Our approval step had one flaw: you could not read the draft

We have said before that AI drafts in our back office wait for a person before they reach a client. That was true. Until the middle of August it was also less reassuring than it sounded.

The approvals screen listed work waiting for review and offered approve and reject buttons. For drafted client deliverables, it did not show the full draft. The person approving saw enough to know that something was waiting, and not enough to judge whether it was right. On 13 August we changed it. The screen now shows every deliverable awaiting review in full, scrollable, with approve, or reject with a note explaining what needs to change. Text that had been cut short in the agent's run details is now shown whole.

Nobody had been approving blind on purpose. The screen made it easy to.

How approval gates go blind

We think this failure is common, and that most firms with an AI approval step have some version of it. A few ways it happens:

The summary instead of the thing. The approver sees a title, a status and the first few lines. The error is on page three. Summaries are convenient precisely because they hide detail, and detail is where AI drafts go wrong: a figure, a name, a date, a promise.

Volume. One draft a day gets read. Forty a day get skimmed, then stamped. An approval gate that receives more work than a person can read is a delay with a button on it.

Trust that grows faster than the evidence. The first twenty drafts were fine, so the twenty-first gets less attention. That is exactly when a quiet failure, such as a model call falling back to a template, slips through.

The approval that nobody hears about. If drafts wait in a queue that nobody is told about, they either sit for days or someone builds a shortcut around the gate. In the same week we fixed the screen, we made sure approvals reach the people who handle them by email, in the app and as a phone notification, so waiting is visible.

What the memory has to say about it

The change to the screen was small. The more durable change was recording the flaw and the fix in our company memory, which our AI sessions read before they work, with the reason spelled out. The rule that comes out of it is plain: an approval step must show the approver the complete item being approved, in the form the recipient will see it. If it cannot, it is not an approval step.

That rule matters because the next approval screen we build will be built by someone, often with AI help, who was not there in August. Without the rule written down, the same convenience will creep back, because summaries look tidier on a dashboard.

There was a second benefit we had not planned for. Rejections now carry a note saying what was wrong. Those notes are corrections in the reviewer's own words, and the useful ones can go into the company memory as rules, so the next draft does not repeat the mistake. An approval step that only says yes or no teaches nothing. One that records why it said no becomes a source of rules.

A full screen is still only a screen

It will not make a tired person read carefully. A full draft on screen can still be approved in four seconds. The gate is as good as the attention behind it, and attention runs out.

Two things help with that, and neither is software. Keep the volume going through any one approver low enough that reading is realistic. And let routine, low-risk actions graduate to running on their own once they have a track record, each one written down with the reason it was allowed, so that human attention is saved for the drafts that commit money, dates or promises.

Look at your own gate this week

If any AI tool in your business drafts something a person approves, open the approval screen yourself and pretend you are the approver. Ask three questions. Can I see the whole thing the customer will see? How many of these do I get in a day? When did I last reject one? If the honest answer to the last question is "never," the gate may not be doing what you think.

Book 20 minutes at pgco.world/book.