Shadow AI is the use of artificial intelligence at work without the organisation’s knowledge, approval or adequate oversight. In a typical case, a customer service employee pastes a complaint thread, complete with names and order numbers, into a personal chatbot account to draft a faster reply. Nobody reviewed the tool, the organisation has no contract with its provider, and nobody knows where the data went.
Shadow AI is not an offence in itself, and there is no fixed fine for it. The legal risk lies in what the use involves, such as unprotected personal data or AI used for a high-risk purpose. Reducing that risk starts with four things: visibility into the tools in use, clear rules, approved tools that meet real needs, and employees with the AI literacy to use them well.
What is shadow AI?
Shadow AI is any AI use that the organisation has not approved or cannot oversee. It takes three forms, and only the first is what most people picture.
The first is unapproved tools and personal accounts, such as a free chatbot or transcription app set up with a private email address. The organisation has no contract with the provider and no say over settings, retention or access.
The second is AI features inside existing software. Office suites, CRM platforms and video-conferencing tools increasingly ship AI assistants, sometimes switched on by an update. The software was approved, but the new feature never went through a review.
The third is approved tools used outside their approved scope. An assistant licensed for drafting internal documents ends up screening job applicants or analysing health information. The tool is sanctioned. The use is not.
Shadow AI is a form of shadow IT, the familiar problem of unsanctioned software. The difference is that AI tools process whatever users enter and produce output that can shape decisions about people. In governed AI use, the organisation knows which tools are used, for what and with which data, and someone is accountable for each choice.
Everyday examples of shadow AI in the workplace
The scenarios below illustrate how shadow AI can show up at work. Each flags a risk to assess, not an automatic breach. Whether a rule is actually broken depends on the data, the contracts in place and how the output is used.
| Example | Potential exposure | Practical safeguard |
|---|---|---|
| Uploading customer records to a personal chatbot account | Personal data shared with a provider the organisation has no contract with; a possible data breach; no control over retention or model training | An approved enterprise tool with a data processing agreement; clear rules on what may be entered |
| Using an unreviewed AI tool to rank job applicants | A high-risk use case under the EU AI Act (rules apply from 2 December 2027); GDPR rules on automated decisions; a risk of discriminatory outcomes | Pause until reviewed; a data protection impact assessment; a person responsible for every decision |
| A meeting assistant recording a confidential HR or board discussion | Recording without proper notice; confidential data held by a third party; emotion features that may fall under an AI Act ban | Approved tools and settings; notice to participants; retention limits; no assistants in sensitive meetings |
| Connecting an AI coding assistant to internal repositories | Code, credentials or personal test data exposed to a supplier; unclear retention terms; more access than the task needs | Reviewed data-use terms; access limited to specific repositories; a secrets scan before connecting |
Can your business be fined for shadow AI?
There is no universal “shadow AI fine”, because no EU law penalises unapproved AI use as such. Regulators fine organisations for breaching specific obligations. Uncontrolled use makes those breaches more likely, since nobody has checked the data, the supplier or the purpose.
Which rules apply depends on the data, the purpose and your organisation’s role, and enforcement differs per country. Under the GDPR you are usually the controller, and under the AI Act usually the deployer. Fines are set case by case, weighing factors such as gravity, intent or negligence, and the steps taken to limit the damage.
GDPR risks when employees share personal data
The GDPR applies whenever an AI tool processes personal data, such as customer names in a prompt or voices in a recording. Sharing personal data with an AI tool is not automatically unlawful, but shadow AI skips the checks that show whether it is lawful.
Those checks start with a lawful basis, a defined purpose and data minimisation. Special category data, such as health information, may only be processed under one of the narrow exceptions in Article 9. The checks also cover a data processing agreement under Article 28, security under Article 32, retention, model training and transfers outside the EU. With a personal account, the organisation has no contract of its own that settles these points.
In 2024 the Dutch Data Protection Authority reported breach notifications caused by employees entering personal data into AI chatbots. One involved a GP practice employee who entered patients’ medical data. The authority warned that chatbot providers may gain unauthorised access to such data, and that notification is mandatory in many cases. Under Article 33, a notifiable breach must be reported to the supervisory authority without undue delay and, where feasible, within 72 hours of becoming aware of it. The EDPB guidelines on breach notification explain how to assess whether a breach must be notified.
Fine ceilings depend on the provision breached. Failings on processor contracts, security, breach notification or impact assessments can reach €10 million or 2% of worldwide annual turnover. Breaches of the core principles or the transfer rules can reach €20 million or 4%, and in both tiers the higher amount is the cap. The GDPR amendments in the Commission’s separate Digital Omnibus proposal have not been adopted, so the current text applies.
Shadow AI and the EU AI Act
The EU AI Act regulates AI by purpose and context, not by product name. The same assistant can be a low-risk drafting aid in one team and part of a high-risk use case in another. Unmanaged use means nobody has checked which is which.
