Taking invoice data entry off an accounting firm

This case has been running for one month. I would rather tell it this way, with what is actually measured, than inflate it with numbers that do not exist yet.

~80%
Less time spent on data entry: ~94%
Field extraction accuracy: 1 month
Running at a real firm: 3
Input formats

The starting point

In every accounting firm there is someone who spends the day keying in invoices. They arrive as PDFs, as WhatsApp photos or as DIAN XML files, each with its own layout, and all of them end up typed by hand into the accounting software.

What hurt

Every invoice is transcribed field by field: tax ID, date, subtotal, VAT, withholdings, description.
No two suppliers invoice the same way, so no single template works for all of them.
A typing error is invisible when it happens: it shows up weeks later, during reconciliation.
It is work that requires no accounting judgment, yet it eats the accountant’s day all the same.

What was built

AI invoice reading

The invoice comes in as a PDF, an image or a DIAN XML file and comes out as structured data, with no per-supplier templates.

Output ready for the software they already use

Exports in the format Siigo, World Office and Helisa expect. There is no need to switch accounting systems to use it.

Human review before anything is booked

The accountant approves or corrects what the AI extracted. Professional responsibility stays with them — the AI only removes the transcription.

The time saving is a pilot estimate, not a closed measurement: that is why it says "close to 80%" and not an exact figure. When the pilot ends, the final number will be published on this page — better or worse.

I want to try it at my firm

Next case study

Marce Anahata

Wellness center · El Poblado, Medellín

Read it →