Shadow AI: Your Employees Already Use ChatGPT. What Is the Organisation Doing?
The first AI usage inventory we run in an organisation always ends with the same finding: management believes AI "has not arrived yet". Employees have been using it for two years, from their phones, with personal accounts.
What shadow AI is
Shadow AI is the use of artificial intelligence tools by employees without the organisation’s approval, knowledge or control. The term comes from shadow IT, but the scale is different. Setting up an unapproved server took skills and time. Pasting a contract into ChatGPT and asking for a summary takes ten seconds and a free account.
The PagerDuty survey published in June 2026, covering 1,250 office professionals in non-technical roles at large enterprises, gives the picture in numbers:
- 66% used AI tools at work while believing company policy did not allow it.
- 88% have shared work information with public AI tools. 43% emails and correspondence, 40% meeting notes, 34% customer data, 31% financial information or confidential documents and strategies.
- 86% say their company has an AI policy. 81% believe leadership plays by different rules.
- 89% first met the tools in their personal lives and brought them to work.
The survey covers the US, UK, Australia and Japan. From what we see in Greek organisations, there is no reason to believe the percentages are lower. If anything differs, it is that the AI policy often does not exist at all.
The paradox
Shadow AI is not a sign of undisciplined staff. It is a sign of people who want to do their job better and faster and found no approved way. 77% in the same survey believe restrictions on AI use limit their professional growth. A ban does not stop the use. It hides it.
What the real risks are
The risk of shadow AI is not that “the AI will make a mistake”. It is that the organisation takes on obligations without knowing it. Four of them have a specific legal or contractual footprint.
GDPR. When an employee pastes a customer email, a list of names or meeting notes containing personal data into a public tool, the organisation has transferred data to a third-party processor without an Article 28 contract, most likely outside the EU, without a legal basis and without the DPO knowing. The free tiers of most tools state explicitly that they may use the data to train models. In a breach or a complaint, “we did not know” is not a defence; it is an admission that there were no measures.
Confidentiality and trade secrets. Financial figures, code, contract terms, strategic plans. 31% of respondents have shared such material. Beyond the competitive risk, many contracts with clients and partners contain confidentiality clauses that this act breaches, whether or not the data ever leaked.
AI Act, Article 4. Since 2 February 2025 the organisation, as a deployer, must take measures for the AI literacy of the people using AI systems on its behalf. The employee using ChatGPT to answer customers is using an AI system on your behalf, whether you approved it or not. If you do not know they are doing it, you have not taken measures. We analyse the obligation in a separate article.
NIS2 and supply chain security. For entities within the scope of Law 5160/2024, every SaaS tool processing the organisation’s data is a supplier. A tool nobody has assessed, with an account nobody manages, is exactly the gap the supply chain security measure exists to close.
There is a fifth risk, less legal but more frequent: quality. Outputs produced by a tool without review, by a person not trained in its limits, end up in reports, offers and answers to customers. When something goes wrong, nobody knows AI was involved.
Why bans fail
The first reaction of many boards is to block the domains. We have seen this approach often enough to know the result: use moves to the phone and the personal laptop, that is, outside any control and any logging. The organisation loses the visibility it had, gains a false sense of security, and tells its people that management does not understand how they work.
The second reaction is a “zero use” policy in a PDF nobody reads. 86% of respondents say a policy exists. 66% break it. The ratio speaks for itself.
The point is that shadow AI is a symptom of demand. People need the tools. The question is not whether they will use them, but whether they will use them inside or outside the frame the organisation sets.
What works
From the AI governance engagements we have run, the approach that brings shadow AI into the light and keeps it there has four elements. The order matters.
1. An inventory without punishment. Before any policy, a short anonymous survey and a few conversations per department: which tools do you use, for what, how often. The message must be clear that the aim is to provide better tools, not to find culprits. The inventory almost always reveals uses management did not imagine and, more often than one would expect, truly useful ones worth making official.
2. Approved tools, quickly. The enterprise edition of the same tools (Microsoft 365 Copilot, ChatGPT Enterprise or Team, Gemini for Workspace, Claude for Work) solves most of the GDPR and confidentiality problems: contractual commitment that data is not used for training, EU processing where available, central account management, logging. The cost per user is small compared with the cost of one incident. If the organisation does not offer an approved alternative within a few weeks of the inventory, use goes back into the shadows.
3. A one-page policy. Not twenty pages of legal language. What is allowed, what is never allowed (customer personal data, credentials, confidential contracts, production code), which tools are approved, how to request approval for a new tool, who answers questions. And, critically, how to disclose that an output was produced with AI in anything that goes to third parties.
4. Training with a record. The policy does not work if people do not understand why it exists. A short training per department, explaining what happens to a piece of data when it enters a public tool, why the model invents sources with confidence, and how to use the approved tool properly, changes behaviour more than any ban. The record per employee is at the same time the evidence for Article 4 of the AI Act and the training measure of NIS2.
The result we see
In organisations that followed this order, AI use did not fall. It moved. From personal accounts to corporate ones, from the phone to the company laptop, from "don't tell anyone" to "look what I built". Management gained visibility, the DPO gained sleep, and employees gained better tools than the ones they had found on their own.
The leadership question
One finding of the survey deserves separate mention: 81% believe leadership plays by different rules. In quite a few organisations it is true. The CEO uses ChatGPT on the phone to prepare a board meeting while the policy bans use by staff. No AI policy survives in that environment. Management must be the first to be trained, the first to use the approved tools, and the first to say openly how it uses them. It also happens to be the first thing every supervisory authority will ask about, for the AI Act as much as for NIS2.
How we help
Northbound’s AI governance service starts exactly here: usage inventory, risk assessment per use, AI policy, selection and configuration of approved tools, and a training programme with evidence. For the baseline training of all staff we use Credenz, so that training happens per department, with assessment and a record per employee, without tying up weeks of HR time.
If you want to start on your own, the first action is free: ask your people, anonymously, which AI tools they used last week. The answer will define what needs to happen next.