Operations · 12 October 2026 · 4 min read

When human review becomes the bottleneck.

AI prepares the work in seconds. It then waits two days for approval. Where did the time saving go?

“A human will check it” sounds reassuring. But it leaves several operational questions unanswered. Which human? Checking what? With which information? And what happens while the work waits?

As AI produces more drafts, recommendations and proposed actions, a growing organisation can create an approval queue faster than it creates useful capacity. The founder still carries the decisions. They simply arrive in a different format.

Give review a specific job

A reviewer needs a decision to make, not a pile of output to inspect. Checking whether a customer name matches a record is different from deciding whether to offer an exception to commercial terms. The first can often be handled by a fixed check. The second may require judgement and authority.

Before adding a review step, ask what it protects: money, a customer commitment, confidential information, quality or something else. Then define the evidence the reviewer needs and the point beyond which work must wait.

For a customer proposal, that might mean confirming the scope, approved price and delivery capacity before sending. An AI draft can organise the evidence and highlight missing information. It should not infer permission from the fact that a similar proposal was approved last month.

A real-world pattern: approval belongs to a version

Software teams have a useful example. GitHub’s protected branch settings allow teams to require automated checks and approval from designated code owners before changes are merged. They can also dismiss an earlier approval when the proposed changes are updated.

The operational lesson travels beyond software: approval should refer to the actual thing that will be used. If an AI revises a customer proposal after someone approves it, the earlier approval does not necessarily cover the revised price, promise or wording.

You do not need a software development system to apply that principle. Keep the reviewed version identifiable. Record who approved it and what they approved. Decide which later changes require another review. Otherwise, a tick beside “checked” can hide an unreviewed commitment.

Make the queue visible

Human review has a cost even when the reviewer spends only a few minutes. Work may wait until tomorrow. Another colleague may chase an answer. A customer may receive a late response.

Measure waiting time alongside review time. Record why items return for correction and whether the reviewer has enough context to decide. A short approval meeting is not efficient if every item needed three messages beforehand.

Do not remove a necessary control just to improve the numbers. Improve the handoff first: complete evidence, a clear owner, a response expectation and a named backup with the right authority. Route exceptions deliberately. Sample completed work where appropriate, while keeping required approval before consequential actions.

In my Blueprint, these choices become part of the operating design: what AI prepares, what fixed rules verify and what an authorised person decides. Build tests that arrangement against the starting process, including correction costs and human effort. Faster drafting alone is not evidence of better operations.

All insightsTell me about it