Shadow AI: what it is, the risks, and how to prevent and remediate it
Shadow AI is the use of artificial intelligence tools inside an organisation without the approval or control of the people responsible for IT, security and data protection. The problem is not that people use AI, but that the company does not know which tools are running, what data they receive and where it ends up. Shadow AI is prevented with approved tools and clear rules, detected through inventory and proportionate monitoring, and remediated like any other data incident.
Key points
- Shadow AI is a form of shadow IT with an added risk: the tool does not just store data, it processes it, may retain it and produces output people treat as correct.
- IBM’s Cost of a Data Breach Report 2025 found that about one in five organisations linked a breach to shadow AI, and high levels of shadow AI added roughly USD 670,000 to breach costs.
- The most common forms are not chatbots but AI features quietly switched on in existing applications, browser extensions and meeting assistants that join calls automatically.
- A blanket ban does not work: it pushes use onto phones and personal accounts, where nothing is visible.
- After data leaks through an AI tool, the same discipline applies as for any incident: contain, assess, notify where the law requires it, and fix the cause.
What shadow AI is and how it differs from shadow IT
Shadow IT is any application or service used without IT’s knowledge: personal cloud storage, a messaging app, a project tool paid for on an employee’s card. Shadow AI is the specific case where that tool contains an artificial intelligence model.
The difference matters for three reasons:
- Data is processed, not just stored. A document uploaded to a chatbot is read and summarised and, depending on settings, may be retained or used to improve the service.
- The output has consequences. A wrong contract summary or generated code with a vulnerability ends up in decisions and in production.
- The tool can act. AI assistants and agents connected to email, calendars or storage can read, send and change data on the user’s behalf.
The forms shadow AI takes
| Form | Example | Why it escapes control |
|---|---|---|
| Public chatbots on personal accounts | A contract pasted in for rewording | No contract with the provider; data retention settings belong to the employee |
| AI features enabled in existing apps | Automatic summaries in the office suite, CRM or helpdesk | Arrive through an update, with no new supplier assessment |
| AI browser extensions | A writing assistant that reads every open page | Access to internal applications open in the browser |
| Transcription assistants and notetakers | A bot that joins every meeting in the calendar | Records confidential discussions and sends them outside the company |
| OAuth connections to AI services | An AI app authorised to read the whole mailbox | Access stays live after the employee stops using the app |
| Subscriptions and API keys on personal cards | A developer wires an AI model into an internal script | Code, data and costs appear in no record |
| Models downloaded and run locally | An open-source model installed on a laptop | Provenance of the model and its libraries is unchecked |
Why it happens
Shadow AI is rarely bad faith. It appears where people are under time pressure and have no approved option: the company offers no tool, approval takes months, or the approved tool is weaker than the free one. On top of that, software vendors add AI features to products already in use, sometimes switched on by default.
The risks, grouped by consequence
Loss of control over data. Trade secrets, offers, client and employee personal data, litigation or security incident documents reach a provider with no contract and no verified safeguards.
Legal breaches. Processing personal data through an unapproved tool can breach the GDPR: no legal basis, no processor agreement (Article 28), and possibly no safeguards for transfers outside the EU. Confidentiality clauses in client contracts may also be breached.
Compromised access. Code containing passwords or API keys sent for review, over-permissioned OAuth connections, extensions that read everything in the browser. Some apps marketed as “AI assistants” are in fact malware.
Decisions based on wrong output. AI output can contain errors stated convincingly. Without human review, they end up in reports, offers and code.
No audit trail. If nobody knows who sent which data where, an incident cannot be investigated or even scoped.
Regulatory non-compliance. Since 2 February 2025, the AI Act requires AI literacy for staff using AI systems (Article 4). For essential and important entities in Romania, GEO 155/2024 requires logging and traceability (Art. 11(4)), asset management, access control and supply chain security (Art. 13). An undeclared AI tool is an asset and a supplier missing from the records. Supplier assessment is covered in our article on NIS2 and the supply chain.
How to detect shadow AI
Detection starts with asking, not surveillance. Recommended order:
- Declarative inventory. A short, sanction-free questionnaire: which AI tools you use, for what, with which account, what data you enter.
