Processing invoices with AI: what to automate and what to review
AI saves the typing, not the responsibility of checking. A method to decide what to automate and what to review on each invoice.
In this article
A language model can read an invoice in seconds: issuer, tax ID (NIF), date, tax bases, VAT rates and total. What it cannot do is answer for it. Responsibility for the VAT being deductible and the entry being correct stays with the firm. So the useful question is not "does it read well?" but "what do I let it do on its own and what do I review myself?"
What data an invoice requires
Spain's invoicing regulation (Royal Decree 1619/2012) sets the content of a full invoice: number and, where applicable, series; issue date; name and tax ID of issuer and recipient; address; description of the transaction; taxable base, VAT rate and VAT amount. That list is the starting point of any automatic reading: if a mandatory field is missing, the problem is not the reading but the invoice.
What can be automated
- Transcription. Turning a photo or PDF into structured fields is the lowest-value task in the firm and the one machines do best.
- Arithmetic checks. That the lines add up to the total, that VAT matches the base and that a withholding is one of the usual ones.
- Formal tax ID validation. The check digit of a DNI, NIE or CIF is calculated; nobody's eyesight needs to be trusted.
- Duplicate detection. Same issuer, same number, or same amount and a nearby date.
- Electronic invoices. If the document carries structured data, it is read as is, without interpreting an image.
What a person must review
- The nature of the expense. Whether it is business or personal, and whether the VAT is fully or partly deductible. The VAT Act conditions deduction on the VAT being incurred in the business activity: that is a professional decision.
- Doubtful readings. A cropped photo, a tax ID with an ambiguous digit or a date whose day and month may be swapped.
- Anything unusual. An uncommon VAT rate or withholding, a disbursement or an unexpected equivalence surcharge.
- Closed periods and distant dates. The right to deduct has a time limit, and the general tax limitation period is four years (General Tax Act).
A practical review rule
- Sort by confidence. Review first what the reading flags as doubtful; the safe ones come after, and much faster.
- Review with the document in view. Comparing the field with the image is faster and more reliable than remembering what the invoice looked like.
- Approve in bulk only what is safe. An electronic invoice, or a high-confidence reading with no warnings, can be approved in a batch — always by a person.
- Leave a trail. Who reviewed, when and what was corrected. That is what lets you explain an entry years later.
- Measure accuracy. If you know which document type fails most (receipts without a tax ID, credit notes, noisy scans), you know where to put more review.
When the system learns
Repeated corrections are information. If a client has a supplier whose tax ID is misread, correcting it once and having later readings come out already corrected, with a warning, saves time without hiding the change.
Where Guilda Work fits
Guilda Work reads invoices with AI, checks tax ID, duplicates and amounts, orders the review by confidence and generates the import file for A3. The reading proposes; a person at the firm reviews, and no invoice is marked as reviewed without human involvement.
Informational content. Deductibility decisions depend on each case: confirm them in the regulations.
Informational content based on the official sources cited. It does not replace tax or legal advice for a specific case; check the data at the source before applying it.