Skip to content

How we work

Nothing runs on its own until you say it can.

Our team sets the platform up alongside yours, using your documents and your rules. Your team checks all of the AI’s work to start with, and decides when that changes. By the end of the project they are running it without us.

Where projects stall

Nobody has written your process down.

Ask three people how the work gets done today and you’ll often get three answers. What the process really does, and who decides when a document is unclear, tends to live with two or three people who have done the job for years. Most of it has never been written down. Writing it down takes your people and ours in the same room, and that’s where every engagement starts.

The usual model

A platform, an API key, and a services partner you hire separately to work out the rest.

Ours

No separate services partner. Our team does that work alongside yours, and hands over a system your people already know how to run.

Setting it up

Five things we build with your team.

Most of this comes from watching the work get done. A questionnaire won’t tell you what a reviewer checks when two numbers don’t agree.

1

How the work runs today

Every step in the order it happens, and how you tell a good result from a bad one at each point. That second part is usually the one nobody has had to explain out loud before.

2

Writing your rules down

The things your team just knows: this vendor always bills Net 45, that one adds freight separately. Written down once, then applied to every document after that, whether or not the person who knew it is in that day.

3

The screens your reviewers work in

We lay them out around the checks your team makes, then sit with your reviewers and keep cutting steps until a case opens with everything they need already on screen.

4

Naming the documents and fields

What counts as each document type, and which fields you need from it. Every check later on depends on this list, so we agree it in writing early and it stops being a moving target.

5

Where the AI stops

Which steps the AI handles, and which ones always end with a person signing off. We agree that shape before anything goes live. How much of it runs unattended is your team’s call later.

Widening what it does

Four phases, and your team sets the pace.

Everything starts at 100% review. Your team decides when a document type is ready for the next phase, and anything that slips goes back under full review.

01Check everything
02Read the numbers
03Let it run
04Keep it current
everything reviewedexceptions only

Phase 1 · Check everything

On day one, a person checks every result.

The AI does the work and your team checks all of it. Every disagreement is logged, with the document and the reason next to it. By the end of the phase your team can say which documents it handles well and which still need a person.

Every document, checked

100% human review. Slower than a demo, and it stays that way for as long as your team wants it.

Phase 2 · Read the numbers

Three numbers tell you when it’s ready.

Every run leaves an audit trail: what the model read, and what the person who checked it changed. You also get three numbers. How often the AI and your reviewer agree, how many corrections per hundred documents, and how long a review takes. They sit in your dashboard from week one, not in a slide from us.

94.2%
Agreement with reviewer
11
Corrections per 100 docs
41s
Median review time
extracted by model · 0.98
checked by S. Okafor · corrected
new rule proposed · approved

Example figures. Our customers’ real numbers aren’t public, but we’ll take you through them on a call.

Phase 3 · Let it run

Document types come off review one at a time.

Your team picks which document types can run on their own, one type at a time, and sets the level of agreement each one has to hold. If a type drops below its level it goes back to full review automatically, the same day.

Standard invoices
98.1%
Runs on its own
Credit notes
96.4%
Runs on its own
Multi-currency invoices
91.2%
Spot-checked
Handwritten remittances
78.0%
Full review

Example figures. Which types come off review first depends on your documents.

Phase 4 · Keep it current

When accuracy slips, you see it in the numbers.

Your team’s corrections keep feeding back in, so there’s less to fix each cycle. Then something outside changes: a form gains a new box and nobody tells you, or a vendor changes how it bills. Agreement is tracked on every document type, and the alert goes to your team while the drop is still small. Nobody has to be watching for it.

Agreement on one document type

Rules out of date. This vendor changed how they calculate tax.

Handover

We’re done when your team is running it.

The engagement is short, and we agree the end date before any work starts. If your team still needs us to run it after that date, we did the handover badly, and that’s on us. Three things have to be genuinely yours by then.

1.5 million activities a month run this way in production. SOC 2 Type II, and already live with customers in regulated industries.

Your rules

The corrections your team made, and the rules that came out of them. It all sits in your own account, in a form a person can read, and it still works when the model underneath gets replaced.

Your setup

Screens, document types and checks your team can change themselves, without opening a ticket with us.

Your audit trail

Every step, from intake to sign-off, on one trail, along with the numbers that show it’s still working.

What we need from you: time with the people who do this work today, and someone senior enough to settle an argument about what counts as right.

Show us the process you’d most like to stop doing by hand.

Tell us which one and we’ll walk you through how the first weeks would run, on your documents. Bring the people who do the work today if you can.