WholesaleGoal: cut operational costs and time14 weeks

Supplier invoices and customer orders, entered into the ERP without retyping

A Polish FMCG distributor — 350 people, 4,000 products, about 2,600 supplier invoices and customer orders a week — was growing faster than its back office could type.

Client: FMCG distributor, Poland
Their team: CFO (owner of the goal), back-office lead, 2 key users, IT admin
Our team: Engagement lead, 2 engineers, part-time designer
Timeline: 14 weeks to production · in run & care since

11 → 2 min

per document, from inbox to ERP

78%

of documents go through without a human touch

4.5 people

moved from data entry to supplier management

5 months

to pay back the project

The business goal

Handle about 40% more volume next year without hiring more people into data entry — and stop paying suppliers late because invoices sat in a queue.

Agreed in the outcome brief

Criteria: Minutes of manual work per document
Baseline: 11 min — measured over 3 weeks on 1,900 documents
Target: Under 3 min, with fewer than 1.5 errors per 100 documents (was 2.3)
Pilot exit: Continue only if 60% of real invoices go through untouched by week 6

Where it was

Documents arrived through five channels: PDFs by email, paper scans, EDI from two large suppliers, customer orders in Excel, and photos sent over messengers. Six people retyped them into the ERP; an accountant checked a sample. Every Monday started with a two-to-three-day backlog.

5 channels

email · scans · EDI · Excel · photos

6 people retyping

11 min per document

ERP

2.3 errors per 100 docs

What we did

1Outcome brief · 2 weeksMeasured the baseline on 1,900 real documents. Supplier invoices were 60% of the volume, so they went first. Agreed the target and the exit rule for the pilot.
2Pilot · 6 weeksExtraction checked against purchase orders and supplier data; a simple queue where people approve only the exceptions. Two weeks in parallel with manual entry to compare. Week 6: 64% went through untouched — above the exit rule, so we continued.
3Production · 6 weeksCustomer orders added. Posting straight into the ERP through its API, audit log for the auditors, accuracy monitoring per supplier, training for the team.
4Run & improve · ongoingMonthly review of the number with the CFO. The system learns new supplier layouts from people’s corrections: 64% at launch, 78% by month four.

What did not work

Handwritten delivery notes stayed below 70% accuracy, so we left them manual on purpose — about 5% of volume. And for the two largest suppliers, switching them fully to EDI turned out cheaper than any AI. Not everything needs a model.

Who did what

The client’s team

Owned the target, cleaned up supplier master data, changed two approval rules, ran the parallel checks.

Our team

Built and ran the system, measured the number every week, raised risks early.

Together

A 30-minute review of the number every week — and the go/continue decision after the pilot.

Before and after

NumberBeforeAfter

Minutes per document112

Without a human touch0%78%

Errors per 100 documents2.30.6

Monday backlog2–3 dayssame day

Late-payment feesbaseline−85%

Under the hood

Email and scan intake → OCR for scans → LLM extraction into a fixed schema → checks against purchase orders, supplier master data and VAT rules → confidence threshold per field → exception queue (web app) → posting via the Comarch ERP API. Evaluation set of 1,200 labelled documents, re-run on every change. Azure, EU region; documents never used for model training.

We did not buy an AI project. We bought back four and a half people’s time — and they now negotiate with suppliers instead of typing.

CFO, FMCG distributor · name withheld at the client’s request