Case 02 · Automated document review

Somebody finally read the fine print.

Before anyone signs, a manager works through a long checklist on every booking, comparing the partner’s paperwork to your own records. It’s slow, mistakes slip through, and it eats their best hours. This began as a sketch of a tireless second set of eyes: gather both sides, compare every line, tick what checks out, and leave the manager only the calls that need them. Today the system gets to every booking in the queue first.

the first sketch↓ what it became

13:45 p.m.

A booking is waiting for sign-off.

Maya finished Daniel’s cottage booking and dropped it in the sign-off queue with the partner’s confirmation attached. A manager will get to it eventually, along with forty others.

Designed forFitting into the queue the managers already use. Nobody has to learn a new tool.

23:45 p.m.

It gathers both sides of the story.

Our own record on one side, the partner’s confirmation on the other. There’s already a problem: one of the attachments belongs to a different booking.

Designed forReal-world mess. Paperwork arrives in every format, and the wrong file gets attached more often than you’d think. It says so once instead of reporting twenty false errors.

33:45 p.m.

It reads the paperwork.

The AI pulls out what matters: who’s staying, the dates, what’s been paid, the cancellation terms. Then the system works out the total and the payout itself, so it has its own numbers to compare. The AI reads. It never decides.

Designed forKeeping the AI in its lane. Reading is what it’s good at; the judging happens separately.

43:46 p.m.

It compares every line.

Name, dates and total all match, even with Daniel’s middle initial thrown in. The deposit doesn’t: we say $200, they say $400. The cancellation note is missing. And the payout is $12 apart, which is a manager’s call.

Designed forJudgment without false alarms. It works from a playbook of each partner’s quirks, so a middle initial isn’t a mismatch, and anything it can’t confirm is said out loud, never quietly passed.

53:46 p.m.

It leaves the real decisions to the manager.

The findings go straight onto the booking, and the boxes that passed get ticked. The manager gets one short email: two things to fix and one to decide.

Designed forRespecting people’s attention. Clean bookings stay quiet, and only real issues reach an inbox.

6later that week

Then it keeps score.

Once the booking is signed off, it checks what actually changed and keeps track: the deposit and the note were fixed. Over weeks, that becomes a scorecard for each agent and a short list of what the team misses most.

Designed forCoaching, not blame. Something only counts as a miss if it really had to be changed, and nobody is scored while the system is still learning.

What it took to make it hold up.

Checking one tidy booking is easy. Checking every booking, in every format, every hour, without crying wolf, is the real job. Three things shaped every decision.

Accurate

The AI reads, but the math and the rules are handled by the system itself. Every finding shows its working, so a manager can see exactly why it was flagged.

Robust

We planned for odd formats, missing files and busy hours. If a booking can’t be checked, it’s tried again and flagged, never marked clean.

Learns as it goes

Managers leave a note on any item to say how it should have been handled. Those notes become written rules, and scoring stays off until the team says training is done.

Managers stay in charge.

One switch decides whether findings are written automatically or wait as drafts a manager can edit.
Any booking can be re-checked with one click after a rule changes, or skipped.
Agent scores stay switched off until the managers turn them on.

Same idea, other places.

It fits anywhere an outside document has to agree with your own record, and today a person checks it by hand.

Invoices against purchase orders
Signed contracts against the agreed quote
Shipping documents against the order
Timesheets against the schedule