How Croatian SMEs Can Use AI Without Hiring a Tech Team

A client in Split forwarded me a quote last winter. A vendor wanted €40,000 to “implement AI” across his 18-person firm. The deck had a robot on the cover. The owner asked me one question: would it answer the emails his office manager spends two hours a day on?

It wouldn’t. Nobody in the deck had asked what she actually did for those two hours.

That’s the whole problem. Most AI sold to Croatian SMEs is sold as a thing — a platform, a license, a robot on a cover slide — to a business that doesn’t have a problem a platform solves. It has a person doing the same task four hundred times a month, by hand, from memory.

What the hype gets backwards

The pitch starts with the technology and works backwards toward your business. It assumes you have someone who can integrate it, maintain it, and notice when it quietly stops working. A 5-50 person owner-led firm has none of those people. The owner is the someone. And the owner is already doing three jobs.

So the honest starting point isn’t “what can AI do.” It’s narrower, and more useful: what can a non-technical owner put to work this quarter, without a developer, without a €40,000 license, without a robot on the cover. The answer is smaller than the hype and larger than the skeptics think.

What it genuinely changes

Four kinds of work, in my experience, are where AI earns its keep in a small Croatian firm.

The repetitive. The office manager answering the same five questions about delivery times, VAT, opening hours, parking. The same reply, retyped, with small variations. This is exactly the kind of task that compounds — four hundred near-identical replies a month — and exactly the kind a language model drafts well, because the variation is small and the judgment is low.

The documentation. The thing one person carries in their head: how an order moves from inquiry to invoice, which clients need a narudžbenica, what the supplier needs to hear. AI is unreasonably good at turning a messy ten-minute voice note into a clean written procedure. Not because it knows your business — it doesn’t — but because it’s patient with structure when you’re not.

The lead handling. An inquiry arrives at 21:40. It sits until morning. By morning the person has emailed two competitors. AI can draft the first reply the moment it lands — acknowledge, ask the two questions you always ask, flag it for a human. It doesn’t close the deal. It keeps the deal warm until you can.

The reporting. The Monday number you assemble by hand from three tools and a feeling — last week’s revenue, which invoices are overdue, which inquiries went cold. AI can read those exports and write you the paragraph. You still decide what to do about it.

Notice what these share. All bounded, all repetitive, all low-judgment. They sit on top of work that already happens. That’s not a coincidence — that’s the filter.

What it can’t

It can’t decide whether to fire a supplier. It can’t tell an unhappy client they were wrong. It can’t price a job that’s never been priced before. Anything that turns on relationship, context, or consequence stays human — and should.

And it can’t fix a process that’s already broken. If five people handle inquiries five different ways, AI doesn’t standardize that. It just produces five kinds of mess, faster. The thing people call an “AI problem” is almost always a clarity problem wearing a robot costume.

The reframe

Here’s the part the deck got wrong, the part worth slowing down for. AI isn’t a chatbot you bolt onto the side of your business and hope it sticks.

It’s a system you install around how the business already works.

The difference is everything. A chatbot is a thing you buy. A system is a sequence you fit to the path an order already travels, the questions a client already asks, the report you already need every Monday. The technology is the last thing you choose, not the first. You start from the two hours your office manager loses — and you work backwards.

This is why the €40,000 deck was the wrong shape, not just the wrong price. It started from the robot. It should have started from her.

What an owner can actually do this quarter

You don’t need a tech team. You need three small steps, in order.

Write down one task. Pick the single most repetitive thing in your week — the one you’d describe as “we always have to…” Write down exactly how it’s done now. Every step, every exception. If you can’t, that’s your real first project, and it isn’t AI.

Test it on that one task, by hand. Take a normal AI assistant. Feed it ten real examples — ten inquiries, ten replies — and see if the draft is good enough that a human edit takes thirty seconds instead of ten minutes. Don’t integrate anything. Don’t pay for anything beyond a basic subscription. Just see if the work gets lighter. Often it does. Sometimes it doesn’t — and finding that out costs you an afternoon, not €40,000.

Only then, make it a system. When one task clearly works, you wire it into where the work already lives — your inbox, your spreadsheet, your Monday. That’s the step where a system gets installed around the business. It’s also the step where most owners should get a second pair of hands, because the wiring is where the quiet failures hide. If you’d rather not guess at the wiring, that’s exactly what a paid Diagnostic is for — a structured read of where to start and what to install first.

The order matters. Task, then test, then system. Never the reverse. Buy the platform first and you’ve bought the robot on the cover — €40,000 of capability aimed at a problem nobody wrote down.

The owner in Split never bought that license. We started with his office manager’s two hours. We wrote down what she did, tested it on a week of real emails, and wired up the half that was genuinely the same reply four hundred times over. She got her mornings back. The robot never showed up. It turned out he didn’t have an AI problem at all — he had two hours, and a question nobody in the deck had thought to ask.