- Apps authorised via OAuth. The Microsoft 365 or Google Workspace admin console lists apps granted access to company accounts and their permissions.
- Browser extensions on company devices, through device management tools.
- Traffic to known AI services, analysed at connection level (DNS, proxy, firewall), not content.
- Expense claims and subscriptions. Recurring payments to AI providers show up in accounting before any technical report.
- Calendar invitations. Transcription bots appear as meeting participants.
- AI settings in existing apps. Check which AI features are active in the office suite, CRM, helpdesk and development platforms.
Monitoring must be proportionate and announced: staff should know what is checked and why. Otherwise the company solves one data protection problem by creating another.
Prevention: what reduces shadow AI
- Approved tools that are real and useful, with company accounts, single sign-on and a contract that excludes training on your data.
- A short AI use policy with examples. The full 10-point structure is in Shadow AI at work: a 10-point AI use policy.
- Data classification, so every employee knows what may and may not go into an AI tool.
- A fast approval procedure, answering in days, not months. Slow approval produces shadow AI.
- Technical controls: blocking unapproved extensions, restricting OAuth consent for new apps, data loss prevention rules for sensitive categories.
- Regular training on cases from your own work, supporting the AI literacy duty.
Remediation: what to do after it has happened
When sensitive data has gone into an unapproved AI tool
- Stop the flow. The user stops sending data; the connection or extension is disabled.
- Preserve the evidence. Screenshots of the conversation, account history, access logs and the list of files sent. Delete nothing before the assessment.
- Establish what left. Data categories, volume, data subjects, whether credentials were involved.
- Rotate exposed credentials. Passwords, API keys, tokens, certificates. Revoke the app’s OAuth grants.
- Contact the provider. Request deletion and confirmation that the data was not used for training; check the terms of the account used.
- Assess notification duties. Where personal data is involved, a breach must be notified to the supervisory authority (ANSPDCP in Romania) within 72 hours of becoming aware of it, unless it is unlikely to result in a risk (GDPR Article 33); where the risk is high, data subjects must also be informed (Article 34). NIS2 entities should check whether the event is a significant incident reportable to DNSC under GEO 155/2024. Check contractual duties towards clients too.
- Document the incident and fix the cause: which tool was missing, which rule was unclear, which technical control would have stopped the upload.
When you find widespread use but no specific incident
- Announce a sanction-free disclosure period. People say what they use, and the company learns where the real needs are.
- Triage the tools found: some are approved with company accounts, some replaced with an approved alternative, some banned with an explanation.
- Move users to the approved tools, then disable old accounts, extensions and grants.
- Bring the tools into the records: asset inventory, supplier register, risk analysis, record of processing activities.
- Recheck after three months. Shadow AI returns whenever a need appears that the approved option does not cover. The Humanᴬᴵ / Orgᴬᴵ calculator, described below, lets you compare the baseline with the recheck.
Shadow AI through the Humanᴬᴵ / Orgᴬᴵ framework
The measures above treat shadow AI as a control problem. The Humanᴬᴵ / Orgᴬᴵ framework, set out in “Humanᴬᴵ and Orgᴬᴵ: A Relational Framework for AI-Mediated Human and Organizational Capacity” (Alexandru-Ciprian Angheluș, 2026, DOI 10.5281/zenodo.22295278), also treats it as a capacity problem.
- Humanᴬᴵ describes a person’s AI-mediated capacity: what someone can do when working together with an AI tool.
- Orgᴬᴵ describes the organisation’s governed AI capacity: what the company can do with AI, repeatably, with controlled data and clear accountability.
Seen this way, shadow AI is Humanᴬᴵ capacity growing outside Orgᴬᴵ. Staff have become more productive, but the gain stays individual, undocumented and unprotected, while the risk falls on the organisation. When the employee leaves, the capacity leaves too. When they make a mistake, the company is accountable for something it never knew about.
The practical consequence for remediation: the goal is not to eliminate use but to turn it into governed capacity. The tools found in the inventory show where real needs and skills already exist; that is where tools should be approved, rules written and training organised. The useful question for management is not “which app do we ban?” but “which decisions do we let AI assist, with which data, and who is accountable for the result?”.
