AI adoption in Romania: what a year of building production systems means for your business

After more than a decade of building custom software, we launched Romania's first applied AI studio in 2025. A year later, we have live systems running in construction, public procurement, healthcare, logistics, distribution, and eCommerce, and roughly as many projects scoped, costed, and waiting for the next round of government digitization grants to open before the client gives the green light.

That second group tells you more than the first, because what decides whether an AI project works in a Romanian company has surprisingly little to do with AI. Most of it never comes up in a sales conversation, because you only learn these things by keeping systems running inside companies, not by building demos.

We have, though, but instead of walking you through our year, we'll walk you through what that year has taught us and what it’s worth to you.

The companies that come to us are rarely chasing efficiency. Most often than not, they're just growing and can't sustain that growth.

One of our clients, a large Romanian retail and distribution group, is working toward doubling revenue this year with a back office of about seven people. The founder has said publicly that for every open position, they interview two or three candidates and one stays. If you look at it, it’s not a productivity problem. They can't just hire their way through the growth, because the people they need aren't on the market.

We'll be honest with you in both directions, because we're honest about it with every company that calls us. If you can hire the people you need at a price you can pay, hire them. A person handles exceptions, learns your business, and takes responsibility. AI does none of those things, and a system costs more to build than a salary costs to try.

However, the calculation changes when the labour market has closed that door, and in Romanian mid-market operations right now it's closed more often than not. When there's nobody to hire, the question isn't whether AI is better than a person. It's whether it's better than leaving the work undone, because that's the real trigger for AI in this market.

Every AI readiness exercise starts by hunting for repetitive work. Unfortunately, that filter is so wide it's useless, which is why those exercises produce a list of 40 opportunities and 0 projects.

The thing is, repetitive work that can be delegated has already been delegated. It probably went to a junior, an assistant, or a shared service centre years ago. Any Romanian company above 50 people has had that conversation with itself.

What's left is repetitive work that's locked to a senior (and usually expensive) specialist by something other than cost.

In a year of production systems, we've found three of these instances:

Legal.

In public procurement, the appointed technical evaluator has to read every requirement line personally, because the law says the evaluator evaluates. He can't hand 600 requirements to an intern no matter how much he'd like to. In healthcare, a diagnosis carries a signature and the signature carries a name.

Knowledge.

At an electrical contractor we work with, the senior engineer was the only person who could tell whether a supplier's cable line and a client's cable line described the same product (voltage class, cross-section, type). So he typed the prices himself. Four thousand copy-paste operations per bid, because nobody else could spot a wrong match.

Liability.

A wrong number inside a commercial proposal is a signed obligation to deliver at that number. Companies keep senior people on work that bores them because the cost of one error lands on the balance sheet.

You can run this diagnosis yourself this week, for free. Take your three most expensive non-executive roles and ask each person to walk you through their last full working day in fifteen-minute blocks. Mark every block that needed none of the expertise you hired them for. Then ask, for each marked block, why it can't go to someone cheaper. If the answer is that you tried to hire for it and couldn't, you've found your project.

And one more free exercise while we’re at it: find out who in your company is already pasting things into ChatGPT without being asked. Whoever it is, they probably picked those tasks themselves, which means the tasks are repetitive enough to be worth the effort and not sensitive enough to make them hesitate.

We hear the same sentence in most first meetings: we need to sort our data out before we can do anything with AI. That sentence buys a year of delay, and the delay is usually wasted.

We've built systems on top of supplier price lists in Excel, Word and scanned PDFs, on manufacturer data sheets that all use different formats, and on thousands of recorded phone calls. Dealing with messy data is part of the job you're hiring us for.

You're also more ready than you feel, and this is something outsiders consistently get wrong about Romania. ANAF spent years forcing businesses here into structured electronic reporting. e-Factura, SAF-T, e-Transport. It was expensive, it was resented, and it felt like compliance work with no upside.

The side effect is that a Romanian mid-market company today holds invoice-level, transport-level, and ledger-level data in machine-readable form, with enforced schemas and enforced deadlines, inside an organisation that will describe itself in the first meeting as "not really digitalised". Companies of comparable size in several Western European markets never went through that. Their operational data sits in places where nothing can be read.

What decides whether the project succeeds

So messy data won't stop a project from starting, but we've also watched projects with perfectly good data fail, which raises a more useful question: what actually separates the AI projects that work from the ones that don’t?

To answer it, you need to know one thing about how these systems work. They don't replace your people and run unsupervised. The AI does the heavy, repetitive part, and one of your employees reviews what it produced, confirming or correcting it. That's where projects are decided: when the AI proposes a result, can your employee look at it and immediately tell whether it's correct?

At DentOP, a dental group we built three connected systems for, a doctor looks at a flagged region on a panoramic X-ray and confirms or corrects it on the spot. Both systems improve every week, because every mistake gets noticed and corrected immediately.

Compare that with a request we get all the time: score our leads by how likely they are to buy. Nobody can look at a number like "82% likely" and know if it's right. You find out in four to six months, maybe, and by then the result has been shaped by what the salesperson did, what the market did and what your competitors did. A system like that never improves, and after a year everyone stops looking at the score.

The data quality is similar in both cases, but the outcomes are opposite. Part of our job, before you spend anything, is sorting your project ideas by that test instead of by how exciting they sound.

