Denis Romanovskiy has a number he is not satisfied with. Artificial intelligence has made individual employees at SOFTSWISS between 20 and 50% more productive, the chief AI officer of the online gambling technology supplier says, and everyone in the company can see it.

"We are now in the stage of individual productivity," he says. "This improvement is transparent, everybody sees it. But we need to move further, because what we see is just 20 to 50% of improvement. To really have an advantage, like 200%, we need to fully give the work to it."

Giving the work to the model is not about eliminating a person. "It means that we are not replacing an individual, but we are changing and broadening what individuals need to do, inside of the company or outside of the company," he says. "And this is more complicated."

The company published its fifth annual iGaming Trends Report this week, arguing that online gambling is an early test bed for what AI does to workforces, security and regulation in every regulated digital business. Romanovskiy is the person inside SOFTSWISS who has to make that thesis true, and he is noticeably less enthusiastic than the accompanying press release.

"We are still in early days of AI," he says. "At this moment, we're not there yet."

Code and prose

Romanovskiy describes four steps to get where the industry now is, and he is careful to say each one had to be earned.

First, AI could research: put in a question, get an answer. Then it was connected to tools, to search and to company data, which made it useful because a user could point it at their own context and get a relevant answer rather than a generic one. Then, he says, "we understood that we can trust it a little bit, so the trust grew". Fourth, the cost of the models came down far enough to run things at scale.

What is missing is not intelligence but architecture. "You need not only the models, you need the tools that know how to behave in particular situations, and how to organise the proper work of what we call code and prose together," he says.

Prose is the instruction, the natural language a model reasons in and the part that can wander off the range. Code is the part that cannot. "The prose manages the code, and then the code manages the prose back, and doesn't let the prose go in wrong directions," he says. "So, you have some deterministic approach and a non-deterministic approach, combined together. This is where AI brings really something new. The new field where you can go into really complex scenarios of automation."

It is a description of engineering discipline rather than of a breakthrough, and it is the opposite of how AI capability is usually sold. The model is the cheap part. The scaffolding that stops it improvising is the product.

Why everyone suddenly wants hands

AI is rapidly transforming from a dialogue box that gives you recipes recommendations and train tables to an agent that can do some work. Asked what an agent actually is, Romanovskiy gives the plainest definition available.

"An agent is the AI that has hands, that can touch something, do something, that can do the real work," he says. "Without hands, it just can give you the text. But when AI takes the command and produces some text that then touches some switches, this is an agent, because it can do the change to the outer world." Without them, he adds, "it's just a brain without hands".

That is why the word is everywhere. "Everybody talks about agents because to do change in the jobs, and really to optimise, you need these hands."

The difficulty is not building the hands. It is what has to be poured into them. "You need to properly train these agents and control them," he says. "And you also need knowledge of different people put into the same agent. This is the hardest thing to do, because in enterprises knowledge is in the heads of many different people, and how can we put it together?"

Which is why the people who have to change are not the engineers. Handing over a workflow means the domain experts have to learn to specify it. "You need to train your people, and not only technical people," he says. "The people who understand the domain, like finance, like legal, like compliance. They need to understand how to write these instructions, these guardrails, and how to trust the information that comes in and goes out."

Trust is a control system

SOFTSWISS's software sits underneath operators handling real money in real time, which makes the question of autonomy less abstract than in most industries. Romanovskiy does not pretend the models are fully trustworthy. He describes buying trust with limits instead.

An agent may be allowed to run financial transactions, he says, but inside a cap on how many it can execute in an hour and on the total value it can move. Breach either and it has to call a human.

"We don't trust them because we trust them," he says. "We trust them because we build guardrails around it. We tested it, we saw that other companies in other industries are doing the same, and the overall opinion is that yes, they are more capable."

The corollary is that managers cannot delegate the judgement. Users "need to learn the tools, try themselves, feel when the model goes wrong", he says. "You trust the tools, you have your people who know how to do it, and you build the system from it, with proper people and systems together."

