AI Moves the Work to Review: Budget for the Checking
A project manager with no coding background asked an AI to make a change to the company's software. The AI wrote the code. A second AI reviewed it and said it was fine. Then a group of developers did a human review on top, because their rules require one for anything an AI produces.
Another developer described where that leads. A colleague outside engineering has an AI picking up every new ticket as it arrives, and the developers review what it produces. "Try reviewing 3 a day every day."
The writing got cheap. The checking did not, and it landed on the people with the least spare time.
Where the hours go
This pattern is not limited to software. Someone in a group of automation builders described their best small win. A job pulls figures from weekly emails into a spreadsheet and drafts the summary. It cut 40 minutes to 5. The 5 minutes is a person reading it before it goes.
A sales rep said the same about AI meeting recaps and follow-up drafts. Checking each one before it reaches a customer takes 5 to 10 minutes. That is still a large saving. It is also time someone has to find.
In software the gap is wider. Some in the software industry argue, in effect, that reviewing code now takes far longer than writing it, and that review is where AI projects stall.
Why rollouts miss it
Most AI business cases count the hours saved on the first draft. Few count the hours added to the review. The drafting is spread across many people, so the saving looks large. The checking lands on a few senior people, so the cost looks small, until their week fills up.
That is how adoption fails while the tool works. The reviewers fall behind. Drafts pile up. People start approving without reading, or stop using the tool to avoid the queue. As we argued in Freed Hours Need a Job, the hours never spread evenly, and nobody plans where they go.
How to plan for the checking
Start with three numbers before any rollout. How many outputs will the tool produce each week? How long does a careful check of one take? Whose desk does that check land on? Multiply them out. If the answer is ten hours a week for your most senior person, plan for that, or change the design.
Then set the rules for what waits. People who run drafting agents settled on one: an unreviewed draft is a no, never a yes. Nothing goes out because nobody got to it. A draft that sits too long gets dropped or rebuilt from fresh information, because yesterday's reply to a thread that has moved on is often wrong.
Train the reviewers too. Checking AI output is its own skill. Staff need to know which errors the tool makes, when to trust it and when to check the source. That is the people pillar in our six pillars of readiness, and it decides whether a person stays in control or just clicks approve.
How to tell if you have the problem
Look for a queue. Drafts waiting for approval, pull requests waiting for review, AI summaries nobody has opened. If the queue grows each week, the tool is producing faster than your people can check, and the saving is on paper only.
Who does this work
Training and adoption is the fourth stage of how we work, included in every engagement. It covers training for the staff who live with the system, an AI use policy people follow, a champions program, and adoption measured like everything else. Part of that is sizing the checking before the tools go out.