Shadow AI explained: a guide for EU SME leaders in 2026
If you lead a European small or medium enterprise, a law firm, an accounting practice or a board of directors, here is a sentence we have heard hundreds of times in the past six months: "We do not really use AI yet, so the privacy debate does not apply to us".
This sentence is almost always factually wrong. According to Microsoft's Work Trend Index 2024, confirmed in the 2025 edition, three out of four global knowledge workers already use AI at work. Seventy-eight percent bring their own tools, often outside any corporate policy. The phenomenon has a name: shadow AI. It is the 2026 evolution of the shadow IT problem boards have been dealing with since the early 2010s, but with a critical difference. The data leaving the organization today is not a file on a personal Dropbox. It is a client balance sheet entering the training pool of a foundation model.
In this article we explain what shadow AI is, why it generates personal legal exposure for company directors and partners, the difference between consumer apps, enterprise APIs and self-hosted models, and the first concrete steps to take control without paralysing day-to-day operations.
The starting point: most companies are exposed without knowing it
The numbers are stark. Microsoft Work Trend Index 2024 (published 8 May 2024 in cooperation with LinkedIn, sample of 31,000 knowledge workers across 31 markets) and the 2025 edition both report:
- 75% of knowledge workers use AI at work.
- 78% bring their own AI to work (BYOAI).
- Many do so without authorization, training or corporate policy.
In Italy, where we are headquartered, the Politecnico di Milano AI Observatory's 2025 data show that 47% of workers use AI in the workplace, but only 19% rely exclusively on company-approved tools. The remaining 81% mix personal and corporate tools.
These numbers do not describe an exotic risk. They describe the median European SME. If your company has 50 employees, statistically dozens of them are using ChatGPT, Gemini, Claude, or an automatic transcription service on a regular basis. The question is not whether, but what data they are exposing.
The cause is not malice. It is perceived efficiency. An employee facing a deadline pastes a draft contract into ChatGPT for a quick summary. Forty minutes saved. No one knows. No one asked. Multiply this by a realistic frequency per week, across all the employees of a typical mid-size firm, and you have your real risk surface.
Why shadow AI is a personal legal risk, not just a corporate one
This is where many board conversations go wrong. Directors often frame shadow AI as a "corporate IT issue" to be delegated to the CIO or DPO. The reality is more uncomfortable: shadow AI generates personal legal exposure for those who hold fiduciary duties.
For accountants and tax practitioners
Client financial data pasted into consumer AI tools can trigger:
- Breach of professional secrecy, a duty defined in every European jurisdiction (in Italy, Art. 622 of the Criminal Code; equivalent provisions in France, Germany, Spain).
- Violation of GDPR Art. 28 (no data processing agreement with the AI sub-processor).
- Violation of GDPR Art. 32 (inadequate security measures).
- Disciplinary sanctions from professional orders (ODCEC in Italy, equivalent bodies elsewhere).
For law firms
In addition to the above:
- Special category data under GDPR Art. 9 (health data, sometimes embedded in case files).
- Judicial data under GDPR Art. 10 (criminal records, ongoing investigations).
- Bar association ethics rules, requiring qualified human supervision in every European jurisdiction.
For company directors and board members
This is where the personal exposure is most acute. In every European jurisdiction, directors hold fiduciary duties to the company and shareholders. In Italy these are codified in Codice Civile articles 2392 and 2394. In Germany in Aktiengesetz Sec. 93. In France in Code de Commerce L. 225-251. In the UK, in Companies Act 2006 section 172. In Delaware DGCL Sec. 141 (relevant for many EU-listed groups with US subsidiaries).
When strategic data (M&A planning, financial projections, contracts in negotiation, customer pricing) is pasted into unauthorized AI tools and a damage materializes (a leak, a breach, a regulator action), directors face personal civil and, in aggravated cases, criminal liability. The corporate veil does not protect against personal failure of fiduciary duty.
This is why shadow AI is not "an IT problem to delegate to the DPO". It is a board-level risk that requires board-level attention.
Three things people confuse under the label "generative AI"
Most bad decisions on AI start from a category error. Three very different things get grouped under the same label "generative AI". We explain each.
Consumer apps
ChatGPT on chat.openai.com, Gemini on gemini.google.com, Claude on claude.ai, Microsoft Copilot with personal account. Free or near-free web interfaces, designed for individual end users.
