Skip to content
BCR GROUP
← All posts

Comarch ERP + AI: how we cut month-end close by 60%

by BCR GROUP
  • #AI
  • #Comarch
  • #automation
  • #case study

Starting point

The client: a trading company with PLN 8M annual revenue, two offices, a 4-person accounting team. Comarch ERP XL in place for 5 years, well-parameterised. Problem: month-end close took 12–14 working days.

Main bottlenecks:

  • categorising cost invoices (~800 per month),
  • matching invoices to purchase orders (~60% auto-matched, the rest manual),
  • verifying intercompany completeness (3 group companies).

What we deployed

Agent 1 — categoriser. AI reads the OCR-ed invoice, the supplier NIP, line items and amounts, and assigns the GL/cost-centre account. Trained on 6 months of the client's historical data.

Agent 2 — PO matcher. Links invoices to POs by NIP, date, amount and PO numbers (across various format conventions).

Agent 3 — completeness controller. Asks daily: are all invoices for the month booked? Do intercompany balances reconcile? Are there expenses without supporting documents?

Everything connected to Comarch ERP XL via API.

Results after 4 months

MetricBeforeAfter
Month-end close duration13 days5 days
Categorisation accuracy100% (manual)96% (review-grade)
PO matching60% auto92% auto
Intercompany errors3–5/month0–1/month

Month-end close from 13 to 5 days — equivalent to one full headcount saved without hiring.

What we couldn't do

  • Full VAT return automation — JPK_V7 still needs a human final check; we didn't trust AI here.
  • Classifying atypical income — EU grant income required specialised knowledge the agent lacked.
  • Full management reporting — out of scope for now (phase 2).

Costs and payback

  • Implementation: 4 months of work + ~PLN 95k.
  • Agent operations: ~PLN 3k/month (model + infrastructure).
  • Payback: 7 months (one accountant FTE saved).

What we'd do differently

  1. Clean the supplier database earlier. AI was learning on noise; accuracy was low at first.
  2. Shorter training backlog — 3 months would have been enough, not 6.
  3. Longer test phase — two months of "shadow" mode instead of one.

We do similar rollouts for accounting offices and finance teams. Every project is different, but the template works: audit → MVP → full implementation. First audit call is on us.

CallBook a call