At some point in an AI project, you get handed a list of features with a price next to each one, and you have to decide what you're paying for. In our experience, the features that earned their cost were almost always the dullest ones on the page, and the ones clients were most relieved to have skipped were the ones that look best in a demo.

Start with the features someone can verify fast

Coni, a construction company we work with, is a good example of where that leads. Site paperwork reaches them in every format imaginable, most of it photographed or scanned, and what we built for them does one narrow thing: it reads each document and fills in twelve specific pieces of information, keeping Romanian diacritics intact, which most off-the-shelf tools destroy. Twelve fields, each either right or wrong, so the person who filed the document knows within a glance whether the system did its job.

The same contract listed more sophisticated features at a price, including a search that understands what you mean rather than matching your exact words, and the client declined all of them in writing. That was the right call. Search of that kind gives you results, but nobody in your office can tell you whether a better document existed and was skipped, so nothing it produces can ever be confirmed or corrected. A feature nobody can check will not earn anyone's trust, however well it presents.

The impressive feature rarely works without a dull one underneath it

The visible feature rarely works on its own. Buy it without the unglamorous machinery beneath it and you have bought a faster route into the same disorder.

Oltwam distributes LPG, and their orders used to arrive in WhatsApp groups. We put the AI where the orders already were, so a customer writes in ordinary language and gets a price, within limits the company set in advance, and a human operator receives something clean. That is the part worth showing anyone. But in the same project they also bought the dispatch system underneath, which is where those orders land so the office can see what is happening and act on it. Without that, faster orders would simply have piled up in the same place the old ones did.

So when a proposal has one exciting line and several boring ones, the boring ones are usually not padding.

Two questions to ask any vendor, including us

First, what is this system not allowed to do? A vendor who cannot answer has not set limits on what the AI can produce, and output without limits cannot be checked by anyone.

Second, which items could I remove and still get what I came for? A vendor who defends every line equally has not worked out which one carries the value. A vendor who tells you to cut three of them has.

An AI system doesn't answer every question with the same certainty. On some documents it's on firm ground, on others it's guessing, and any system built properly knows the difference and can be told what to do about it. In one of our systems, more than 80% of the work goes through with no human involved at all, and fewer than 2% of those come out wrong.

Where you draw those lines looks like a technical setting, but it isn't. It's a price, and you're the one who should be setting it. Push the line so the system handles more on its own, and you save staff hours and accept more mistakes. Pull it back, and you catch more mistakes and pay for the hours. The same choice shows up everywhere in a system.

With Oltwam, the LPG distributor whose orders arrive by WhatsApp, the equivalent decision is how far the AI can move on price before a human has to approve it, where a wider band closes more orders at a thinner margin. It depends on what one wrong answer costs you, and that cost is not remotely the same for a €200,000 procurement bid as it is for an X-ray a doctor is going to look at anyway.

Most vendors set these thresholds by feel, or wherever the demo looked best. Ask any vendor directly, including us: what does one wrong answer cost me in this process, and what limit did you set based on that number? Anyone who has run systems in production has done the math.

Here's something we learned that no Western playbook mentions: in this market, projects don't die (contrary to popular belief), they just queue.

A Romanian company that has decided AI is worth doing frequently doesn't proceed to buy it. The project gets scoped, costed, and parked until a European or national funding call opens. Digitalisation grants for SMEs run into the tens and low hundreds of thousands of euros, with co-financing between 10% and 50%. Half the projects we've scoped this year are sitting in exactly that queue.

A year of watching companies win and lose these races taught us three things, and getting them wrong wastes a year.

The grant won't pay for your idea

Several current calls will not consider a project that is still on paper, and require you to show the technology working at the moment you apply. So the order most companies assume (secure the money and then build) is backwards for those calls. A small prototype paid for out of your own pocket is often the thing that makes the application eligible at all, and it happens to be the cheapest way to find out whether the idea works before anyone commits money to it.

Eligibility criteria will try to redesign your project

Grant guides come with a list of reimbursable costs, and it might be easy to start building what’s on that list rather than what solves the problem you identified initially. That produces systems that get funded, get installed and never get used. The advice we give to clients is to let funding decide the timing and what gets built in which order, but never let it shape the scope.

The window is weeks, not months

Funding for digitalisation has decentralised into eight regional programmes with separate budgets, calendars and scoring grids, so you're competing against firms in your own region rather than the whole country. A guide gets published and applications close within weeks. A company that starts scoping when the call opens doesn't make it.

This means the work has to be done before the money exists. The process diagnosis, the technical specification, the cost breakdown, and the supplier quotes need to be finished, ready to attach to an application the week a guide appears.

That's a service we now build into how we work, because we watched unprepared companies miss the window. One caveat: calendars and eligibility rules change by ministerial order, so verify against mfe.gov.ro and your regional programme rather than any summary, including this one.

A year of production systems means the expensive lessons have already been paid for, on projects that weren't yours. You don't need to discover that lead scoring never improves, or that the flashy features are the wrong ones to buy first, or that the funding window closes faster than a scoping process runs. We already know, and that knowledge is most of what you're buying when you work with a team that has shipped rather than piloted.

If any of this sounded like your company, the diagnosis costs nothing. Run the fifteen-minute-block exercise on your three most expensive people, or book a discovery call and we'll run it with you. Thirty minutes, no obligation, and if the honest answer is "hire someone instead", that's the answer you'll get.