How they treat data. On the free or consumer tier, content entered by the user can be used to improve the models, unless an explicit opt-out is activated. Opt-outs exist but are often hidden, not retroactive, and not on by default.
The concrete risk. A client balance sheet pasted into consumer ChatGPT enters a system that has:
- No data processing agreement signed with your company (GDPR Art. 28 violation).
- No contractual guarantee of non-training.
- No accessible audit logs.
- No guaranteed European residency.
- No clauses on sub-processors.
The risk is not "the balance sheet shows up tomorrow in a response to a competitor". The risk is more subtle: patterns, formulations, peculiar numbers can be learned and can resurface. Proving disclosure ex post is impossible. Proving non-disclosure, equally so.
Enterprise APIs
OpenAI API, Anthropic API, Google Gemini API via Vertex AI, Azure OpenAI Service, AWS Bedrock. Programmable interfaces used inside business software and platforms.
How they treat data. Data sent via enterprise APIs is not used for model training. This is the contractual default of all four major vendors. DPAs are signed (click-through or negotiated). Retention windows for abuse monitoring are limited:
- OpenAI: 30 days default (zero for EU Projects or approved Zero Data Retention).
- Anthropic: 7 days since 14 September 2025 (zero with approved ZDR).
- Google Gemini: 55 days default, configurable on Vertex AI.
- Azure OpenAI: configurable, ZDR available.
On request, Zero Data Retention nulls even the abuse monitoring window.
The critical difference. When a professional says "I use ChatGPT for analysis", the right question is: "Through the consumer browser with a personal account, or inside business software calling the enterprise API?" That difference accounts for nearly the entire risk delta.
A system like TDS AI Systems uses enterprise APIs exclusively. Client data never passes through consumer interfaces. It passes through programmatic calls under signed DPAs, with training opt-out active, with auditable logs, and with European residency selected where available (Vertex AI europe-west1/europe-west4 as the default for EU-sensitive data).
Self-hosted models
AI models (Llama from Meta, Mistral, Qwen from Alibaba) installed on the company's own servers or fully controlled cloud infrastructure (Modal, AWS Bedrock with dedicated instances, Vertex AI with customer-managed encryption keys).
How they treat data. Data does not leave the organization's perimeter. No third-party sub-processor sees the request. Maximum control.
The cost. Significant. GPU hardware, ops capacity, LLMOps expertise. Models that, while progressing, do not always match the frontier of closed-source providers on complex tasks.
When it makes sense. Not for everyone. It makes sense for genuinely top-tier data: board documentation for M&A, healthcare data under GDPR Art. 9, judicial files for sensitive matters, contracts with explicit no-sub-processor clauses.
For most use cases in an SME or professional firm, enterprise APIs with signed DPAs and training opt-out cover the risk adequately. Self-hosted is for ultra-sensitive data or specific contractual requirements.
The applicable European regulatory framework
Three overlapping layers, all in force.
GDPR (Regulation EU 2016/679)
In force since May 2018. Applies to any processing of personal data, including via AI tools. Key articles for AI use:
- Art. 5: lawfulness, fairness, transparency, purpose limitation, minimization, accountability.
- Art. 6: lawful basis required for any processing.
- Art. 9: special categories (health, biometric, religious, etc.) reinforced regime.
- Art. 13-14: information duty toward data subjects.
- Art. 28: written DPA with every sub-processor.
- Art. 32: appropriate security measures.
- Art. 35: DPIA for high-risk processing.
- Art. 44-49: extra-EU transfers (Standard Contractual Clauses, Data Privacy Framework, adequacy decisions).
EU AI Act (Regulation EU 2024/1689)
In force since 1 August 2024 with progressive enforcement. Applies to anyone who places on the market, puts into service, or uses AI systems in the EU.
Key articles for SMEs and firms:
- Art. 6 and Annex III: classification of high-risk AI systems.
- Art. 9: continuous risk management for high-risk systems.
- Art. 13: transparency requirements for high-risk systems.
- Art. 14: meaningful human oversight for high-risk systems.
- Art. 50: transparency obligations for generative AI, deepfakes and chatbots. Enforcement from 2 August 2026. Mandatory labelling of AI-generated content and disclosure when interacting with AI systems.
National implementations
Italy passed Law 132/2025 ("Provisions and delegations to the Government on artificial intelligence"), in force since 10 October 2025. It complements (does not replace) the EU AI Act. We cover the Italian framework in detail in our dedicated article (Italian, contesto normativo italiano IA 2026).
