Discover
We look at a sample of real documents and the system they go into, noting the formats, fields, and common problems.
EXAMPLE PROJECT
AI that reads invoices, receipts and orders
Retyping invoices, receipts, or order emails is slow and error-prone. AI can read the documents and prepare the entries, with a person checking anything unclear.
01 · THE PROBLEM
The problem
Documents arrive as PDFs, scans, photos, email bodies, or spreadsheet exports. Someone reads each one and types the supplier, dates, amounts, and line items into accounting or order software.
The work piles up at busy times, and a small typo can cause a real problem later. Layouts vary by sender, which is where fixed templates break.
02 · SIGNS
Signs you have it
- Someone spends a regular part of the week retyping documents.
- Documents come in many layouts, so simple templates do not keep up.
- Typos in amounts, dates, or reference numbers lead to corrections later.
- The destination system has an import option or a supported connection.
03 · A FIRST VERSION
What a first version looks like
An example of a focused scope. Your first version, price, and timing are agreed in writing for your project.
- One document type, such as supplier invoices, receipts, or customer orders.
- AI extraction of the agreed fields as documents arrive.
- Checks against rules you set, such as a known supplier, a matching total, or required fields.
- A review screen for anything unclear or low in confidence, where a person confirms or corrects it.
- Hand-off to your system through a supported import or connection.
- A record of what was processed and what was changed by hand.
04 · HOW WE’D DO IT
How we’d do it
The same four steps as our approach, applied to this kind of project.
Design
We agree on the fields, the checks, what counts as an exception, the AI provider, and data handling. We also check whether plain rules would do without AI.
Build
We build and test against the agreed sample, including the messy documents, and compare results with the manual process before launch.
Improve
We confirm the agreed criteria and review the exceptions together. More document types are scoped separately.
05 · QUESTIONS
Common questions
Related: Workflow automation · AI integration
What if the AI reads a number wrong?
It can. Fields are checked against your rules, and anything that fails or looks uncertain goes to the review screen instead of your system. Testing on real documents before launch shows where it struggles.
Will it handle every document?
No. The goal is to prepare the routine ones and make the unusual ones easy to spot and fix by hand.
Can it put entries straight into our accounting software?
It depends on whether the software offers a supported import or connection. We check this during discovery.
Where do our documents go?
To the AI provider agreed in design, ideally under your business’s own account. What is sent and how long it is kept are agreed first.
LET’S BUILD SOMETHING USEFUL
Have a version of this problem?
Tell us how the work happens today and what a useful first version would change. A plain-language description is enough to start.