The Review Burden AI Shifts, Not Removes
Work Moves From Production to Verification
The usual pitch is that AI removes work. What it reliably does is move work from producing text to checking text — and those two activities have very different properties. Production is predictable: a competent drafter knows roughly how long a first draft takes. Verification is not: checking a draft that is ninety per cent right takes far longer than the ten per cent would suggest, because you cannot know in advance which ten per cent is wrong, so you must examine all of it with equal attention. Verification is also cognitively harder to sustain. Reading for errors that are usually absent is a vigilance task, and human performance on vigilance tasks degrades quickly. The net time saving is real in many workflows, but it is smaller than the drafting time saved and it is unevenly distributed.
- Drafting time falls; verification time rises — the net saving is real but routinely overstated
- A nearly-correct draft is expensive to check because the errors are unlocated
- Vigilance degrades with time on task; error-hunting is not the same skill as reviewing your own work
- Measure the whole loop when evaluating a tool, not the generation step in isolation
Who Absorbs the Shifted Work
The review burden does not land evenly. Drafting work has historically been how junior practitioners learn — the slow production of a first draft builds the judgement needed to later review one. If juniors move straight to reviewing machine output, the training pipeline that produces competent reviewers is being consumed to fund present efficiency. Meanwhile senior practitioners find that reviewing plausible-but-unverified material is slower and less pleasant than reviewing a junior's work, because a junior's errors are patterned and a model's are not. Firms that adopt without redesigning supervision often discover both effects at once: partners doing more checking, and juniors developing the skills more slowly than the firm assumed.
- Drafting is how judgement is built — removing it from juniors has a delayed but compounding cost
- Machine errors are unpatterned, so reviewing them is harder than reviewing a known colleague's work
- Supervision load tends to migrate upward unless the workflow is deliberately redesigned
- Plan explicitly for how juniors will still learn to draft, or accept a weaker bench in a few years
Designing for the Burden Rather Than Denying It
Workflows that hold up are the ones that budget for verification instead of assuming it away. That means selecting tasks where checking is cheap relative to producing, preferring outputs that carry a traceable source so verification is a lookup rather than a reconstruction, and building the check into the process as a step with a named owner and a visible artefact. It also means being honest in fee and deadline conversations: if a matter is priced on the assumption that AI removed the work, the verification step is the thing that gets cut, and it is the only control that was doing anything. Adoption that ignores the burden does not eliminate it; it relocates it to whoever is least able to refuse.
- Prefer tasks where checking is cheap relative to producing — that ratio is the real adoption criterion
- Traceable, source-linked output turns verification into a lookup instead of a reconstruction
- Give the verification step an owner and an artefact so it survives deadline pressure
- Pricing and deadlines that assume the work disappeared will quietly delete the only real control
Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.