A Montreal accounting firm with 12 employees reached out in November 2025. Their problem: 800 supplier invoices per month to manually enter into Acomba for their SMB clients. Each technician spent about 6 hours a week on this task. At $65/hour billable, that was $100,000 of revenue “wasted” on routine data entry.
The diagnostic
Three weeks of analysis. We looked at a sample of 200 invoices from different clients. Finding: 65% followed a predictable format (PDFs with structured fields), 25% were scanned PDFs (OCR required), 10% were edge cases (handwritten invoices, exotic formats, multiple currencies). The first 65% were automatable with good accuracy. The rest had to stay manual.
The design
Four-step pipeline:
1. Reception: a dedicated email inbox receives invoices. A filter rule sorts PDFs from other attachments.
2. Extraction: Claude Sonnet reads each PDF and extracts supplier, date, invoice number, subtotal, taxes, total, suggested accounting line(s) per the client’s chart of accounts.
3. Validation: extraction is cross-referenced with the supplier list registered in Acomba (fuzzy name matching). If confidence is below 90%, the invoice goes to the manual queue. Otherwise, it’s pre-entered into Acomba.
4. Human approval: a technician sees a list of pre-entered invoices with their sources, validates or corrects with one click, and ships to production.
The deployment
Six weeks, in three waves: first three pilot clients to calibrate prompts, then ten additional clients to validate scalability, then the rest of the portfolio. Continuous adjustments during the first eight weeks. Team training in two half-days: how to validate, how to escalate an odd case, how to add a new supplier.
Results at four months
60% of invoices are fully automated (extraction + pre-entry without human intervention). Another 25% are pre-extracted with one-click manual validation. 15% remain fully manual processing (edge cases). Average time per invoice: dropped from 7 minutes to 2 minutes (on average, including manual cases). Error rate: dropped from 4% (manual entry) to 1.5%.
Operating cost: about $280/month in tokens and infrastructure to process 800 invoices, so $0.35/invoice. Team time freed up: about 50 hours/month, or $3,250 of additional billable capacity. ROI reached in three and a half months.
The key point: the 60% automatable portion was predictable from the diagnostic. That’s why the initial investment paid off. Without that honest evaluation phase, we could’ve sold an 80% automation project and delivered 40% disappointment.