Everyday drafting and summarising is not high-risk, but some purposes are. Annex III lists AI used to recruit or select people, including filtering applications and evaluating candidates. If a team repurposes a general-purpose tool for that job, Article 25 may make the organisation the provider of a high-risk system. That role carries much heavier obligations. Under the Digital Omnibus on AI, the Annex III rules apply from 2 December 2027, while the GDPR already covers the same recruitment use today.
Some practices are banned outright. Since 2 February 2025, Article 5 has prohibited AI that infers the emotions of people at work, except for medical or safety reasons. The Commission’s non-binding guidelines on prohibited practices read the ban as covering emotions inferred from biometric data, such as voice or facial expressions, and not from text alone. Check any meeting tool that claims to read emotions before it is switched on.
The highest AI Act fines, up to €35 million or 7% of worldwide annual turnover, whichever is higher, apply only to prohibited practices. Breaches of other listed obligations, such as those of deployers of high-risk systems or the transparency duties in Article 50, can reach €15 million or 3%. For SMEs and start-ups, the lower of the two amounts applies.
AI literacy obligations for organisations
Article 4 of the AI Act has applied since 2 February 2025 to all providers and deployers of AI systems. That includes most organisations whose staff use AI at work. The Digital Omnibus rewrote it with effect from 27 July 2026. Organisations must now “take measures to support the development of AI literacy” among staff and others using AI on their behalf. Those measures should reflect people’s knowledge, experience and training, the context of use and the people affected. No specific level of AI literacy has to be guaranteed for any individual.
The Commission’s AI literacy Q&A confirms that a company whose staff use ChatGPT for advertising copy or translation is in scope. Those employees should be informed about risks such as hallucination. The Commission also notes that simply relying on a tool’s instructions for use might be ineffective, and points to training and guidance tailored to each group’s knowledge and context of use.
National market surveillance authorities supervise Article 4 from August 2026. The AI Act sets no fine amount for it, so penalties depend on national law and must be proportionate. In practice, enforcement also depends on each Member State designating its authorities. The Netherlands, for example, opened a consultation on its draft implementing act in April 2026, which ran until 1 June. The draft gives the Dutch Data Protection Authority (AP) and the Dutch Authority for Digital Infrastructure (RDI) coordinating roles alongside sector regulators. Check who supervises your sector in each Member State where you operate.
You cannot tailor literacy measures to tools you do not know about, which is why literacy work starts with knowing what is in use. Our EU AI Act guide covers the wider timeline and obligations.
Seven steps to reduce shadow AI risks
These steps form a proportionate approach to shadow AI governance. Most are recommended practice rather than explicit legal requirements, and we flag where a step supports a specific obligation.
1. Identify the AI tools employees actually use
The quickest first step is to ask. A short, blame-free survey or a few team conversations will show which tools people use and why. Then check the admin consoles and release notes of the software you already pay for. Expense claims, browser extensions and app permissions in your identity platform fill the remaining gaps.
Technical discovery completes the picture, but keep it proportionate. Monitoring traffic is itself processing of employee data, so focus on metadata, inform staff and involve your works council where national law requires it. DataNorth’s AI Security & Shadow AI Assessment maps AI traffic through metadata such as destination, volume and frequency. It only reads individual prompts if you explicitly enable deep packet inspection for high-risk users. That option is intrusive enough to need its own assessment, including a DPIA where required.
2. Assess the data and purpose of each use case
For each tool or use case you find, record the following:
- the tool, the account type and the business owner
- the data entered or connected, including personal, special category and confidential data
- access rights and integrations with other systems
- the people affected, such as customers, applicants or employees
- the intended purpose
- whether outputs serve as a draft, inform a human decision or make the decision
This record shows which rules may apply. Screening applicants, for example, may need a data protection impact assessment under Article 35 of the GDPR and an AI Act risk check.
3. Provide approved AI alternatives
People use shadow AI because it helps them work faster, so a ban without an approved route mostly drives the use out of sight. Offer at least one general-purpose assistant under an enterprise agreement, with data processing terms, single sign-on and admin controls. Publish the approved list and commit to a realistic turnaround for new requests.
4. Establish clear AI usage rules
Keep the policy short enough that people actually read it. It should cover:
- which types of data may go into which approved tools
- prohibited uses, such as inferring staff emotions or deciding about individuals without human review
- the approval route for new tools and new AI features
- which outputs need human review before they are used or published
- how to report data entered by mistake, so a possible breach is assessed in time
No law prescribes this exact policy, but it helps you demonstrate GDPR accountability and can be one of your AI literacy measures.
