Approval Is the Product

The autonomy pitch is seductive and mostly dishonest. "Set it and forget it" sells because it promises the thing everyone actually wants, which is not automation but absence, permission to stop thinking about the task. Then the first hallucinated customer name goes out under a founder’s byline and the whole category loses a year of trust.

The interesting realization is that the approval step, which everyone treats as the concession they had to make, is the highest-value surface in the system.

What a review actually generates

Every time a human looks at a draft, they produce a judgement that cost real expertise to form. Most products throw it away. They record approved or rejected, which is one bit of information from an expert who just spent ninety seconds reasoning.

Capture the reason instead and you have a labelled dataset of your own quality standard, produced free, in the normal course of work.

• ·wrong_claim. The fact is not true, or not ours to state.
• ·not_my_voice. Accurate but reads like someone else wrote it.
• ·no_edge. True, safe, and pointless.
• ·bad_timing. Right piece, wrong week.
• ·too_close. We said something like this recently.

Five buttons. Six weeks later you can rank your failure modes and you know precisely what to fix. Ours, embarrassingly, was no_edge at 41%. The system was producing correct, publishable, forgettable work. No amount of model upgrade fixes that; it is a brief problem, and we would never have known without the labels.

Approval fatigue is a UI bug

The standard failure is a queue of twelve drafts in a slow interface at 6pm on a Friday. Everything gets approved, because rejecting requires effort and approving requires none. Then quality collapses and someone blames the model.

Things that measurably fix it:

1. Pre-sort by confidence. Show the two the system is least sure about first, while attention is fresh.
2. Show the evidence inline. The proprietary fact the draft rests on, with its source, next to the claim. Verification should take three seconds, not three tabs.
3. Diff against the last version. Reviewers read a full draft once; after that they should only read what changed.
4. Cap the queue. Ten items maximum. An eleventh means the system generated more than the human can honestly assess, which is a throughput bug.
5. Make edits first-class. Most rejections are 80% right. An edit captured as a diff is a better training signal than a rejection.

Autonomy you earn

The right shape is graduated. Start with review-everything. Track agreement rate per content type. When a category runs above 95% approval over fifty samples, promote it to auto-publish with post-hoc review. Anything that dips gets demoted automatically.

That is a policy, not a slider, and it means autonomy expands where the evidence supports it instead of where the marketing does. More on the framework: autonomy levels.

The uncomfortable implication

If approval is where the product is made, then the reviewer is staff, not overhead, and the honest sales pitch is not "we replace your content person." It is "we make one content person as productive as five, and they still read everything."

Less exciting on a landing page. Considerably more likely to still be true in a year. See also: nobody wants another copilot.

More on approval design, including the queue layouts we tested, at Alex Cinovoj↗ (opens in a new tab). The build side of it happens at TechTide AI↗ (opens in a new tab), where approval is the first screen we design, not the last.