Your AI Is Saving Time. So Why Is Everyone Checking Its Work?

AI is supposed to remove work. That is one of the clearest reasons companies are investing in it.

An agent handles the request. A copilot drafts the response. A system reviews the claim, recommends the next action, or completes a task that used to take a person several minutes. Multiply that across thousands of interactions and the value looks obvious. Faster decisions, lower cost, more capacity, better service.

But there is a problem showing up inside many AI programs that the standard performance metrics do not capture.

The AI completes the task, and then a person checks it.

An employee reviews the recommendation before acting. A manager validates the output before it reaches a customer. A customer contacts a human to confirm that the answer is right. A team keeps the old process running alongside the new one, just in case. The AI has technically removed a step, but the organization has quietly added another one around it.

That is the trust tax.

It is the extra work created when people do not fully believe the AI will get the outcome right. It shows up in verification, repeat effort, escalations, overrides, duplicate processes, and the constant need for human reassurance. None of those activities necessarily appear as a failure in the AI dashboard. The system worked. The answer was generated. The interaction was completed. Usage may even be going up.

But the work did not disappear. It moved.

Some of that checking is appropriate. No one should accept an AI decision blindly, especially when the stakes are high. The problem is the uncalibrated, invisible kind: the checking people do because they have no reliable way of knowing when the output can be trusted, so they check everything.

The Work Didn’t Disappear. It Moved.

Research from BetterUp Labs and the Stanford Social Media Lab offers a useful view of what that transfer can cost. Forty percent of US desk workers said they had received AI-generated work that looked complete but lacked the substance needed to move the task forward. On average, each incident took nearly two hours to resolve, which the researchers estimated at roughly $9 million a year in lost productivity for a 10,000-person company. The person creating the output may have saved time. The person receiving it inherited the work.

This is why productivity gains from AI can look much stronger in a business case than they feel inside the operation. The model measures the time saved on the task. It often does not measure the time spent checking, correcting, explaining, or redoing the output afterward.

The gap between those two numbers can be striking. In a controlled trial by METR, experienced developers expected AI to speed them up by around 24 percent, and afterward believed it had made them 20 percent faster. In reality, they were 19 percent slower when using the tools. It was a single study in one setting, but it is one of the most rigorous looks we have at a pattern that should give every AI sponsor pause: perceived productivity and measured productivity can point in opposite directions.

A task that once took ten minutes may now take two minutes for the AI to complete. That looks like an eight-minute gain. But if an employee spends four minutes reviewing it, another two minutes correcting the language, and a manager still needs to approve it, the productivity story changes quickly. The technology is faster. The total process may not be.

The same issue appears in customer service. An AI agent answers immediately, so response time improves. The interaction is counted as contained, so automation rates improve. But the customer is not confident in the answer and contacts the company again through another channel. The first interaction looked efficient. The second reveals that the problem was never fully resolved.

Fast Is Not the Same as Trusted

Fast is not the same as trusted. Automated is not the same as adopted. Completed is not the same as resolved.

This distinction matters because trust is often treated as a soft experience issue, something to address through better communications or change management after deployment. In reality, trust has a direct operating cost. When people do not trust an AI output, they change their behavior around it. They verify more, escalate more, avoid the tool, and build parallel workflows, leaning on a small group of experienced employees to confirm what the system produced.

Over time, those behaviors can absorb the savings the AI was supposed to create.

The Problem Isn’t Resistance

Our own research suggests the problem is not resistance to AI itself. In the 2025 Amdocs Studios EX20™ study, which surveyed 1,000 employees and 355 business leaders across 14 industries, 58.2 percent of employees said they were excited about using generative AI. Yet 47 percent said they lacked the training to apply it, and only 38 percent trusted how leadership was using it. The interest is there. The confidence to act without hesitation is not.

Those figures do not tell us how often every AI output is being manually checked. They tell us why checking becomes a rational response. When people are asked to use a system they do not fully understand, inside a strategy they do not fully trust, they create their own safeguards around it.

The problem is not that people are being irrational or resistant to change. In many cases, they are responding to what the organization has taught them. Perhaps the AI gave confident answers that turned out to be wrong, or could not explain how it reached a recommendation. Perhaps it works well for common cases but becomes unreliable when the situation is unusual. Perhaps employees do not know when a human has reviewed the output, or who is accountable if the decision causes a problem. Perhaps customers have learned that pushing past the agent is the only way to get something fully resolved.

Trust does not break because people fail to understand the opportunity. It breaks because the experience gives them a reason to hesitate.

Once that hesitation becomes normal, the organization can end up with the worst of both worlds. It carries the cost of the AI and the cost of the human process built around checking it.

You’re Measuring the System, Not the Behavior

This is especially easy to miss because most AI measurement begins with the system itself. Accuracy, latency, uptime, completion rate, containment, usage, and cost per interaction all matter. But they do not tell you whether people act on what the AI tells them without needing another layer of confirmation.