The framework is explained in full in Humanᴬᴵ / Orgᴬᴵ: what capacity has the organisation built.
Measure capacity with the Humanᴬᴵ / Orgᴬᴵ calculator
The paper comes with a local-first interpretive calculator (DOI 10.5281/zenodo.22884721): the assessment runs on the user’s own computer, without sending answers to an external service. For a topic like shadow AI, an assessment tool that does not export data is a requirement, not a detail.
The calculator fits two points in a remediation programme:
- After the inventory, as a baseline: management sees what AI capacity the organisation has actually built, beyond the list of apps.
- At the three-month recheck, to see whether individual use has been brought into governed capacity or simply pushed back into the shadows.
Step-by-step use, with screenshots, is in the Humanᴬᴵ / Orgᴬᴵ calculator guided tour, and the current version is presented in the calculator launch article.
How ProDefence can help
- CysAI is ProDefence’s AI security intelligence platform, with a catalogue of AI agents and tools and their risk profiles, useful when assessing tools found in the inventory.
- Our cybersecurity training and the CysEdu platform can include safe AI use modules tailored to roles in your company.
- NIS2 consulting brings AI tools into the asset inventory, supplier assessment and access control policies.
- Penetration testing checks the exposure of applications that gained AI features and their connections to external services.
Frequently asked questions
What is shadow AI?
Shadow AI is the use of artificial intelligence tools in an organisation without approval or control by IT, security or data protection. It includes public chatbots used with personal accounts, browser extensions, transcription assistants and AI features enabled in existing applications.
What is the difference between shadow AI and shadow IT?
Shadow IT covers any application used without IT’s knowledge. Shadow AI is the specific case of AI tools, which do not just store data but process it, may retain it and produce output used in decisions.
How can I tell whether my company has shadow AI?
Through a declarative inventory and by checking OAuth-authorised apps in Microsoft 365 or Google Workspace, browser extensions, traffic to known AI services, subscriptions in expense claims and transcription bots in meetings.
Is banning AI tools enough?
No. A blanket ban pushes use onto phones and personal accounts, where the company sees nothing. What works is approved tools, clear data rules and training together.
What should I do if an employee put personal data into a public chatbot?
Stop the use, preserve evidence, establish what data left, ask the provider to delete it and assess whether the breach must be notified to the supervisory authority within 72 hours under GDPR Article 33. If the risk to individuals is high, inform them as well.
How does shadow AI relate to the Humanᴬᴵ / Orgᴬᴵ framework?
In the Humanᴬᴵ / Orgᴬᴵ framework, shadow AI is individual AI capacity (Humanᴬᴵ) growing outside the organisation’s governed capacity (Orgᴬᴵ). Remediation aims to turn that use into governed capacity, not just to remove it. The local-first interpretive calculator that accompanies the paper helps assess the position before and after remediation.
Does shadow AI matter for NIS2 compliance?
Yes. For essential and important entities, an AI tool processing work data is an asset and usually a supplier. It must be inventoried and included in risk analysis, access control and supply chain assessment, in line with Articles 11 and 13 of Romania’s GEO 155/2024.
Sources
- IBM, Cost of a Data Breach Report 2025 (July 2025)
- Regulation (EU) 2024/1689 on artificial intelligence (AI Act), Article 4
- Regulation (EU) 2016/679 (GDPR), Articles 28, 33 and 34
- GEO 155/2024, consolidated text, Legislative Portal
- A.-C. Angheluș, “Humanᴬᴵ and Orgᴬᴵ: A Relational Framework for AI-Mediated Human and Organizational Capacity” (2026), DOI 10.5281/zenodo.22295278
- A.-C. Angheluș, HumanAI / OrgAI Framework Calculator v1.2.0 (2026), DOI 10.5281/zenodo.22884721
Alexandru Angheluș, ProDefence. Author of the Humanᴬᴵ/Orgᴬᴵ framework and of “Artificial Intelligence and Civilizational Transformation”; works on AI security and governance in organisations.



