more serious as AI moves into higher-stakes interactions. An inaccurate response about store hours is frustrating. An incorrect explanation of an insurance policy, unauthorized change to a financial account, or mishandling of sensitive healthcare information can have much more serious consequences.
The more consequential the interaction, the higher the standard of proof should be. An agent that can reliably answer routine questions isn’t necessarily ready to make decisions involving money, sensitive data, or customer accounts. Enterprises need to establish clear boundaries for what AI can handle independently and when human oversight is required.
The business deploying the AI agent owns that customer relationship. It must also own the responsibility for proving the agent can be trusted.
The future of agentic CX depends on proof
The opportunity for agentic AI in customer experience is enormous. Businesses can make support more accessible, resolve issues more efficiently and free human teams to focus on interactions where their expertise matters most.
But scaling automation without scaling accountability is a dangerous trade-off.
Cyara’s research shows that consumers aren’t simply accepting AI at face value. They are questioning its accuracy, limiting what they share, changing how they communicate and, in many cases, abandoning interactions altogether.
Enterprises should view these behaviors as an early indication of what is at stake.
The companies that succeed with agentic AI won’t necessarily be those that automate the most interactions or deploy the most sophisticated models. They’ll be the ones that can independently verify their agents are behaving as intended. As enterprises rely on more AI vendors and models, they cannot depend solely on those same providers to evaluate whether their technology is working.
Trust isn’t something an AI agent earns once, when it passes a test or goes live. Agent behavior can change as models, workflows, and customer expectations evolve. Enterprises need to continuously evaluate how agents perform in real-world interactions, identify when behavior drifts from expectations, and turn those failures into tests that prevent the same issues from recurring.
And ultimately, the question isn’t whether your AI agents are capable of doing more. It’s whether you can prove they’re ready to be trusted with more.
Survey methodology: This online survey was commissioned by Cyara and conducted by Dynata in September 2026 among 1,000 U.S. adults, with demographic representation across gender, age and region.