Automated invoice and document reading: what works, what fails
Automated reading works well on clean digital documents, much less on damaged scans or handwriting. Type by type: what works, what fails, and where a person must check.
Yes, automated invoice reading works on clean, regular digital documents; it becomes much less reliable on poor-quality scans, handwriting and unusual layouts. In every case, a person must stay in the loop for uncertain fields and for anything that triggers a payment. A document-reading project starts at 12,000 DH excl. VAT for a prototype measured on your documents.
We do not publish a “universal” accuracy rate: there is no such thing. The rate that matters is the one measured on your documents, field by field. What follows is a qualitative reading, document type by document type.
Updated on 17 November 2026.
How does a reading process work?
Four steps. The document arrives (email, upload, scanner). The text is extracted: directly if the PDF is digital, by character recognition (OCR) if it is an image. A model finds the useful fields: supplier, ICE (Moroccan company identifier), date, number, amounts excl. VAT, VAT and incl. VAT, line items. Finally, rules check consistency and a screen shows a person the doubtful fields next to the document. This is the setup described on our automated document reading page.
The key point: quality depends as much on step two (is the text readable?) as on step three (does the model find the right field?). An excellent model cannot read an illegible scan.
What works well?
- PDF invoices generated by software. The text is already digital and the layout is clean. This is the most favourable case.
- Repeated documents from the same issuer. A supplier that always sends the same invoice becomes very predictable, and errors are fixed once and for all.
- Fields with a checkable format. A date, an ICE, an amount, an invoice number: you can check they have the right form, and that the total incl. VAT equals the amount excl. VAT plus VAT.
- Classification by type. Telling an invoice from a delivery note or a credit note is generally reliable.
What fails, and why?
- Poor-quality scans and photos. Skewed pages, phone photos, stamps over figures, folds: character recognition gets it wrong, and the model reads what it is given. An 8 read as a 3 in an amount is the typical error.
- Handwriting. Sometimes readable, often unpredictable. A handwritten form almost always needs a human check of the written fields.
- Long or complex tables. Invoice lines over several pages, merged cells, intermediate subtotals: this is where lines get lost or duplicated.
- Multilingual documents. An invoice in French with notes in Arabic, or a bilingual contract, can be read but needs specific testing: do not assume Arabic will be read as well as French.
- “Smart” documents that need interpretation. A contract can be read (parties, dates, amounts), but judging the scope of a clause remains human work.
Table: expected reliability by document type
| Document type | Expected reliability (qualitative) | Human check |
|---|---|---|
| Digital PDF invoice, regular supplier | Good | On fields flagged as uncertain and before payment |
| Good-quality scanned invoice | Good to medium | On amounts and line items |
| Phone photo, damaged scan | Medium to low | Always on amounts |
| Handwritten form | Low to medium on handwritten areas | Always on handwritten fields |
| Contract, amendment | Good for parties, dates, amounts | Always on interpreting clauses |
| ID document | Good if the image is sharp | On numbers and dates; personal data to be handled under the law |
Digi4 estimate, qualitative, based on the nature of the documents; it does not replace a measurement. The real rate is measured on your sample during the prototype.
Where must a person always check?
On three categories of fields, whatever the measured rate:
- What triggers a payment: amount, bank details (RIB), payee. An error here costs money, not just time.
- What has legal or tax weight: ICE number, VAT, due date, contract clause.
- What the system flags as uncertain, or what the consistency rules reject (total that does not add up, unknown supplier, duplicate number).
The gain does not come from removing the check; it comes from the person no longer re-keying everything: they confirm what is certain at a glance and only correct what is flagged. The rules that do this sorting are described in AI checks and validation.
How much does it cost?
| Step | Range (DH excl. VAT) | Assumption |
|---|---|---|
| Reading prototype | 12,000 – 60,000 | One or two document types, real sample, field-by-field measurement. Source: Digi4 price list |
| Integration into your software | 18,000 – 200,000 | Validation screen, consistency rules, writing to the ERP. Source: Digi4 price list |
| Prior audit (optional) | 15,000 – 30,000 | If several processes are candidates. Source: Digi4 price list |
What makes the price vary: the number of document types, the variety of issuers, scan quality and the software to feed. The model usage cost, billed by volume by the provider, comes on top; it is measured on the sample.
When it is not the right solution
- Low volume: a few dozen invoices a month are keyed in faster than a process can be built.
- Data already structured: if your supplier can send you a file or go through an API, import it. Reading a PDF generated from data you could receive directly is a detour.
- Off-the-shelf software is enough: many accounting and ERP packages offer built-in invoice reading. For standard invoices and medium volume, it is often the most economical option. Custom work is justified when the documents are specific (reimbursement files, claim documents, business forms) or when your own software must be fed with your rules.
What about personal data?
A supplier invoice contains little personal data; an ID document, a payslip or a medical claim form contains a lot. If the model is hosted abroad, sending it is a transfer governed by Moroccan law 09-08: see our guide AI and law 09-08. The choice of model and hosting is made during scoping.
What we have built, and what we do not claim
Our published platforms, such as LA MAS (member portal, healthcare partner area and back office for staff) or Doctifile (flow of medical documents), handle documents every day. They are not automated reading projects, and we publish no rate measured at a client. On your project, the first deliverable is precisely that figure, measured on your documents. To get started: request a quote.
Frequently asked questions
Does automated reading replace data entry?
It replaces re-keying, not checking. A person validates uncertain fields and anything that triggers a payment.
Do you need thousands of documents to start?
No. A representative sample of your common and difficult cases is enough to measure quality; its size is set during scoping.
Are invoices in Arabic read?
They can be, but quality must be measured separately on those documents before it is promised.
What happens when a supplier changes its invoice?
Fields are flagged as uncertain more often and go to validation. That is why monitoring quality over time is part of the integration.
Related articles
Artificial intelligence
How much does an AI project cost in Morocco?
From 12,000 DH excl. VAT for a prototype to 200,000 DH for AI built into a business workflow. The four cost drivers: volume, model usage, integration, human review.
Artificial intelligence
AI or automation: how to tell what you are being sold
Much of what is sold as “AI” is rule-based automation, and that is not a flaw. A simple test to tell the difference, and why it changes the price and the risk.
Artificial intelligence
AI agency in Morocco: the questions to ask before you sign
Before handing an AI project to an agency, ten questions on quality measurement, data, code ownership and payment. And what a reasonable first engagement costs in DH.
Newsletter
Get our upcoming articles straight to your inbox.
