Organise the inbox before automating it

In a small accounting practice, the inbox can look like a sorting depot: messages arrive in several mailboxes, attachments have inconsistent names, and one client may send both purchase and sales invoices in the same email. An eager robot intern would gladly file everything. Without rules, it would just as eagerly file things in the wrong place.

That is the problem described in a case study published on 14 July 2026 by the European Digital Innovation Hubs network about RAMAJO ASESORES, a micro-firm in Salamanca, Spain. This is not a demonstrated return on a deployed AI system. It is a case of preparing an implementation. The distinction matters: the work produced a technical roadmap, but invoice automation had not yet entered the firm's production workflow when the case was reported.

What was measured

According to the case, staff manually handled 30–60 emails with invoice attachments each day. Two or three employees were involved, spending a total of 30–60 minutes a day on this administrative task. More than three Outlook mailboxes were in use. They had to distinguish purchase from sales invoices, extract details, and place the documents in the correct accounting workflow.

For a manager, the useful number is not a claimed AI saving—none has yet been measured—but the baseline. Assuming 22 working days per month, 30–60 minutes a day equals a modelled 11–22 hours of manual work per month. This is arithmetic based on the case's baseline, not a verified saving at RAMAJO. Some time would remain for exceptions, review and corrections.

The firm first completed two practical AI training sessions, then mapped its process with DIGIS3 and considered three implementation options. The recommended design combined Power Automate, Azure Document Intelligence and computer-vision models. The case authors estimated five to seven weeks for implementation and €0–50 of additional monthly operating cost, relying heavily on Microsoft software already available to the firm. That is a project estimate in Spain, not a universal price and certainly not a quotation for a Russian business. Directly copying it into a local budget would be an accounting trick worthy of a disappearing attachment.

Which parts need AI—and which do not

An email connector receives messages and attachments. Deterministic rules check file type, count and size, sender and duplicate delivery. A document extraction service then reads invoice number and date, counterparties, amounts and other fields. A classifier helps identify document type and transaction direction. But an authorised employee should retain the final decision on accounting entry, document status and disputed fields.

Microsoft's invoice-processing documentation lists common extractable fields, including invoice ID, date and amount due, as well as confidence scores. The vendor also notes that unusual documents may need a custom model or extra fields. A confidence score is a routing signal, not proof that an invoice is correct. Even a correctly read amount has to be checked against the source document, order and accounting rules.

For a Russian small business, a comparable design could remain inside a controlled environment: corporate mail, malware and file-format checks, local OCR or a document model, a task queue, an employee review screen and an accounting-system integration. A local language model belongs only where templates and conventional OCR cannot cope with varied wording. It should not invent a missing tax ID or amount. For regular documents, rules plus specialised extraction can be simpler, cheaper and easier to audit than an open-ended agent.

Do not confuse a pilot with an outcome

The case reports an increase in the firm's overall digital maturity assessment from 44% to 49%, and its automation-and-AI score from 20% to 36%. These are measures of skills and readiness after training and consultation, not measurements of processing time saved after a launch. The authors explicitly say invoice automation had not been implemented when the service ended. It would therefore be wrong to claim that the firm already saves 60 minutes every day.

Before implementation, agree on a labelled test set and metrics:

  • the share of messages whose attachment is actually an invoice;
  • accuracy of purchase-versus-sales invoice classification;
  • accuracy for each mandatory field, rather than a single overall AI score;
  • the share routed for manual review;
  • time from email arrival to a verified accounting-system record;
  • cost per processed document, including licences, compute, maintenance and reviewers.

Create a separate list of stop conditions: several invoices in one email, a duplicate file, a corrected invoice, a low-quality scan, a message without an attachment, a password-protected archive, a new supplier, or mismatched amount and currency. Assign an owner to each decision. Otherwise, impressive automation rates may simply mean that errors have moved further down the workflow.

Data, access and economics

A pilot needs a sample of real emails and attachments that is anonymised or otherwise authorised for processing, a target-field schema, document types, access to a test mailbox and a test accounting environment. Do not send production financial documents to an external service until the legal basis, processing location, contractual restrictions, retention periods and access rights are settled. This matters particularly for clients' financial records. An architecture appropriate to a Spanish firm using Microsoft Cloud cannot automatically be treated as suitable for a Russian organisation with different data and infrastructure requirements.

Economics starts with the cost of error. For a modelled estimate, multiply the baseline 11–22 hours per month by the share of work genuinely removed and the internal hourly cost, then subtract service, maintenance and human-review costs. If, hypothetically, half the baseline work disappeared, the gross time released would be 5.5–11 hours a month before review and exceptions. This is a scenario, not a RAMAJO result or a payback promise. At low volume with uniform documents, implementation may cost more than the time it saves; first tidying mailboxes, naming conventions and filing rules may be the better investment.

A safe first step

Take 100–200 recent messages from a single workflow, prepared lawfully for testing. Label invoices versus other attachments, mandatory fields, and the reasons employees override an initial decision. Record the baseline time per document. Then compare simple email rules and OCR with an option where a model helps handle exceptions. Do not enable automatic posting to the accounting system until accuracy has been tested on held-out examples and disputed documents require human approval.

The lesson from RAMAJO ASESORES is not that AI has already conquered accounting paperwork. The verified outcome is smaller and more useful: the firm measured a recurring task, received three alternatives and obtained a technical path before investing. A good pilot separates a promise from a working process—even if the robot intern has already prepared the folders.