Assisted billing

Invoice extraction is the use case that wins the matrix most often: high volume, high pain (VAT, deadlines, duplicates), data that already arrives as PDF, manageable risk if nobody posts blindly.

It is also the one most often sold as magic (“99% accuracy”). On real Spanish supplier invoices the useful accuracy is: fields well pulled plus rules that block a total that does not add up. The demo percentage does not buy coffee.

This is how we set it up so that on Monday admin reviews, rather than types.

Fields and rules (before the model)

• Issuer, tax ID, number, date, net, VAT rate, VAT amount, total, due date if present.
• Rule: net + VAT ≈ total (one-cent tolerance). Else exception.
• Rule: tax ID format and letter. Else exception.
• Rule: duplicate number+issuer in 24 months. If yes, exception.
• Rule: a new issuer is not created alone. Human.

Those rules prevent disaster even if the model misreads a 7. The model proposes. The rule decides whether it may become “pending post”. The human confirms.

The queue, not blind posting

Three trays: ok (rules pass), doubt (partial fields), alarm (broken total, duplicate, new issuer). Admin starts with alarms. The rest is confirmed in a batch. The original PDF is stored next to the id.

If the ERP has no draft, the queue lives in a sheet or board in month one. Writing straight into the books is how you dirty the quarter.

Ugly suppliers and scans

The pretty demo invoice is not the crooked photo from a freelancer. Budget a sample of twenty of your documents, not the vendor’s. Include receipts, foreign invoices, credit notes and a locked PDF. That is where you see if the case holds.

Foreign and exempt invoices have different rules. Do not mix them into pilot “success” until domestic ones measure.

Privacy and vendors

An invoice carries company data and sometimes personal data. Processor, residency, no-training, deletion. Do not upload a five-year archive “to train” on a free account. The pilot is this week’s flow, under contract.

Metrics and programme

Minutes per invoice, exceptions per week, duplicates caught, share that goes to batch. At day 30 you should see less typing and not more VAT errors. If errors rise, switch off automatic posting — if you had switched it on, which you should not have.

Fits process and management solutions. Bring weekly volume, the ERP and three ugly PDFs. That is what we quote from.

A sample of twenty: how to pick it

Include: three clean domestic invoices, three crooked scans, two credit notes, one exempt, one intra-EU, one till receipt, one locked PDF, one of your own duplicates, one new issuer, and the rest from a real week. If the vendor only lets you try the three clean ones, the percentage is useless.

Hand-label the true fields in an hour. That is the truth you measure against. Not the demo label.

If eight of twenty are real alarms (broken total, not a model whim), the case still holds: admin will go to alarms. If eighteen are alarms, the extractor is not ready or the batch is too dirty for month one.

Credit notes, exempt invoices and the pride of “99%”

A misread credit note doubles or cancels badly. An exempt invoice posted as 21% opens a tax-office problem the demo does not show. That is why they are not in month-one success. They get their own tray or stay manual until domestic invoices measure.

A vendor who insists on mixing them into the percentage is selling an average. You need trays. The average does not post.

A new issuer is created with the same controls as without AI: who they are, tax ID checked, payment terms. The extractor does not “know” a supplier because it read a PDF. Knowing is an admin act. If you skip it, the model will have created a ghost creditor.

Keep the original PDF next to the id even if the ERP “already has the scan”. When the tax office or the supplier disputes an amount, you want the file the extractor saw, not a later thumbnail. That pair (id + original) is part of the design, not an archive extra.

A 30-day review that still mixes domestic invoices with receipts and credit notes in one success rate is not a review. Split the sheet. If domestic batch-ready share is high and the rest is still alarms, you have a pilot. If everything is one blended percentage, you have a vendor slide.

Frequently asked questions

Does this replace the accountant?

No. It tidies what reaches the accountant or admin. Tax judgement stays with people.

What about till receipts?

Another flow, dirtier. Do not mix them with supplier invoices in month one.

Do we wire the bank the same day?

No. Extract and post first. Matching is the second cycle, once the books are not dirty.

Next: automating ops and legacy. Send volume and a sample PDF (no extra sensitive data).