Equivalent national implementations are in progress across other EU member states. France, Germany, Spain, the Netherlands are at various stages. Penalties scale on the EU AI Act grid: up to 35 million euros or 7% of global annual turnover.
What you should do this week
If you lead an SME or professional firm in Europe, here are the first concrete steps. None of them require external consultants. None of them require waiting for the perfect framework.
Step 1: shadow AI audit
Anonymous internal survey (Tally, Google Forms, equivalent). Key questions:
- Which AI tools have you used in the last week?
- Personal account or company account?
- Did you enter client data, contracts, financials, internal documents?
- Do you know the difference between consumer AI apps and enterprise APIs?
Aggregate answers in numbers. Classify usage by traffic light: green (authorized on safe tools), yellow (tolerable if formalized), red (consumer with client data, to block).
Total time: half a working day. Output: an internal two-page document that you can show to a regulator or an enterprise client on request.
Step 2: internal AI policy
A two to three page document, distributed to all staff with signed acknowledgment. Minimum content:
- What is allowed (sanctioned tools and approved use cases).
- What is not allowed (consumer apps with client data, off-list tools, PII anywhere unapproved).
- List of authorized tools (few and well-defined; better 3 approved than 30 tolerated).
- Authorization process (whom to ask, response time, criteria).
- Sanctions (disciplinary; in serious cases, breach of confidentiality clauses).
Total time: one day of writing plus half a day of review.
Step 3: minimum training
One hour session, in-person or video, for all personnel. Content: shadow AI, consumer vs enterprise vs self-hosted distinction, what counts as sensitive data, what to ask before using a new tool, how to report incidents.
Mandatory record-keeping: attendance list signed, date, material archived. This is what evidences the measures adopted under Art. 4 of the AI Act (Reg. (EU) 2024/1689, as replaced by Reg. (EU) 2026/1744 in force since 27 July 2026: an obligation to adopt proportionate measures, not to guarantee a given level per person), and it supports Art. 5(2) GDPR accountability. Where a DPO has been appointed, staff awareness and training fall among their tasks under Art. 39(1)(b) GDPR.
Step 4: authorized infrastructure
Replace red-flagged shadow AI with sanctioned tools:
- Microsoft Copilot M365 Business Standard if you are already on M365. DPA included, EU residency for many regions.
- ChatGPT Enterprise / Team, Claude for Work for familiar UX under business DPA.
- Vertical platform like TDS AI Systems for multi-model access under one interface, enterprise APIs, PII redaction upstream, centralized audit logs, vertical-specific configuration (law firms, accounting practices).
In every case: signed DPA, training opt-out active, accessible audit logs, EU residency where available.
Step 5: quarterly review
Half a day per quarter. Review: AI market changes, vendor updates, new regulator decisions, internal feedback. Continuous capability, not project.
Practical checklist
What to do in the next 30 days:
- Designate an internal AI governance owner (with name and surname, 4 hours per week).
- Run the shadow AI audit (half a day, no consultant needed).
- Write and distribute the internal policy (one to two days).
- Deliver minimum training to all staff with attendance record (one to three days, including prep).
- Select authorized infrastructure and sign vendor DPAs (timing varies; 1-3 weeks).
- Schedule the first quarterly review at 90 days.
What to avoid:
- "We do not use AI" without an audit. Statistically wrong in the vast majority of cases.
- Generic bans without alternatives. Generates more hidden shadow AI.
- Waiting for the "perfect framework". An applicable framework already exists (GDPR + AI Act + national laws).
- Delegating everything to an external consultant without internal ownership of day-to-day operations.
- Confusing "we bought an enterprise license" with "we did governance".
Shadow AI is not an exotic risk. It is the median state of European SMEs and professional firms in 2026. The distinction between exposed and protected organizations is not between those who know everything and those who know nothing. It is between those who have started a documentable governance journey and those who have not.
For the deeper Italian-language treatment, including the legal framework specific to Italy (Codice Penale Art. 622, Codice Civile Art. 2392-2394, Law 132/2025), see our article on shadow AI: il rischio nascosto che il tuo studio gia' corre.
Continue your journey
If you want our next deep-dives on AI governance, compliance and automation for European professional firms and SMEs, subscribe to TDS Weekly: one email every two weeks, no spam, explicit confirmation.
Related reading: