Document reading
The form already exists. Nobody should have to type it twice.
A lot of operational work is reading one document and typing it into another system: a submission into the broker's system, an invoice into the books, a floor plan into a bill of quantities.
a person signs off the totals
What this produces, shown with sample data.
The short version
It reads the document, keys the fields into the system you already use, and sends back a short list of what's missing or doesn't add up, written the way a person on your team would ask.
Where money depends on the numbers, a qualified person signs off. It saves them the typing, not the responsibility.
Where it runs
Where this kind of tool runs today, and where we'd build it next.
| Industry | Job | Status |
|---|---|---|
| Distribution | Supplier price lists into product records, at Thebo | Live |
| Insurance | Submission intake into the broker's system | Playbook ready |
| Real estate | Quantity takeoff from floor plans | Demo |
| Accounting | Bookkeeping exceptions | Playbook ready |
| Freight | Delivery proof and invoice matching | Playbook ready |
Start to finish
- 1
Write down correct
We pull a few hundred real documents, including the messy ones. Each one gets the answer a person on your team would sign. That set becomes the test it is graded on.
- 2
Build inside your tools
It works where the work already happens: WhatsApp, email, your POS, CRM or practice software. Nobody learns a new tool.
- 3
Run it with a person approving
For the first weeks a person approves what it sends. We loosen that only where the scorecard says it has earned it.
- 4
Grade it every month
The test re-runs monthly and whenever the model underneath changes. You get the score and the list of what it got wrong.
Two dates on one invoice
The test set always includes the documents that confuse people: two dates, a handwritten correction, a scan upside down. A tool that only reads clean forms isn't finished.
If nobody can write down what correct looks like, we don't build it yet.
Tell us the job your team hates. We'll tell you if AI can do it.
A 30-minute call with the people who would build it. No technical knowledge needed: just describe the job, how often it happens, and how you check it's done right.