5. Develop role-specific AI literacy
Article 4 expects measures that fit people’s knowledge and context, so one generic e-learning module rarely does the job. Build a shared baseline, then add what each group needs:
- HR teams need to know that AI used to screen applicants, decide on promotions or evaluate performance is listed as high-risk in the AI Act, with obligations from 2 December 2027, and how the GDPR already limits
- Marketing teams need rules on customer data and fact-checking, and need to know the Article 50 duty to disclose deep fakes, which has applied since 2 August 2026.automated decisions about people.
- Developers need to keep credentials and personal data out of prompts and repositories, and to review generated code before merging.
- Managers need to spot shadow AI in their teams, approve use cases within their authority and know when to escalate.
6. Apply technical and supplier controls
Technical controls make the approved route the easy one and limit the damage when something goes wrong. Depending on your risks, consider single sign-on, new AI features switched off until reviewed, least-privilege permissions for connectors and centrally managed retention settings. Data loss prevention rules can catch obvious cases, such as bulk customer exports.
Supplier review covers the contract: processing terms, sub-processors, data location, security and whether the provider trains on your inputs. No single control solves every risk, and controls only reach the tools you know about.
7. Document decisions and review regularly
Name an owner for AI governance and a business owner for each approved tool. Keep a register of tools, use cases, assessments, guidance and training provided, and corrective actions. These records support GDPR accountability and show your AI literacy measures if a regulator asks.
Reassess when a supplier adds a feature or changes its terms, or when a team finds a new use for an approved tool. A scheduled review, such as a quarterly scan, catches changes nobody reported.
AI literacy turns policy into everyday practice
AI literacy is how rules on paper become habits at work. A policy says what is allowed, and literacy helps people recognise the moments it applies to.
It starts with spotting AI. A sales employee who sees a new “summarise with AI” button in the CRM checks whether the feature is approved before using it on client notes. The second skill is handling sensitive data. Before pasting a spreadsheet into an assistant, an employee checks it for health or salary details and removes them.
Literate employees also check outputs. A finance analyst, for example, verifies the largest figures in an AI-generated variance summary before it reaches the board. They choose the approved tool even when a free one is a click away. And they escalate concerns, raising a colleague’s new transcription app through the approval route rather than letting it spread.
That is how literacy complements discovery and technical controls: discovery shows which tools are in use, controls limit what can go wrong, and literacy guides the small decisions in between. DataNorth’s AI Literacy Workshop is a two- or four-hour session, online or on location, and can be tailored further by having our AI experts interview participants beforehand.
A practical shadow AI checklist
Use this checklist to find gaps in your current approach. It is a working tool, not proof or certification of compliance.
- We know which AI tools and AI features are in use, including features inside existing software.
- Each use case has a record of its data, access, affected people, purpose and role in decisions.
- Employees have approved alternatives for their main AI tasks and a quick way to request new ones.
- Our usage rules cover permitted data, prohibited uses, approvals, human review and incidents.
- Employees receive AI guidance that fits their role and tools, and we keep a record of it.
- Each approved tool has reviewed supplier terms covering data processing, retention and training on inputs.
- Every approved tool has an owner, and staff know how to report an AI incident or a possible data breach.
- We reassess tools and use cases when suppliers, features or purposes change.
Frequently asked questions about shadow AI
Is using a personal AI account for work considered shadow AI?
Using a personal AI account for work is usually shadow AI. If the organisation has not approved the account, it has no contract with the provider and no control over settings, retention or access. The same product under an enterprise agreement, with defined purposes and oversight, can be governed use.
Can approved software still create shadow AI risks?
Approved software can still create shadow AI risks. Vendors regularly add AI features to existing products, sometimes switched on by default. An approved tool can also be used for a purpose nobody assessed, such as screening applicants with an assistant bought for drafting. Review new features and new purposes, not only new tools.
Can banning AI tools solve the problem?
A ban rarely solves the problem on its own. A blanket ban without alternatives tends to push use onto personal devices and accounts, where you see even less of it. Targeted blocks work better when they come with approved tools, clear rules and a fast route for new requests.
Is AI literacy training enough to ensure compliance?
Training alone does not ensure compliance. It supports the AI literacy obligation, but it does not settle questions about lawful basis, supplier contracts, security or high-risk use cases. An enterprise licence or a written policy has the same limits. Compliance depends on the whole approach, from knowing what is used to documenting decisions.
Take control of AI use in your organisation
Shadow AI shrinks when you can see it and offer people a better route. The most useful first step is an accurate picture of the AI already in use and the data flowing into it.
DataNorth’s AI Security & Shadow AI Assessment maps existing AI use and checks each tool it finds against a database of more than 500 AI privacy policies. It ends with a Keep, Replace or Block roadmap and sanctioned alternatives for high-value workflows, and it can be repeated as a quarterly scan. Our AI literacy workshops then help employees apply the rules in their daily work. Get in touch to discuss your situation and request a quote.
