What shadow AI is, and how big it really is
Shadow AI is any AI tool your team uses for work that the company either does not know about or does not pay for.
The best measurement on this is a university-led survey covering 47 countries with 48,340 respondents (roughly a thousand people per country; the data was collected between November 2024 and January 2025). The key question: how do you get access to the AI tools you use for work?
| Answer | Share |
|---|---|
| I use a free, publicly available tool | 70% |
| I use a tool provided by my employer | 42% |
| I use a public tool that I pay for myself | 18% |
(More than one answer was allowed, which is why the total exceeds a hundred.)
More than twice as many people use a public tool as a company-funded one. That is shadow AI in a single line.
What the policies actually look like — and this is where the argument is settled
The same survey asked whether the company has a policy on generative AI use:
| Answer | Share |
|---|---|
| there is a policy that guides use | 34% |
| there is a policy that bans it | 6% |
| there is no policy | 41% |
| I do not know whether there is one | 19% |
Two things can be read out of this, and both contradict the usual story.
The first: bans are rare. Six per cent work under a banning policy. The "everyone bans it and staff route around it" narrative describes a small minority.
The second: almost one person in five does not know whether a policy exists — more than the number working under a ban. Shadow AI is not rebellion. It is a missing and an unknown rule.
The most telling figure supports this: 38% of respondents used AI in their work without knowing whether they were allowed to. The rest of the shares come from the same survey:
| What they did (sometimes or often) | Share |
|---|---|
| put company information into a public AI tool | 34% |
| uploaded copyrighted material | 34% |
| used AI in a way that breached the policy | 31% |
| hid the fact that they used AI for their work | 42% |
| relied on the output without checking it | 42% |
| made a mistake in their work because of AI | 31% |
Do bans work? What we know, and what we do not
This has to be said honestly, because the market is very confident here and the evidence is not.
What argues against banning:
- A 1,250-person study across four countries (April 2026, non-IT office staff at companies with revenue above 500 million dollars) found that 86% of respondents say their company has an AI policy — and 66% of them still used a tool they believed was banned. The mere existence of a policy holds nobody back.
- The same study also measures what restriction costs: 77% of respondents say it holds back their professional development, and 75% would move to a company with better AI training. Banning is not a neutral move.
- A 2026 security survey of 602 breached organisations found that 68% of companies have no tooling whatsoever for detecting shadow AI — and that share got worse, up from 63% the year before.
And here is what has to be said from the other side: I found no measurement that compared banning and non-banning companies on the same sample. Which means the sentence "bans do not work" has no control-group evidence behind it. What does exist: plenty of people break the rule, and most companies cannot see what is going on.
One thing, however, provably works, and it is not banning use: the same security survey found that 92% of the organisations that suffered an AI-related incident lacked proper access controls — role-based access, multi-factor authentication. Restricting access has evidence behind it. Banning use is the part with none.
What it costs
The 2026 security survey found that the share of security incidents involving shadow AI rose from 20% to 43% in a single year across the organisations studied, and that the average breach cost in those cases rose from 4.63 to 5.39 million dollars.
That figure cannot be scaled down to a small company — it describes 602 large organisations, in cases involving between 2,590 and 115,380 compromised records, and the study was funded by a company that also sells security services. What can be taken over is not the amount but the structure: 63% of the cost comes from detection, escalation and lost business. At a twenty-person company too, it is time and customer trust that disappear, not fine money — although there was a fine in roughly one case in five.
The national numbers: why your question is a different one
Two things tend to get called by the same name, and they are worth separating.
The EU statistics office measures AI use at company level, country by country. The international surveys above measure individual use, typically at large employers. The two are not the same thing — and that gap is precisely what shadow AI is.
The useful part is that the same office also measures individual use, so the two numbers can be put side by side. Open the one that applies to you:
Hungary — 10.4% of companies, 14.9% of people
10.4% of Hungarian firms with more than ten employees use AI technology at company level. Meanwhile 14.9% of Hungarians aged 16–74 used generative AI for work in the previous three months, rising to 15.9% among those aged 25–64.
In other words, more people use it for work than there are companies admitting they use it. That is not a contradiction, it is the cleanest official measurement of shadow AI there is — and both numbers come from the same statistical office, for the same year.
One more figure confirms it: 50.3% of Hungarians who use AI at all also use it for work. That is among the highest shares in the Union — so in Hungary AI is not primarily entertainment but a work tool, just not one the company hands out.
Spain — twice as many companies, and a bigger gap even so
20.27% of Spanish firms with more than ten employees use AI technology: nearly double the Hungarian rate, and just above the EU average. At first glance, then, sanctioned use is in better shape there.
