Why model price is not the cost of automation

When estimating document-processing economics, teams often compare an API price with an employee’s salary. But a model does not retrieve an invoice from email, investigate a mismatch, or take responsibility for posting it. For a business, the useful metric is the cost of one correctly processed document: extraction, model, integration, manual review of exceptions, and error correction.

That is not a reason to avoid automation. It means the calculation should cover the entire document journey, not just one line in a pricing table.

Define the process boundary first

Consider incoming supplier invoices. Before a pilot, agree on what “processed” means: supplier, number, date, currency, total, VAT, and line items are extracted; the data passes validation; a record reaches the accounting system; and disputed fields receive an owner and status.

Separate the flow into three outcomes:

  • **accepted automatically** — critical fields were extracted and passed business checks;
  • **reviewed by a person** — low confidence or conflicting rules routed the document to an employee;
  • **rejected / corrected** — wrong document type, duplicate, missing purchase order, or a total that does not reconcile.

Count all three groups. If a report shows only documents the model accepted automatically while the work on the remaining exceptions disappears, the savings will be more accounting than real.

A model calculation for 1,000 invoices per month

This is an illustration, not company data or a price forecast. Replace the assumptions with measurements from a pilot:

  • manual handling takes 4 minutes per invoice on average;
  • with automation, employee time averages 1.5 minutes per invoice across the entire flow, including exceptions and corrections;
  • fully loaded labor costs ₽900 per hour;
  • model and OCR processing cost an assumed ₽6 per document;
  • infrastructure and service cost ₽3,000 per month;
  • support and process oversight take 4 hours per month at ₽900 per hour;
  • integration development costs ₽120,000 and is spread over 12 months for comparison.

Time freed: 1,000 × (4 − 1.5) / 60 = 41.7 hours. Its estimated value is about ₽37,500. Ongoing costs are ₽6,000 for processing, ₽3,000 for infrastructure, ₽3,600 for support, and an assumed monthly share of ₽10,000 for implementation. That leaves about ₽14,900 per month before taxes, cost of capital, and the impact of errors.

Under the same assumptions, simple break-even is about 527 invoices per month: each document provides ₽37.50 in estimated labor-time value and incurs ₽6 in variable costs, while fixed monthly costs total ₽16,600. This is not a universal threshold: changes in review time, labor rate, or error costs will move it.

One more important caveat: freeing 41.7 hours of capacity does not automatically reduce payroll. It becomes cash savings only if the business avoids overtime or hiring, or reallocates the time to valuable work.

Where confidence helps—and where it can mislead

Field confidence can help route extracted values, but it is not a guarantee of correctness. Microsoft describes confidence as an estimate and recommends human review for accuracy-critical scenarios. Its Content Understanding documentation calls example thresholds for critical, important, and non-critical fields illustrative: teams should determine thresholds experimentally for their own workflow.

Set separate pilot rules:

  • for total and VAT, extract the fields, reconcile arithmetic and purchase-order data, and stop automatic posting if they conflict;
  • for invoice number and supplier, compare against the register and known identifiers;
  • for comments, a lower threshold may be acceptable if the field does not trigger a financial action.

Measure more than average confidence: track incorrectly accepted documents, manual-review share, employee minutes per exception, and correction cost. A single amount error can cost more than several minutes of manual work.

Run a short pilot

Use several hundred previously processed invoices, including scans, different layouts, credit notes, and supplier errors. Compare system output with approved data without changing the live posting process.

Over 2–4 weeks, record:

  • manual processing minutes before and after;
  • the share automatically accepted versus sent for review;
  • accuracy of critical fields separately from non-critical fields;
  • duplicates, corrections, and stopped postings;
  • variable OCR, model, and storage costs;
  • time spent supporting the integration and updating templates.

Recalculate the break-even point using the measured exception rate. If the system reads typical invoices well but every unusual document still requires lengthy manual work, it may be a useful tool for part of the flow without being full automation.

Practical takeaway

Compare the cost of a correctly completed process, not “token price” against salary. Start with one document type, establish time and error baselines on a representative sample, then add OCR, the model, and validation one step at a time. Let the AI intern sort the paperwork; the process owner should still keep the calculator nearby.