The AI repair queue: count the work after the output

AI can make the visible part of a task look faster while quietly adding work after it.

A draft arrives sooner. A summary appears instantly. A list of recommendations is ready before the kettle boils. Then someone has to check the claims, repair the gaps, rewrite the awkward bits, find the missing context, explain the decision, and answer the follow-up questions the original output created.

That downstream work is the AI repair queue. If you do not count it, you can mistake activity for productivity.

The output is not the finish line

A useful way to assess an AI workflow is to stop asking, “How quickly did the tool produce something?” and ask, “How much human effort did it take to reach a usable result?”

That includes more than prompting and editing. It includes:

• checking facts, calculations, and citations;
• supplying context the system did not have;
• correcting tone, structure, or formatting;
• deciding what to keep and what to discard;
• repairing errors after the work has been handed off;
• answering questions caused by an incomplete or overconfident result.

The last two items are where the queue becomes expensive. A person may save ten minutes creating a first draft, only for a colleague to spend fifteen minutes figuring out what the draft actually means. The time was not removed. It changed owners.

Harvard Business Review gave this kind of low-substance output a memorable name: workslop—content that looks like work but does not meaningfully advance the task. The label is useful because it points to the real failure: not that AI produced text, but that the workflow treated production as completion.

What the research says about “time saved”

There is real productivity value in AI. The sensible response is not to deny it. The sensible response is to measure the whole chain.

A field experiment covered by the National Bureau of Economic Research followed 7,137 knowledge workers across 66 firms. Among the 80% of treated workers who used the AI tool, email time fell by two hours per week and time spent working outside regular hours also declined. But the researchers did not detect a change in the quantity or composition of workers’ tasks from individual-level access alone.

That is an important distinction. A worker can save time on email without the organisation automatically producing more valuable work. The saved time needs somewhere useful to go, and the surrounding workflow has to support the change.

Microsoft’s 2026 Work Trend Index makes the repair queue visible from another direction. In its survey of 20,000 AI-using workers across ten countries, 50% said quality control of AI output is a human skill becoming more important, followed by critical thinking at 46%. And 86% said they treat AI output as a starting point rather than a final answer.

That is not an argument against AI. It is a description of the job it creates around itself. The first output is often cheap. Judgment remains part of the cost.

Research from MIT Sloan and Microsoft reaches a similar conclusion at the workflow level. The “Chaining Tasks, Redefining Work” paper models production as a sequence of interdependent steps. Its empirical results suggest that AI-executed steps tend to co-occur in chains, while scattered AI-exposed steps are less likely to be automated at the job level.

In plain English: AI works better when you redesign the route, not when you sprinkle a faster tool onto one isolated step.

Run a two-week repair-queue audit

Choose one recurring workflow. Keep it narrow: a weekly report, customer-reply draft, meeting follow-up, sales proposal, or research brief. Do not audit “AI at work.” That is how spreadsheets become small monuments to avoidance.

For two weeks, record the following for every run:

1. Baseline effort: how long the task normally took before AI.
2. Generation time: prompting, uploading context, and waiting for the output.
3. Verification time: fact-checking, checking numbers, and comparing against the source material.
4. Repair time: rewriting, restructuring, filling gaps, and fixing wrong or invented details.
5. Handoff time: explaining the output, answering questions, or adapting it for the next person.
6. Downstream rework: minutes spent fixing a problem after the output was accepted or shared.
7. Outcome quality: whether the result met the same standard as the old process.

Also tag the repairs. “Needed more context” is a different problem from “made a factual error.” Useful tags include wrong, incomplete, unsupported, off-brand, wrong format, and unclear owner.

At the end of the two weeks, calculate the total human effort, not just the tool’s production time.

Net time saved = old-process effort − (AI generation + verification + repair + handoff + downstream rework)

The formula is deliberately boring. Boring is good here. It makes the hidden work harder to wave away.

Decide what to do with the result

Your audit should produce one of three decisions.

Keep the workflow

Keep it when total effort is lower, quality is stable or better, and the remaining repairs are predictable. Then document the good path: the input the assistant needs, the checks a human must perform, and the conditions that trigger escalation.

Redesign the workflow

Redesign it when the tool is useful but the repair queue is too large. Common fixes include giving the system better source material, narrowing the output format, adding a verification step before handoff, or moving AI earlier in the process so a human can shape the brief before generation.

The MIT research points toward this kind of redesign: connected AI steps can reinforce one another, while scattered automation creates coordination costs. If every handoff needs a fresh explanation, the workflow is not yet a workflow. It is a relay race in which everyone keeps dropping the baton.

Stop using AI for this task

Stop when the output regularly creates more checking, clarification, or rework than the old method—and when the problem is structural rather than a missing prompt trick. Not every task benefits from an extra layer of machinery.

For higher-consequence work, add a second test: can the action be reversed if the output is wrong? Let an assistant prepare, classify, compare, and draft where mistakes are cheap to correct. Keep a human decision-maker on commitments that are costly, public, sensitive, or difficult to undo.

The practical rule

Do not ask whether AI made the first step faster. Ask whether it shortened the path to a finished result.

The first output is a hypothesis. The repair queue is where the workflow tells the truth.

Run the audit on one recurring task this week. If the numbers show a real gain, standardise the path. If they show a transfer of work, fix the handoffs before adding more automation. More AI will not rescue a workflow that nobody has bothered to measure.