Shadow AI: how to map actual AI use in 30 days?

Shadow AI: how to map actual AI use in 30 days?

Why shadow AI is the company's primary AI risk

The inventory of AI systems is the starting point for all AI Act obligations: it is impossible to train users (Article 4), rule out prohibited practices (Article 5), ensure transparency (Article 50) or prepare for the high-risk regime on uses that are unknown. On the GDPR side, each prompt containing personal data is an undocumented processing activity; on the trade secrets side, each entry of confidential information into a public tool weakens legal protection (see the article ChatGPT in business). And experience shows that the gap between declared use and actual use is massive: AI entered through individuals, not through the CIO.

Week 1: voluntary declaration, with amnesty

Launch a short declarative survey (5 minutes), anonymised at team level, with a clear message from management: the objective is to equip, not to punish; no blame will be attached to past uses. Three questions are enough: which tools, for which tasks, with which data. Amnesty is not naivety, it is the condition for honesty: a punitive survey produces an empty map and even more hidden uses.

Week 2: technical measurement

Cross-check the declarations with actual traces:

  • Proxy, firewall and DNS logs: connection volumes to the domains of the main AI services (chatgpt.com, claude.ai, gemini.google.com, perplexity.ai, mistral.ai and others).
  • SaaS console and discovery tools (CASB, Microsoft 365 and Google Workspace reports): third-party applications connected to accounts, installed browser extensions.
  • Embedded AI features in SaaS already under contract (CRM, support, videoconferencing, office suites): often enabled by default, they are part of the map.

The measurement remains aggregated: the aim is to map uses, not to monitor individuals (see the FAQ on the legal framework).

Week 3: qualify each use

Classify each tool x use pair on two axes:

  1. The level of regulatory risk: prohibited practice (Article 5: to be stopped immediately), future high risk (Annex III: HR, credit, education: to be framed as a priority), transparency (Article 50: content and agents exposed to the public), minimal risk.
  2. The sensitivity of exposed data: personal data, sensitive data, trade secrets, client data under confidentiality clause, public data.

Also determine the company's role for each system: simple deployer, or de facto provider if a tool has been substantially modified or rebranded (requalification is the classic trap).

Week 4: decide, equip, frame

  1. List of validated tools by use case, with enterprise versions under data processing agreement: the best weapon against shadow AI is an official alternative as convenient as the hidden tool.
  2. AI use policy short and enforceable (see the 8-point charter), integrated into the IT charter.
  3. Article 4 training planned and tracked for all identified users, with enhanced module for sensitive uses.
  4. Targeted blocking reserved for indefensible cases (tools without any guarantees, prohibited practices), explained and accompanied by an alternative.
  5. Continuous governance: an AI officer, a quarterly review of the map, a simple channel to request validation of a new tool in less than a week: if the official process is slow, shadow AI returns.

The indicators that prove it is under control

  • Share of AI uses covered by a validated tool (target: over 80% at 6 months).
  • Number of validated tools in the catalogue and average time to validate a request.
  • Rate of trained employees (Article 4) and date of the last map update.
  • AI-related incidents reported and handled (a figure rising at the start is a good sign: we can finally see).

Take action

Manage your compliance with Data Comply One: all-in-one software and outsourced DPO/CISO experts.

Discover the AI Act software →Test my AI compliance →

Sources

  • Regulation (EU) 2024/1689 (AI Act), Articles 3, 4, 5, 25, 26 and 50, EUR-Lex.
  • Regulation (EU) 2016/679 (GDPR), Articles 5, 6, 13 and 35, EUR-Lex.
  • The French data protection authority (CNIL), recommendations on workplace monitoring tools and on AI systems; French Labour Code, Article L.2312-8.

Frequently asked questions

Would it not be simpler to ban AI outright?

This is the strategy that fails fastest: use shifts to personal phones and the company loses both visibility and the productivity benefit. Governance with validated alternatives is safer than prohibition without alternatives.

Are we allowed to analyse employees' connection logs?

Yes, provided GDPR and employment law are complied with: legitimate and declared purpose (security, mapping of use), proportionate and aggregated analysis, prior information to employees, and consultation of the works council if the system constitutes activity monitoring.

What should be done with discovered uses that fall under prohibited practices?

Immediate cessation, documentation of the discovery and remediation, and analysis of past exposure. Article 5 practices (social scoring, emotion recognition at work except in provided cases, etc.) are sanctioned at the highest ceiling of the AI Act.

Must the recorded uses be declared to an authority?

No for a deployer company in 2026: the inventory is an internal compliance tool. It will become the basis for the high-risk obligations of December 2027 for uses in Annex III.

How often should the exercise be repeated?

The map is maintained continuously via the validation channel; a quarterly review and an annual re-survey are then sufficient. The first inventory is the most costly, the following ones are updates.