AI Agents Were Supposed to Save Labor. New Survey Shows Workers Still Cleaning Up After Them

Picture handing a new hire a stack of reports and asking them to double-check every number before it goes to a client. Now imagine doing that same review, every single day, for a piece of software that never sleeps and keeps generating more work to check. That's the reality a growing number of companies are describing when they talk about deploying AI agents, the software systems designed to complete tasks on their own with minimal human input.

A new Collibra-sponsored Harris Poll of 306 U.S. data management, privacy and AI decision-makers found that more than three-quarters of organizations hit roadblocks trying to move AI agents from small pilot programs into full production over the past year. More than half said their staff spent significant time reviewing and correcting what those agents produced before anything went live. Nine in ten said their organizations are already deploying autonomous agents in some form, so this isn't a fringe experiment. It's the mainstream approach, and it's already generating friction.

That friction has a name now. Collibra calls it the "hallucination tax," a term for the hidden cost of manual oversight, rework and risk that builds with every new agent a company puts into production. It's a useful way to think about it. An AI hallucination happens when a system generates information that sounds confident but is wrong, incomplete, or simply invented. Someone has to catch that before it reaches a customer, a regulator, or a decision-maker. That someone is a person, and that person's time isn't free.

Who Ends Up Doing the Work

The survey's numbers get more pointed when you look at company size. At larger companies with annual revenue of $100 million or more, 64% of respondents reported significant employee time spent reviewing and correcting agent outputs. Across the full survey pool, that figure was 51%. Bigger companies, running more agents across more workflows, are seeing more of their workforce pulled into cleanup duty rather than less.

The survey doesn't say how many hours this amounts to, which is a real gap in the data. But the direction is clear enough. As deployment scales up, so does the burden of checking the work. Nearly nine in ten respondents, 87%, said their teams regularly verify whether the information feeding these agents is still accurate and current. That's a lot of people spending a lot of time making sure the raw material an agent is working with hasn't gone stale.

Felix Van de Maele, Collibra's co-founder and CEO, put it plainly: "We have found that every enterprise scaling AI today is paying a hallucination tax, a hidden cost of manual oversight, rework, and risk that grows with every new agent put in production." It's worth sitting with that phrase. A tax is something you pay regardless of outcome. It doesn't matter how well the agent performs on its best day if the average day still requires someone to redo part of the job.

The survey also found that 72% of respondents believe poor or unaligned data is almost always behind their organization's disappointing AI results. This points to something deeper than a software problem. It's a data problem, and data problems are notoriously unglamorous to fix. Separate research from Gartner, surveying 223 data and analytics leaders in March, found that 60% cited cultural resistance as a reason governance initiatives fail, compared to 40% who cited funding constraints. Weak business engagement and limited understanding of what governance actually delivers were also flagged as obstacles.

None of that should surprise anyone who has worked inside a large organization. Keeping records accurate and agreeing on shared definitions takes time from people who already have full plates. The payoff feels abstract, until an application starts spitting out answers nobody trusts, and suddenly everyone remembers why that unglamorous work mattered.

There are signs some companies are responding. Collibra found that 53% of respondents said their AI function's reporting line had moved closer to the primary data organization over the past year. Among companies with at least $100 million in revenue, that figure climbed to 62%. The logic tracks: teams building and deploying agents often depend on data they don't own or maintain, so putting them closer to the people who manage that data gives them a faster path to flagging and fixing problems.

But a reorganization chart doesn't automatically fix an agent's output. The survey offers no evidence that moving reporting lines, by itself, improves performance. What matters more is what happens next: are the same errors showing up again and again, who's actually resolving them, and are employees spending less time on corrections as a result. Those are the questions that determine whether this is a real fix or just a rearrangement of the org chart.

Why It Matters

Human oversight of AI systems isn't inherently a red flag. A task that once took a person several hours might still be worth automating if all that's left for a human is a quick glance to confirm the output is right. That's a legitimate productivity gain, and it's exactly the kind of trade-off automation has always promised. The math changes entirely when the employee has to redo a meaningful chunk of the work rather than simply approve it. At that point, the agent isn't replacing labor. It's just relocating it, and often adding a layer of frustration on top.

The survey doesn't tell us which of these two experiences is more common across the economy right now, and that gap matters. It leaves companies expanding their AI deployments with a basic accounting problem they can't avoid. Counting how many tasks an agent completes tells you very little on its own. What actually matters is how much work is left over for the human who receives the results, and whether that remaining work is shrinking over time or quietly growing. Until organizations can answer that honestly, claims about automation's labor savings deserve real scrutiny, especially from the workers doing the checking.