That requires a different set of questions. How often do employees override the recommendation, or copy the output into another tool to check it? How many cases are escalated after the AI has already provided an answer? How often do customers repeat the same request through another channel? Are teams still maintaining the old workflow next to the new one? Does usage reflect genuine adoption, or are people using the AI because they are required to and then completing the work somewhere else?

Those are not just trust metrics. They are productivity metrics. They show whether AI is actually removing effort from the system or simply relocating it to places the dashboard does not see.

There is evidence that this relocation is already happening. The Upwork Research Institute found that 39 percent of employees using AI said they were spending more time reviewing or moderating AI-generated content than before. The tool may have produced the first version faster, but the responsibility for making that output usable stayed with the employee.

Trust Is Not the Same as Adoption

This also exposes an important difference between trust and adoption. Adoption tells you whether people are using the AI. Trust tells you whether they are willing to rely on it.

A system can have high usage and low trust at the same time. Employees may be required to use it, customers may be directed into it, and teams may report strong activity. But if every output is checked, overridden, or taken to a human, usage alone creates a misleading picture of success.

The strategy may be clear at the top. That does not mean the people expected to make it work understand where AI fits, what they can trust it to do, or what responsibility they retain when they act on its output.

The Goal Is Calibrated Trust

The goal is not blind trust. People should not accept every AI decision without judgment, especially in high-stakes situations. The goal is calibrated trust: where people understand what the AI can handle, when its output is reliable, when a human is involved, and where the line for escalation has been deliberately drawn.

That line cannot be left to individual employees or customers to figure out for themselves.

Organizations need to decide which tasks the AI can complete independently, which require human review, and which should remain human-owned from the start. They need to make those boundaries clear in the workflow, not just in a governance document. They need to explain what the AI did, what information it used, and what happens if the result is wrong. And they need a visible feedback loop, so repeated problems lead to changes rather than becoming accepted friction.

Design Trust In From the Start

Most importantly, organizations need to involve the people who understand how the work actually happens. Yet 72 percent of leaders in our EX20 research acknowledged that AI adoption was being led without enough input from the employees it would directly affect.

That is how trust problems get designed into the system before it ever launches. When employees are not involved in deciding how AI fits into the workflow, the organization misses the exceptions, workarounds, and judgment calls that define the real process. The formal workflow may suggest that an AI output can be accepted immediately. The employees doing the work may know there are five situations where it cannot.

If that knowledge is not built into the design, people will create their own safety mechanisms afterward. They will double-check everything, escalate too much, or stop using the system for anything that matters. Those behaviors are understandable, but they are expensive.

The trust tax also compounds. One unreliable deployment does not stay contained to that use case. It shapes how employees and customers approach the next one. A customer who received a confident but incorrect answer from one agent arrives at the next interaction more skeptical. An employee who was held responsible for following a poor AI recommendation becomes less willing to trust the next one. Every new program starts with the memory of what happened before.

That means trust cannot be treated as something to rebuild after the technology is live. It has to be designed into how the AI works from the beginning.

What Leaders Should Actually Ask

This is one of the conditions our upcoming X20™ framework is built to measure. It looks beyond whether AI is available or being used and examines the organizational gaps that determine whether it creates real value. Trust is one of those gaps, but it does not operate alone. It connects directly to adoption, communication, service, accountability, measurement, and the balance between human and machine decision-making. If the AI cannot explain itself, people check it. If no one owns the outcome, people escalate it. If the service does not resolve the problem, customers repeat it. If the organization only measures activity, the extra work stays invisible.

The result can look like progress from the system’s point of view while feeling like more work to everyone around it.

So when leaders evaluate whether an AI deployment is delivering value, they should look beyond how much work the system completed and ask what happened next. Did people act on the output? Did the task stay completed? Did the customer accept the answer? Did the employee move forward without checking another source? Did the interaction remove work from the operation, or create a new layer of work around itself?

AI creates value when people can rely on it within clear, deliberate boundaries. Until then, the time saved by the system may simply be showing up somewhere else.

Your AI may be working. The more important question is whether everyone else still has to work around it.

References

• Amdocs Studios. EX20 Global Report. 2025. Global study of 1,000 employees and 355 business leaders across 14 industries.
• BetterUp Labs and Stanford Social Media Lab. Workslop: The Hidden Cost of AI-Generated Busywork. 2025. Online survey of 1,150 full-time U.S. desk workers.
• Becker, Joel, Nate Rush, Beth Barnes, and David Rein. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. METR, July 2025. Randomized controlled trial with 16 experienced open-source developers across 246 real-world tasks.
• Monahan, Kelly, and Gabby Burlacu. From Burnout to Balance: AI-Enhanced Work Models for the Future. Upwork Research Institute, July 2024. Survey of 2,500 executives, employees, and freelancers across the United States, United Kingdom, Australia, and Canada.
• Amdocs Studios. X20 Framework. Proprietary research framework, 2026.