First pass, human approve

Why it wins

1. ·The value arrives immediately. Reading and confirming a completed draft is several times faster than producing it, so the saving is real from week one even though nothing is fully automatic.
2. ·Failure is bounded. A wrong draft is corrected in ten seconds. A wrong posting is a reconciliation and an apology.
3. ·It generates exactly the data you need. Every correction is a labelled case, and the correction rate is the number that tells you when to automate.
4. ·It is adopted. Operators asked to check something are collaborators; operators told a system has replaced their judgement are not, and their cooperation determines whether the project survives.

It also makes the eventual automation argument evidential rather than aspirational. "This category has run at a 1.2% correction rate over 800 cases" is a much better case than a projection.

Make the review genuinely fast

The whole pattern rests on review being much cheaper than doing. If it is not, you have added work, and people will route around it.

1. 1.Pre-fill everything, including the fields the model is unsure about, marked as uncertain. An empty field costs more attention than a wrong one.
2. 2.Put the evidence next to the claim. The invoice page beside the extracted total, not behind a link.
3. 3.Sort the queue by how likely it is to need attention, so the doubtful cases are dealt with while the reviewer is fresh.
4. 4.Support bulk approval for a filtered set. Twenty routine cases from a known vendor should be one action.
5. 5.Keyboard first. Operators doing hundreds a day will never use a mouse, and the difference is a factor of several in throughput.

Watch the correction rate, by category

This single number drives everything that follows, and it must be read per category rather than in aggregate.

Under 2% for 500 cases

A candidate for automation with sampling. Take it to the ratchet you agreed at the start.

2% to 10%

Working as intended. Keep the gate, and read the corrections for a pattern worth fixing.

Over 20%

Something is wrong with the specification, not with the model. Go and watch somebody do the work.

Falling to zero suddenly

Almost always rubber stamping rather than sudden excellence. Sample and check.

Moving off it, deliberately

The pattern is a starting position, not a destination. Move off it one category at a time, on evidence, with the ratchet agreed in advance.

1. ·Automate the category, not the workflow. Known vendor plus standard format plus under the threshold, that combination goes automatic; everything else stays gated.
2. ·Keep sampling the automated stream at a rate you can sustain, one in twenty is typical. Without it you lose your only quality signal.
3. ·Keep the correction path open and obvious. An operator who spots a bad automated posting must be able to fix it in the same interface.
4. ·Be willing to go back. Re-gating a category after a bad week is a normal operational action, not an admission of failure, and treating it that way keeps everyone honest.