AI Rework
Last updated
AI rework is the correction, verification, and redoing of AI output that happens after the AI interaction ends. It is invisible to usage dashboards because it falls outside the interaction they measure.
What AI Rework actually means
A dashboard measures the interaction. A prompt was sent, a response returned, a session lasted a certain length. Everything after that is out of frame, which is a problem because that is where the cost of a wrong or partial answer is paid.
The rework takes several forms. Output that is checked line by line because it cannot be trusted unverified. Output that is regenerated because the first attempt missed context the model did not have. Output that is accepted, then corrected downstream by someone who noticed the error later, usually at higher cost than catching it immediately. None of these register as a failure in the interaction that produced them.
The consequence is that a deployment can show a genuine time saving at the point of generation and a net loss across the process. Drafting time falls, review time rises, and the total is worse than before, but the only measured segment is the one that improved. This is the most common way an AI program reports a gain that the operating numbers do not show.
Making rework visible requires observing the work around the AI rather than the AI itself. What matters is whether the process containing the AI got faster and more accurate, which is a question about the process, not about the model.
Examples
Drafting faster, reviewing longer
A generated first draft takes ten minutes instead of forty. Review takes fifty minutes instead of twenty, because the reviewer cannot tell which parts were reasoned and which were plausible. Net process time is worse and the drafting metric improved.
Verification that never existed before
A classification step is automated. A new checking role is created to sample its output, because the confidence bands are not trusted. The checking cost is real, ongoing, and appears in no AI dashboard.
Correction found downstream
An automated coding decision is accepted and surfaces as an exception three steps later, where unwinding it costs several times what reviewing it at the point of generation would have.
Frequently asked questions
Related terms
Adoption Metrics
Adoption metrics measure whether an AI tool is being used: seat counts, logins, prompts per user, satisfaction scores. They are frequently reported as ROI and are not ROI, because usage says nothing about whether the work improved.
Incrementality
Incrementality is the measure of whether outcomes actually improved against a true baseline, as distinct from whether a tool was used. Applied to AI, it asks whether the work got faster, cheaper, or better because of the deployment.
Baseline
A baseline is the measured record of how a process ran before anything was changed, used as the reference point for whether a deployment improved it. A baseline that stops updating stops being a reference point.
