The Hours AI Saves Go to Checking Its Work

In a controlled trial, experienced developers handed an AI coding assistant finished their tasks 19 percent slower than the ones working alone, and reported on average that the tool had made them 20 percent faster ^[1]. The stopwatch and the self-report disagreed by nearly forty points, in opposite directions. They accepted fewer than half of its suggestions and spent close to a tenth of their hours cleaning up what it produced. The researchers described that output without flattery ^[1].

The majority of developers who participated in the study noted that even when they get AI outputs that are generally useful to them—and speak to the fact that AI generally can often do bits of very impressive work, or sort of very impressive work—these developers have to spend a lot of time cleaning up the resulting code to make it actually fit for the project. — Nate Rush

The developers could see the code wasn't right for the job. What they couldn't see was that fixing it had already spent the time they thought they'd banked. That gap is where the entire enterprise buildout keeps falling: the people footing the bill don't experience paying it, so it never becomes a budget line. Scale it to a company and the arithmetic holds. Workday's study of thousands of employees found people gained one to seven hours a week with AI and then handed back 37 to 40 percent of it reworking errors, roughly $186 per worker per month, with only 14 percent of users consistently coming out ahead ^[2]. Workday's own researchers put the finding plainly.

It is the cost of implementing AI without investing in the humans who use it. — Workday

The tools keep improving while the return at the firm barely moves; one industry executive has watched the models get better year over year as the value companies actually capture stays flat ^[3]. That raises the question no budget answers: who ends up doing the checking, and why does it fall on the most experienced people in the building? There is a rival explanation worth stating honestly, that the stall sits in the fragmented data underneath the models rather than in checking costs ^[4]. Leave it hanging. The reversals answer it. The expensive part is where checkers come from. Recognizing output that points the right direction but misses what was actually needed used to be learned free, as a byproduct of the routine work beginners did. Automate that work and you automate the training. One software chief calls it the hidden apprenticeship tax: remove the entry-level tasks that were the classroom for professional judgment, and the people who will one day have to supervise AI results never learned how ^[5]. The labor market has already priced the lesson. Entry-level postings in the fields most exposed to AI fell 35 percent between January 2023 and June 2025, sharpest in software, data engineering, and financial analysis ^[6]. The new hire who used to learn by doing the grunt work no longer gets a seat at the workbench ^[6].

What that practically means, though, is that that junior analyst, junior banker, junior educator doesn't get a shot at participating in the work anymore because they are optional. — Matt Beane

One industry shows the double bind cleanly. British law firms have adopted AI almost across the board, and the billable hour is collapsing, with hourly billing's share expected to fall from 72 percent to 44 percent ^[7]. The same firms warn that AI has stripped away the repetitive junior work that once trained new lawyers, even as they keep paying people to manage hallucination risk ^[7]. The work AI creates to fix its own errors doesn't rebuild the pipeline either: the new remediation roles, humans hired to correct hallucinations and judgment gaps, are deskilled and lower-paid, and the people in them say the monitoring adds to their load ^[8]. The retraining everyone claims to be doing isn't reaching the workers whose jobs changed, with 66 percent of leaders calling training a priority while only 37 percent of the most affected employees can actually get it ^[2]. A three-month MIT study of patent lawyers found the same split: only senior lawyers with seven or more years on the job built better judgment from AI, while juniors walked away confident in skills they hadn't actually developed ^[9]. So the industry is spending down the one asset that turns model output into results, and not replenishing it. The walk-backs are what the bill looks like when it arrives. Up to 55 percent of companies that replaced staff with AI now admit they moved too quickly ^[10]. Commonwealth Bank reversed the termination of 45 customer service staff, and its explanation is on the record ^[10].

did not adequately consider all relevant business considerations and this error meant the roles were not redundant — Commonwealth Bank

Taco Bell is reassessing its drive-through voice AI ^[10]. What each reversal quietly restores is the oversight layer the automation displaced, the human check that had been doing the verification all along, unpaid and unbudgeted. Watch the next one, and the tell is concrete. Read what the rehire requisitions actually ask for: a body to answer the phone, or someone whose judgment can catch what the machine gets subtly wrong. And watch whether verification ever appears as a line item, or keeps hiding inside everyone's hours. If the gains show up without budgeting the checking or rebuilding the people who know how to do it, this argument is wrong.