In almost every Colombian accounting firm there is one person — sometimes the most junior, sometimes the most expensive — who spends the day typing invoices into the accounting software by hand. Tax ID, date, number, subtotal, VAT, withholdings, description. One at a time.
It is work that requires no accounting judgment, and that is exactly the problem: it is being done by someone who has plenty of it.
Why it is so hard to automate
If every invoice arrived in the same shape, this would have been solved twenty years ago with a template. The real problem is the variety:
- The large supplier sends the e-invoice XML, which is the clean case because the data already comes structured.
- The mid-sized supplier sends a PDF with its own layout: the subtotal in one corner, the VAT in another, the description in a table that changed last month.
- The client sends a photo taken with a phone, crooked, of a crumpled piece of paper, over WhatsApp, on a Sunday.
Any solution built on per-supplier templates breaks on the third new supplier. And there are new suppliers every month.
What changed with AI
Current extraction models do not depend on where the data sits on the page: they understand it from context, the way a person does. That means there is no template to configure per supplier. The same logic reads the XML, the PDF and the photo.
What remains non-negotiable is human review. A system that books entries with nobody looking is a system that will one day put an error in your filing and you will find out three months later.
The five questions to ask any vendor
If you are evaluating a tool — mine or anyone else’s — these questions separate the serious ones from the ones that merely sound good:
- What does it do when a field is illegible? The right answer is “it flags it.” If it guesses or fills in a likely value, that is an error you will discover during reconciliation.
- Does it export to the software I already use? Siigo, World Office and Helisa cover most of the Colombian market. If the tool forces you to switch accounting systems, the real cost is not the license: it is the migration.
- Does it need a template per supplier? If the answer is yes, every new supplier will be a support ticket.
- What happens to my documents? Your clients’ invoices are sensitive information. Ask where they are processed, how long they are stored and whether they are used to train models.
- Can I test it with my own invoices before paying? This is the one that filters hardest. A demo with sample invoices proves nothing: sample invoices are always the pretty ones.
What to actually expect
Be suspicious of round numbers. Accuracy depends on the quality of your documents: a DIAN XML file reads almost perfectly, a blurry photo of a faded thermal receipt does not.
In the pilot I have running with a firm in Medellín, field extraction sits around 94% accuracy, and we estimate the time saving at close to 80% — but that last figure is still an estimate, not a closed measurement, and I say it that way on purpose. When the pilot ends, the number will be published, better or worse.
What does not get automated
It is worth saying plainly, because anyone promising otherwise is selling you smoke:
- Classification that requires judgment — which account, which cost center, which treatment — stays with the accountant.
- Professional liability before the tax authority cannot be delegated to software.
- Decisions on edge cases that a human resolves by asking.
What disappears is the transcription. Which is no small thing: it is the part of the job nobody studied five years to do.