ATAILA Newsroom · Budapest · 2026-08-10

If 74% of knowledge workers already use AI, what is still missing inside companies?

hvg360 — the premium tier of the Hungarian news site hvg.hu — used its business-intelligence newsletter to summarise the Boston Consulting Group's fourth global AI at Work study, and it states something a major consultancy rarely says this plainly: the bottleneck in AI adoption is no longer the technology. For us that is not news but confirmation — this is exactly the gap the ATAILA platform was built for.

What hvg360 reports

First the numbers, as the newsletter presents them.

One footnote on the second number: BCG's own release words the same finding more cautiously — more than 20 percentage points over two years.

The point of the article, though, is not the adoption number but what is missing behind it. Organisational adaptation has not kept pace with the speed of use: at many companies nobody has clarified what new skills are expected, who answers for an algorithm's mistake, how to lead people and AI agents in one team, or what should happen to the time freed up. The line separating executives' AI use from that of everyday employees is what BCG once called the “silicon ceiling” — and according to the article, that ceiling has now visibly cracked.

Which is where the joy paradox comes in: work gets better and harder at the same time. The study's public summary puts numbers on it — two-thirds of regular users are more satisfied with their work, while roughly two-fifths report higher cognitive load, and two-thirds get little or no guidance on what to do with the time they save.

And one number worth reading twice: according to BCG's summary, a clear strategy raises the odds of measurable business impact by 25 percentage points — a better tool on its own by roughly 5.

It is not the models that fell behind

The diagnosis is accurate, and uncomfortable for most of the sector. It is not the models that fell behind. Not the licences either. It is the operations that fell behind.

We would sharpen the wording on one point. If the conclusion is that AI adoption is a leadership task, it is a short step to believing the fix is a strategy workshop and a policy document. Our experience is that every one of the four open questions lands, in practice, as an operational question:

• Who answers for the algorithm's mistake? → You can assign accountability where there is authentication, authorisation and an audit log. That is a platform capability, not a paragraph in a policy.
• What new skills do we expect? → The skill that builds is the one the team practises daily, on a safe tool. Without one, the practice happens in private accounts.
• What happens to the freed-up time? → It can only be reinvested if there is somewhere for it to go: the next workflow that actually reaches production.
• How do we lead a mixed team? → You have to be able to see what the agent did. That is logging and environment management.

The gap is visible from the other side too. In an earlier response, on the back of a KPMG study, we wrote that most enterprise AI projects fail not at development but in the chasm between prototype and production. BCG is now measuring the other bank of the same chasm: the people have crossed; the organisation has not.

What ATAILA does with this gap

1. Access is not handed out as an experiment

The 74% is not a forecast; it is the present tense. At most companies the question is no longer whether people use AI, but where — and typically the answer is public accounts, with no oversight. Our answer is the Private ChatGPT: the same user experience, but on EU GPUs, operated by ATAILA, with the data never leaving the platform. A prompt is data processing in its own right — it deserves to be treated as such.

Your team already uses AI. Give them a private one.

2. The freed-up time needs an outlet

According to BCG's summary, two-thirds of regular users get little or no guidance on what to do with the time they save. That is not a motivation problem. A team gaining a day a week does not make the path from idea to production any faster for a new application — that path still burns weeks or months on infrastructure, environments and release management.

That is the part we automated. The Factory provisioning engine builds, deploys and audits the application from a manifest written once, and the Release Manager walks it through the environments to production. The description is written once; the environments and the release derive from it. The time you win does not evaporate — it goes into the next workflow.

3. One layer does not make an operation

Most of the market sells a single layer: some sell GPUs, some models, some development, some hosting. The company is left holding the integration and the operations — which is precisely what BCG says fell behind.

We ship the whole row: a sovereign EU cloud (Cloud), a provisioning engine with release management (Factory), and a private AI development environment (Studio) — as one system, operated. Not a trial environment: it runs on the same foundation our own products run on in daily production. The value is not in the individual layers but in the connections between them.

4. Predictable cost, predictable exit

Two practical things that make an executive decision easier. One: a fixed monthly fee that includes the private AI, the hosting and the people who operate it — we do not bill per query. Two: the code and the data remain yours, and exporting them does not require a separate negotiation.

This is not theory: our open*.hu reference products run on this same platform, in production — the newest of them went live in hours by reusing finished modules.

What we commit to

Where the boundary runs

We do not claim the platform solves the joy paradox. Cognitive load, the “what happens to my role” question, designing the accountability structure and leading mixed teams are human and organisational work. That remains the job of the company's leadership, not its vendor. We are not change-management consultants, and we will not write your AI strategy for you.

Nor do we promise that a bad workflow becomes a good one because AI runs on it. A bad process, automated, is just bad faster. That is why the first conversation includes the question of whether there is a real problem and real data behind it — and if there is not, we say so.

What we do commit to is narrower, but verifiable: we supply the layer BCG’s research says is missing — the operations. The tool where the team can practise safely, the environment where the freed-up time turns into a new workflow, and the log that shows who and what did what.

This offer is not for everyone. It typically works at regulated or data-sensitive companies of 20–200 people, where there is an expensive, repetitive workflow — and where the data cannot be copied into a public AI.

Where it runs, and who operates it

If your team is already part of that 74% — and it very likely is — then the decision is not whether to have AI, but where it runs, who operates it, and what makes it into production. We are happy to talk about that confidentially and without obligation: tell us the one workflow whose data you will not let out of the company. We will tell you honestly whether it can reach production — and if so, roughly how long it takes.

Get started →