When a small company says "let's do something with AI", the first thought is almost always a chatbot on the website or an "assistant" that drafts emails. But the fastest payback from AI in an SME usually sits somewhere completely different and far more boring: in the pile of supplier invoices someone retypes into the ERP by hand every morning.
A chatbot looks great on a slide. Pulling data out of invoices is dull as dishwater — and that's exactly why it works. It doesn't ask the customer to trust a machine in a conversation. It only asks the machine to read a PDF as well as a tired bookkeeper at 3:30 p.m. does, but a hundred times faster and without a typo in the tax ID.
This is the kind of AI we recommend SMEs start with: narrow, measurable, with a human at the end of the process. Not "digital transformation", just one concrete task that's eating half a full-time job today.
Where AI actually earns its keep in an SME
The best AI use cases in a small company share three traits: there's a lot of it, it's repetitive, and the result can be checked. Supplier invoices fit perfectly, but they're not the only example. The same mechanism — read the document, extract the data, push it into the system — works across a whole family of boring paperwork:
- supplier invoices (PDF, scan, photo from a phone),
- goods-received and delivery notes,
- customer orders arriving by email in any format,
- bank statements to reconcile against receivables,
- transport documents (CMR, waybills).
The common denominator: today someone opens a file, reads it with their eyes, and retypes the numbers into another window. The error is costly but detectable — and that's the key. A chatbot holding a conversation with a customer runs the opposite way: the output is open-ended, hard to measure, and a slip-up damages your reputation before anyone notices. So we say start with documents, not conversations.
How it actually works — and where it breaks
Behind "AI reads invoices" there are four steps. First, OCR — turning the pixels of a scan into text. Second, an extraction model that recognises this string of digits as a tax ID, that one as an invoice number, and "12,345.67" as the gross amount. Third, validation: does the tax ID match a supplier in your master data, do the line items add up to the invoice total, does the VAT compute. Fourth, writing to the ERP — usually through an API.
And here's the thing AI vendors don't say at the demo: the model doesn't understand the invoice, it recognises patterns. The more typical the document, the better it works. The more unusual, the more often you have to fix it by hand. It breaks on crooked and faded scans, on handwritten notes, on a new supplier with an unfamiliar layout, on foreign documents, on a photo taken at an angle in the warehouse.
So the sensible metric is not "does it work" but what percentage of invoices go through untouched. Early on, with repeat suppliers, that's realistically 70–85%. The rest goes to a person to fix — and that's how it should be. Anyone promising 100% automation from day one is selling you a demo, not a deployment.
Example: 900 invoices a month
At one of our clients — a 22-person building-materials wholesaler — accounting retyped roughly 900 supplier invoices a month into Comarch Optima by hand. Each invoice took about 3–4 minutes: open the PDF or scan, retype the tax ID, number, date, line items, net and VAT amounts, link it to the right order, save. Two people in accounting were spending around 50 hours a month on this — and mostly at month-end, in a rush, which is exactly when mistakes are easiest.
We rolled out invoice data extraction — OCR, a model recognising the fields, validation (matching the tax ID to the master data, checking line-item totals and the invoice number against duplicates) and writing into Optima through the API. The human stayed at the end: every invoice lands in an approval view, fields already filled in, with anything the model wasn't sure about highlighted.
Handling time per invoice dropped from about 3.5 minutes to roughly 40 seconds — because "retype everything" became "check and approve". Fifty hours a month turned into about twelve. The roughly 38 reclaimed hours went into chasing overdue receivables, which there had physically been no time for — and that, as it happened, earned the company more than the typing savings themselves.
Worth stressing: nobody was let go, and that was never the goal. What disappeared was the overtime scramble at month-end, and the number of wrongly keyed amounts — the kind that used to surface only at correction time — went down.
What not to automate blindly
The most expensive mistake in a rollout like this is taking the human out to make it "fully automatic". A supplier invoice posted automatically with the wrong VAT, or duplicated, is more expensive than an invoice retyped slowly — because you pay for it in corrections, filings, and a bookkeeper hunting a week later for where the discrepancy came from.
A few rules we always apply:
- Human at the gate — permanently, not just at launch. The model proposes, the person approves. This is not a transitional phase.
- Hard validation before saving. Line items must add up to the gross, the invoice number is checked against duplicates (OCR can read the same file twice), the tax ID is matched to the master data.
- A caution threshold. Invoices above a set amount, or from a new and unknown supplier, always go to manual review regardless of how confident the model is.
- Don't start with the hardest documents. Foreign, handwritten, atypical invoices — leave them for last, or with a person.
What needs to be in place before you start
AI doesn't fix mess — it amplifies whatever you already have. Before you let a model loose on your invoices, three things have to be in order.
First, an ERP that exposes an API. Comarch Optima, enova365 or Subiekt can be integrated; if your system offers no interface at all, that's what we solve first, not the OCR. Second, a clean supplier master record. If the same supplier appears under three names and two tax IDs, automation will replicate that chaos on every single invoice. Third, a defined approval process — who signs off on a cost, and at which stage. If invoices circulate by email today and "whoever's around approves them", we sort out the process first and only then put AI into it.
It's the same truth as with any sensible AI rollout: clean data, integrated systems, a smooth process. Without that, invoice extraction is just a faster way to enter errors.
Where to start
If your accounting team is drowning in invoices at month-end, do it in this order:
- Measure today's state. How many invoices a month, how many minutes each, who does it. Without that number you'll never prove AI achieved anything.
- Pick one document type and one or two of your biggest suppliers. A repeatable layout means high accuracy from day one and quick proof that it works.
- Run a 4–6 week pilot with full human review. Measure what percentage of invoices went through without a correction.
- Only expand once pilot accuracy is stable — more suppliers, more document types.
- Keep a human at the end forever. The goal is less retyping, not zero control.
This isn't a quarter-long project or a six-figure investment. It's a narrow, measurable improvement to one process — exactly the kind of AI that actually pays off in an SME, instead of ending up as a slide in a deck.
If you want to check whether the numbers add up at your place, start by counting the hours. And if you'd rather have someone walk through it with you — the first call is free and takes half an hour.
The best AI in a small company doesn't hold conversations. It retypes invoices — faster, cheaper, and without a typo in the tax ID.