Without that, he argues, the capability is wasted. Companies have tools "that probably are really capable", but "we need to do the organisation around it that can really manage it". The bottleneck is institutional, not technical, which is the same conclusion the report reaches about the industry as a whole.

Thirty teams, three months, three questions

Put to him that investment banks are pouring money into AI to build an edge, and asked whether gambling technology is in the vanguard of the same race, Romanovskiy declines the frame entirely.

"Our approach is not like we try to be a pioneer. We try to be efficient," he says. "My goal from my chief executive is to make an economic effect from implementing AI in the company. So, I'm not looking into very big projects that will change everything. I look at the very small initiatives of smaller teams and individuals, how they can use AI to really change their work."

The method is deliberately unglamorous. The company ran a hackathon with about 30 small teams which produced some 30 initiatives, which Romanovskiy is now working through department by department, going to managers to find candidates small enough to finish. He caps each one at roughly three months.

Every proposal has to answer the same three questions before it starts. Will it save time? Will it improve the quality of the output? Will it stop you hiring? "And from these questions we go into: should we run it, or should we not?"

His own job is being reshaped by the same logic. He is building an internal route for anyone in the company to take an idea for a new product in any market, use AI to draft the business plan, build a prototype, and then ask the business to back a trial.

He is blunt about what that does to roles. "In the past you could do just one job. Now with AI you can do more jobs that were previously not possible for you to do, because AI can replace some people," he says. When he builds an application himself, "it can replace the product owner, the tester, and maybe even the designer, and I can create the full tool".

The limit arrives at scale. Getting from a working prototype to a feature serving hundreds of thousands of players still requires engineers thinking about architecture and performance. "That's a big story," he says. "It's slow."

The Brazil problem behind the technology question

The technical conversation is happening against a commercial shock. Brazil dominated discussion at the report's launch, and not in the way the industry expected a year ago.

President Luiz Inacio Lula da Silva issued a provisional betting ban on September 25, which Congress has 120 days to approve or let lapse, in what had become one of the world's largest online betting markets. The report records 415 new illegal betting sites created after the ban, roughly 300 of them on a single Saturday.

Romanovskiy says the commercial question is not his. What he will say is that the market changed the company. SOFTSWISS was focused on regulatory compliance and localisation for Brazil, he says, describing heavy investment in tailoring products for the market, in marketing, and in assembling the data on how to operate there properly. SOFTSWISS does not run gambling operations or make games; it supplies the platform the operators run on.

He is frank about the disappointment. The industry "had great hopes because of Brazil", he says, on the reasoning that if Brazil worked other markets would follow, and operators treated it as a lever for the whole sector. "Now we need to step back a little bit and think about it. Maybe it will get resolved. I don't know, because I'm a technical guy."

Asked whether a political shock of that size is an argument for diversifying out of gambling, into the fintech work SOFTSWISS has already begun, he does not answer as a strategist. He answers as the person who would have to build whatever came next.

"My goal is to let the company build whatever products very quickly," he says. That is what the internal prototyping route is for, and it is a more revealing answer than a yes. The hedge against political risk in one market is not a plan for another market. It is the capacity to stand up a product in a new one before the window closes.

Not there yet

The report gives the industry an AI maturity score of 45 out of 100, even though 85% of betting and gaming companies globally have adopted the technology in some form and 52% of the professionals surveyed named AI the leading macro trend for 2027. Romanovskiy's account explains the gap between the adoption rate and the maturity score better than the score does.

The models are capable and cheap enough. The guardrails exist. What is missing is the organisational layer: domain experts who can write instructions precisely enough to be trusted, managers who have used the tools closely enough to know when one is failing, and processes designed around handing over a workflow rather than assisting a worker.

Until that is built, the gains stay stuck at the level of the individual, which is where SOFTSWISS says they are now. The company's 2027 iGaming Trends Report is available in full online.