Documents · Automation · Validation
Documents that read themselves: what it gives you, and where it still fails
An invoice arrives on WhatsApp, someone opens a file and copies numbers. Then a form arrives, and then a contract that gets forgotten in a drawer. This is repetitive work that can be reduced, but not removed.
What changed
Until recently, getting a system to read a document was clumsy and not accurate enough to rely on. In recent years document-reading models have improved clearly, including on Hebrew documents and on photos of medium quality. I use the Google Gemini API for this, and connect it to a process with validation checks.
What once needed a large project is now feasible for a small business too, but it is not a download-and-go app.
How it works in practice
You upload a document, and the system extracts fields: document number, date, amount, supplier or customer name. From there the data goes into one place - a spreadsheet, accounting software or a CRM - with a link to the original file, so it can be checked later.
Typical cases: supplier invoices that arrive through several channels, inquiry forms that become a customer record, and contracts from which payment, renewal and notice dates are extracted.
Where it breaks
A skewed scan, a blurry photo, a table split across two pages or a column read incorrectly. The model can return a number that looks plausible and is wrong. Without a check, the error enters the system and nobody knows.
That is why layers that have nothing to do with AI are built around the model:
- Validation: is the amount positive and plausible for this supplier, is the date within a sensible range, does the VAT amount match the total.
- Human approval: a document that failed a check is flagged, not accepted quietly.
- Tracking: which documents were captured, which failed, and what happened afterwards.
Without those three there is no system, only a demo.
Who it fits
A business that receives documents regularly: a shop with supplier invoices, a clinic with forms, an office with contracts, a contractor with payment demands. If documents arrive rarely, there is probably nothing to fix here.
What is not promised
There is no promise of zero errors, and no promise that nobody will check. What can be built: less typing, fewer manual copies, and a clear flag on whatever needs review. If someone offers you a system that reads everything by itself with no setup and no checks, that offer deserves a closer look.
Document processing is not included in any of the three packages; it is a separate scope priced by the number of document types and the checks required.
Want to see how this works in your business?
A short call, no obligation. We'll figure out where leads fall through: and make a decision.
Check fit for your businessRead also: How to Check If ChatGPT Recommends Your Business · What real lead automation looks like, end to end