How to Build a Human-Approval Flow

What Is the Problem?

Undesigned collaboration rots in three shapes.

Approval becomes a stamp

The human appears to “check” every output, but after dozens of daily approvals starts waving them through unread. A line that is human-supervised on paper and unsupervised in fact is born — the most dangerous state, because assurance is assumed. The stamp lives on the corpse of the review.

The human does the work twice

At the opposite pole, a human who distrusts the machine rewrites every draft from zero; the promised time gain turns negative and the invoice lands on the tool. Yet the culprit is usually neither tool nor human — where the machine’s work ends and the human’s begins was simply never drawn.

Rejections evaporate unrecorded

The human fixes an output and the flow moves on; what was rejected and why gets written nowhere. The same fault returns next week, the same fix gets made again — the flow does not learn, the human tires. An unrecorded rejection is the line’s costliest waste: effort spent, lesson binned.

Why Does It Happen?

The skipped design has three causes.

The word “together” passes for a plan

“The AI drafts, we review” sounds like an arrangement, yet contains no approval point, no threshold, no record. When the word substitutes for the design, everyone builds a private routine and a ten-person team runs ten different flows — none teaching anything, because none is shared.

Real usage already lives in this pattern

Anthropic’s Economic Index, examining millions of real interactions, finds that much of individual use is not full delegation but back-and-forth refinement loops. “Together” is already lived; what is missing is its conversion into institutional order. A habit that forms by itself does not scale until it is written.

Approval cost never enters the arithmetic

Delegation math gets done as “drafting time falls to zero”, forgetting that reading and approving take time too. Unbudgeted approval means no human capacity gets reserved for the line; unreserved capacity breeds the stamp. The honest formula: gain = old time − (machine time + approval time) — and on some tasks that difference is smaller than hoped.

How Is It Done?

The setup is three lines of drawing; all fit one page.

Split the line into three zones

The task flow is drawn on paper and painted in three zones: the preparation zone where the machine runs alone, the approval zone where the human decides, the dispatch zone that proceeds automatically after approval. An unpaintable step is an undefined step and gets clarified before the line opens. For tasks arriving from the delegability test as “together”, the approval zone sits exactly on the criterion that scored low — verification if verifiability was weak, the decision point if judgment was present.

Reduce the approval point to one question

A good approval point is not “a general look” but a single question: do the figures match the source, is there a commitment sentence, are the client name and terms right? An approver who knows what to look at decides in a minute; one who does not looks at everything or at nothing. The question derives from the task definition and is printed at the top of the approval screen — the quality of an approval equals the sharpness of its question.

Make rejection real, write the exceptions in advance

A rejection must trigger two things: the output returns to the machine with its reason, and the reason enters a one-line rejection log. The log is scanned weekly; a recurring reason gets written into the task definition or the template — every rejection thus works to prevent its own repeat, and the flow learns without tiring the human. Onto the same page, the exception rules go on setup day: which output types may pass unapproved, which need double approval, when the line halts. Unwritten, each exception moment becomes a private judgment call and the line lives its weakest version on its busiest day; an exception rule is a panic plan written on a calm day.

How Long, at What Cost?

The setup’s bill is small; the neglect’s bill is sneaky.

One drawing session per task

Three zones, the approval question, the rejection log and the exception rules — a two-hour session for one task, output one page. With the skeleton in hand, the second task’s line takes half the time; after the third, the team knows the skeleton by heart and setup becomes routine.

Approval capacity gets budgeted

Real time is reserved for the line’s human side: approvals gather into fixed windows and are read in blocks rather than dripping through the day. Scattered approval is both slow and tiring — the batched window is the cheapest antidote to the stamp.

The return reads in two lines

First, time: drafting hours fall. Second and lasting, quality: the template fed by the rejection log produces a little less error each week. The second line hides from standard reports but shows in the audit data: a falling rejection rate is a maturing line.

The Common Mistake

The flow gets built; three reflexes break it.

Giving approval to whoever is free

Approval goes to whoever knows the error’s price, not whoever has calendar space. An available but unaccountable approver is the stamp in institutional form — where there is no responsibility, there is no attention.

Widening the line silently

Neighbouring tasks get added to a built line because “they’re similar”; the approval question now fits none of them exactly and loses its edge. Every task wants its own line; the skeleton is copied, the question is not.

Turning the rejection rate into a punishment

Management that blames the approver or the tool when rejections rise will find, a month later, that not the rejections but the records have fallen. Rejection data is a diagnostic instrument; the day it becomes a disciplinary one, it goes blind.

Frequently Asked Questions