The individual side, however, is further ahead still: 37.9% of Spaniards aged 16–74 used generative AI in the previous three months (EU average 32.7%), and 17.9% of them for work — 20.5% among those aged 25–64. Which means about as many people use it for work as there are companies admitting to any use at all.
Seen from the company side the gap is wider still: a summer 2026 survey that included 501 Spanish office workers found 82% of respondents using AI for daily tasks, and 82% of those had used a tool the company did not provide, without telling anyone. Company AI training had reached 44% of them.
What to take from it: higher corporate adoption does not remove shadow AI. Twice as many Spanish companies use AI as Hungarian ones — and undeclared tool use is still there, measurably so.
Translated to a twenty-person service company [My own estimate, using the shares above]:
| Step | Number |
|---|---|
| headcount | 20 people |
| of those, use AI for work | ~10 people |
| of those, also use a free public tool | ~7 people |
| of those, have already put company information in | 2–3 people |
Two or three people. Shadow AI is not an "everybody does it" problem, it is a "two or three people you do not know about" problem. And that is exactly why it stays invisible.
A US telemetry study (222 customer companies, via a browser extension, for the 2025 calendar year) adds one more important detail: with the less well-known tools more than half the usage runs on personal accounts, while for the best-known chatbot it is around a third. Shadow AI is not biggest where you would look for it — it is biggest where nobody at the company ever thought of buying a subscription.
How to map it in a day
I found no official guidance for this, so the sequence below is my own — but every step works from systems you already run, and one person at a 10–50 person company can walk through all of it.
| # | Source | What it shows | What it misses | Time |
|---|---|---|---|---|
| 1 | Bank and card statements for 12 months — search for the tool names | individual subscriptions put through the books | free use, payment on a private card | 1 hour |
| 2 | The admin console of your office suite: third-party apps and access grants | every tool that signed in with a company account — and what it was granted access to | sign-ups with an email and a password | 1–2 hours |
| 3 | Searching the company mailboxes for sign-up confirmations | wherever someone registered with a work address | sign-ups from private accounts | 1–2 hours |
| 4 | Firewall or DNS logs, if you run a managed network | traffic leaving the company network | home and mobile use — today that is the majority | half a day |
| 5 | An anonymous five-question survey | what the other four never see: use on personal accounts | whatever people will not admit | 1 hour |
The questions for step five:
- Which AI tools have you used for work in the last 30 days?
- Which of them with a company account, and which with a private one?
- Do you pay for any of them out of your own pocket? How much?
- Have you ever pasted in a customer name, contract text, a price, source code or personal data? (yes/no)
- What is the one thing the company tool cannot do, which is why you use something else?
On the fourth question, anonymity is methodology, not politeness. If an answer can be traced back to a person, the answers are worthless — the international survey found that 42% habitually hide their AI use.
The fifth question is the most important one, because it tells you what you have to allow in order for shadow AI to stop existing. One of the most telling figures in the large-company study is that 72% of respondents believe they understand AI better than their own IT team. You can argue with that — but its meaning is unambiguous: employees are not blind to the risk, they are short of tools.
What comes after the survey
The order matters: the policy has to be cut to fit the survey result, not the other way round. Write it first and it will describe what you imagine — not what people actually use.
Four steps, in this order:
- Allow a single tool with a company account for whatever most people use AI for.
- Write down what must never go into it — on one page, item by item. There is a ready-made skeleton for that: the one-page AI policy lays out all twelve lines.
- Set up the access controls. This is the only control with measured evidence behind it.
- Train people. On top of everything else this has been a supervised EU obligation since August 2026 — the details are in the team's AI literacy plan, session by session.
Two misunderstandings
"We have no shadow AI, because we do not allow it." By the data this is the worst possible state: no permitted route, and no detection either. The two or three people are still there regardless.
"The firewall will catch it." In a hybrid team most of the usage never touches the company network. Step four is useful, but it is not enough.
What to take away
Shadow AI is not a mass phenomenon but a handful of people you do not know about — and its most common cause is not rule-breaking but the fact that there is no rule, or nobody knows it. The survey takes a day, and it does not require buying any new software.
There is no evidence that banning holds usage back, and there is evidence that it carries a retention risk. What there is data on: regulating access works, and two thirds of companies have no detection tooling at all.
And the most important part: the fifth question of the survey is not about risk, it is about what to give your team. What they are allowed to upload in the first place is broken down item by item in GDPR and AI, with seven decision questions.