What an AI agent actually does in a small company
Everything sold this year has „AI” on the box. This is not about what the technology will do in five years, but what it does today in a ten-person company — and above all where it should not be put. The most useful example from our own product is a decision not to use AI, and to read instead of guessing.
1. The difference between an assistant and an agent is one word: does
An assistant answers. An agent acts: opens the invoice, fills fields, sends an email, marks something paid. All the value and all the risk come from the same property. So the right question is not „how clever is it?” but what is it allowed to do on its own, and what happens when it is wrong?
2. Three things it does well today
Reads documents — pulls who, how much, when, which tax ID out of an invoice or a receipt. Drafts text — a dunning email, a summary of the month. Finds things — „where is the contract with that company in Cluj?”. Notice what they share: in all three, the human sees the result before it matters.
3. The question that sorts everything: what does a mistake nobody sees cost?
| Task | Fit for an agent? |
|---|---|
| Pull the amount and tax ID from an incoming invoice | Yes — a mistake is visible next to the document, immediately |
| Draft an email to a late-paying client | Yes — you read it before sending anyway |
| Compute a filing deadline | No — it must be right 100 times out of 100, and that is a rule, not an estimate |
| Decide whether an expense is deductible | No — the answer depends on context the agent does not have |
| Submit something on the company's behalf | No — these are actions you cannot take back |
4. The counter-intuitive part: most of what a small firm needs is rules, not AI
The tax calendar, payroll, the five-working-day e-invoicing deadline, the 24-month minimum wage limit — each has one correct answer, written in law. A rules engine is boring: you write it, you test it, it returns the same result every time. That is precisely why it is the right tool. Putting a language model on top of these does not solve a problem; it replaces a certainty with a probability and calls it innovation.
5. The best „AI” decision in our product was not to use AI
We need the data out of incoming supplier invoices. The reflex is a paid document recognition service. But Romanian supplier invoices almost always arrive as PDFs generated by invoicing software, and the text is already inside. It does not need to be recognised. It needs to be read.
So we read the text layer first — free, instant, works offline, and more accurate, because it does not guess characters. The paid recognition service stays for what genuinely needs it: phone photos and scans. Cheaper, faster and more correct, in that order.
6. Four questions for anyone selling you „with AI”
- What does it do on its own? If the answer includes „sends”, ask to be able to turn that off.
- Where do I see what it did? There must be a trail — and for an agent, a why.
- What happens when it is wrong? „It rarely is” means the question was dodged.
- Where does my data go? A document sent to a model in another country is a data transfer, whatever the button is called.
If one thing is worth keeping from this article: ask first whether your problem has a correct answer. If it does, you want a rule. AI is for the rest.
Sources
Not a legal article but a position, based on what we built and what we refused to build. The PDF text-layer example is the product's real behaviour. The engines that compute deadlines, payroll and tax obligations are deterministic and tested — not estimates.