Your AI Rollout Needs a Review Budget

Imagine an AI rollout that doubles the number of customer implementation plans a team can draft. The launch report looks good. By Friday, forty plans are waiting for someone to check them.

Nobody has become slower. The team bought more drafting capacity than it could review.

This is a synthetic example, not a result from my work. It raises a question I would want answered before expanding an agent rollout: how many minutes of human attention does each additional output commit us to?

This week's releases point to the same constraint

On October 16, Anthropic published an analysis of its own workplace, combining internal data with employee interviews. It describes growing agent output and human attention becoming a constraint. The authors also flag rapid company growth and other factors that complicate causal claims. Its estimated reading burden is a calculation, not a record of time people actually spent reading. Anthropic's workplace study.

Two days earlier, Google introduced Deep Review for feedback on technical documents. Its system checks sections and synthesizes findings, offering a plausible way to automate some review work. The public product has document-upload and text-entry options. That establishes a product workflow, not evidence that it saves time in your team. Google's announcement.

There is a useful research check on the temptation to equate activity with delivery. The September revision of Writing Code vs. Shipping Code reports that gains in coding activity translate only partly into releases and usage. This is software research using a matched event-study design, not proof that document review is everyone's bottleneck. Two authors disclose paid Microsoft consulting work. NBER working-paper abstract and disclosures.

The rollout decision is narrower than a prediction about employment. If required review constrains delivery, budget that work before increasing production. Change course if a measured reduction in review effort, or a different bottleneck, makes the budget irrelevant.

Put minutes beside the launch target

Take the fictional implementation-plan team. Two reviewers each protect two hours a day, giving them 240 minutes in total. Each plan needs fifteen minutes to check the proposed configuration, verify customer constraints, and resolve corrections. The team currently receives twelve requests a day.

That consumes 180 review minutes. There is room for interruptions.

Expand the service to twenty-four requests a day and review demand becomes 360 minutes. The same people can finish sixteen plans. Eight wait each day, leaving forty unfinished after five working days, even though drafting keeps up.

Here is the complete tabletop model. All inputs are invented. Plans are equal-sized, arrive at the start of each day, and can be drafted immediately once admitted. Both queues begin empty. Every reviewed plan is assumed to meet the quality bar so that this exercise isolates capacity, not model accuracy.

The protected-capacity case is deliberately uncomfortable. It clears the review queue by leaving thirty requests outside it. That may be a defensible way to preserve time for urgent work. It is not faster service for those thirty customers.

An intake limit needs a visible consequence: schedule the remaining work, decline it explicitly, or fund more capacity. Moving the start line cannot improve the customer's experience by itself. The Kanban Guide makes workflow boundaries explicit; for this rollout, I would also measure elapsed time from the original request, before admission.

An AI reviewer has to retire work

Deep Review makes the strongest counterargument to a fixed human-review budget: perhaps review itself can become much cheaper. Some low-risk outputs may need only sampled checking. Keeping a person on every item can preserve an unnecessary step.

But a second model's comments are still work for somebody unless they remove more effort than they create. Checking a false alarm, reconciling conflicting suggestions, and confirming a proposed correction belong in the time estimate. Count that work even if it happens outside the review screen.

The ten-minute case is therefore a hypothesis. If it holds, the fictional team can deliver all twenty-four plans, although ordinary variation could still create a queue. If the actual total becomes sixteen minutes, adding the precheck has made capacity worse than the fifteen-minute baseline. Better-looking feedback is an expensive consolation.

Do not preserve a human step just because this model has a row for it. Remove or sample checks when the risk permits, then measure whether accepted outcomes remain acceptable. Review minutes are a constraint to understand, not a target to maximize.

Make the expansion earn its budget

I would start with one repeatable document type and a bounded trial. Compare the current workflow with an AI-prechecked workflow on comparable requests, using the same release criteria. Avoid giving one group all the easy cases. Decide the allowable error rate and delay before examining the results.

Record hands-on review and correction time separately from elapsed waiting time. Keep request arrival timestamps, including for work that never enters the drafting queue. Check a blinded sample against the agreed quality standard and follow released plans for costly corrections. A quick approval that sends the customer back tomorrow has not saved the promised time.

Expand only if completed, acceptable plans increase within the protected review hours, without breaching those quality and request-to-delivery limits. If work waits somewhere else, investigate that constraint instead. The simple model cannot forecast variable task sizes, absences, or error costs; the trial has to expose them.

The Exception Queue Is the Product covers what a reviewer needs once work reaches them. Adoption Is Not an AI Value Metric covers the outcome beyond usage. This budget belongs earlier, beside the rollout promise. Before approving another agent, put the extra review minutes on someone's calendar. If they do not fit, which part of the promise changes?

Source and product review: October 18, 2026. The capacity model is synthetic; no Deep Review accuracy or time-saving experiment was performed. The cover is an AI-generated conceptual illustration.