Why the model tariff is not the process economics
A model example in the Telegram note estimated about RUB 14,900 remaining per month for 1,000 invoices after an assumed amortisation of implementation. This is neither a quote nor a forecast for every company: the result depends entirely on the assumptions. Recalculate it using your own invoices, labour costs and processing rules.
The API price alone does not describe the process. An invoice must be recognised, its fields extracted, matched to an order or supplier, checked for required details, approved and transferred to accounting software. An error can send the document back for manual work, delay payment or create an incorrect entry. In finance, “the model answered” does not mean “the invoice was processed.”
How to calculate the cost of a correct outcome
Split the workload into three groups: standard documents, low-confidence documents and exceptions requiring a person. For each group, measure the share accepted without correction and the review time. Then include all costs:
- page recognition and field extraction;
- model calls, retries and error handling;
- integration with ERP, accounting and document storage;
- manual review, correction and escalation;
- maintenance of templates, access controls and monitoring;
- the cost of delays or incorrect postings.
A useful unit is the cost per document accepted without another manual pass. The numerator includes monthly operating cost and the allocated implementation cost; the denominator counts only documents that passed the required checks and actually entered the books. If employees review almost every model output, the savings may be far below expectations.
Confidence is not a substitute for control
Microsoft’s Document Intelligence documentation describes a confidence score as an estimate of the model’s confidence in an extracted field. It suggests using the score to decide whether to accept a result automatically or send it for review. A confidence score is not a guarantee of correctness, and not every field necessarily has one. Check not just extraction output but also scan quality, document type and errors on a labelled sample.
Set separate thresholds and validation rules for invoice amount, due date, tax ID and bank details. If confidence is low, route the invoice to a person rather than letting it enter accounting by default.
A small pilot before automation
Take 100–200 de-identified invoices from different suppliers. Record the correct fields, manual processing time and exception type for each. Run the model in shadow mode: compare suggestions with the human result, but write nothing automatically. Calculate per-document cost and correction rate by supplier and scan type.
The Telegram note cited an estimated monthly remainder of about RUB 14,900 in one scenario, but without the full calculation assumptions the figure cannot be transferred to a company budget. Reconstruct your own assumptions first, compare the scenario with manual processing and only then decide whether to buy or automate invoice posting.
