Shadow AI: risks of using ChatGPT without controls

Practical case at Impulsa3: channelling Shadow AI through a shared system

When we expanded the use of AI at Impulsa3, we understood that banning tools would not solve the problem. Our response was to channel adoption through I3OS: a working context for each client, an inventory of more than 80 skills, authorised sources, clear boundaries and human review. Our experience is teaching us that the alternative to Shadow AI is not more surveillance, but a useful, secure and shared way of working. This turns knowledge from something dependent on private conversations into a capability that can improve the entire organisation.

We can tell you how we did it at Impulsa3

68% of employees who use generative AI at work do so without their company’s knowledge. This is not a productivity problem — it is a problem of data, regulatory compliance and governance that you need to address before it becomes an incident

Imagine this scenario: a salesperson at your company copies the customer database from the CRM, pastes it into ChatGPT and asks it to generate personalised follow-up emails. Productivity soars. The salesperson closes more deals. Everything looks like a success. Until you realise that they have just shared your customers’ names, email addresses, purchase volumes and contact details with a third-party service without consent, without an impact assessment and without the company even knowing.

This is Shadow AI: the use of artificial intelligence tools by employees without approval, oversight or organisational governance. And it is not a marginal phenomenon. According to recent studies, 68% of employees who use generative AI at work have not told their managers. In companies without an AI policy, the figure rises to 80%.

The paradox is that Shadow AI often works. The employees who adopt it tend to be the most productive and proactive. The problem is not the tool: it is the lack of a framework that allows it to be used safely, traceably and in compliance with the regulations.

In this article, we explain why Shadow AI is growing, what the real risks for your company are, how to implement an AI usage policy that works without killing productivity and why the AI Act makes this issue urgent.

Why Shadow AI is growing (and will not stop)

Shadow AI is not an act of rebellion. It is employees’ rational response to an asymmetric situation:

The tools are accessible: ChatGPT, Claude, Gemini, Copilot… any employee with a browser can access a generative AI model in 30 seconds. They do not need to install anything or ask the IT department for permission.
Productivity is immediate: An employee who uses AI to draft emails, summarise documents, create drafts or analyse data saves between 30 minutes and 2 hours a day. The reward is immediate and tangible.
The company offers no alternative: Many organisations have still not deployed corporate AI tools. While the company deliberates, employees are already using the free version of ChatGPT with their personal accounts.
There is no clear policy: If the company has not communicated which tools are allowed, what data may be shared and which uses are acceptable, the employee interprets it as “it is not prohibited, therefore it is allowed”.

Key fact: Shadow AI often generates better ROI than formal AI projects, precisely because it starts with a real employee need. The challenge is not to eliminate it, but to channel it within a governance framework.

The real risks of Shadow AI for your company

The risks of Shadow AI are not theoretical. They are operational, legal and reputational:

Confidential data leakage

When an employee enters information into ChatGPT (the free version or Plus without enterprise configuration), that data may be used to train the model. This includes customer data, commercial strategies, source code, financial information and internal documents. In 2023, Samsung discovered that several engineers had entered proprietary code into ChatGPT. Apple and JPMorgan banned the internal use of ChatGPT for the same reason.

Regulatory non-compliance

Sharing customers’ personal data with an AI service without a legal basis (GDPR), without a DPIA and without a data processing agreement (DPA) is an infringement that can result in penalties. Under the AI Act, the lack of traceability regarding which AI systems your organisation uses adds another layer of risk. You cannot meet transparency and governance obligations for systems you do not even know exist.

Inconsistency and undetected errors

An employee who uses AI to generate content, reports or analyses without supervision may produce hallucinations (false information generated with confidence) that become part of official documents. If nobody reviews the output, a fact invented by the model may end up in a sales proposal, a financial report or a customer communication.

Loss of intellectual property

AI-generated content may have intellectual property implications. If an employee generates a design, advertising copy or code with AI without documenting it, the company may face claims over ownership of that content.

You cannot govern what you do not know exists.

AI usage policy: the solution is not to ban, but to govern

Banning generative AI at work is like banning the internet in 2005: you can try, but you will only make employees use it secretly. The solution is an AI usage policy that balances productivity with security.

We call this AI Lite Policy: a minimum viable framework that you can implement in one week and that covers the critical risks. It is based on five principles:

Proportionality: Not all AI tools need the same level of control. Classify tools by risk level: low (draft generation without sensitive data), medium (internal data analysis), high (processing customers’ personal data). Each level has proportionate controls.
Accountability: The employee who uses AI is responsible for the output. AI is a tool, not a substitute for professional judgement. All AI-generated content must be reviewed before being used externally.
Grounding: AI outputs must be checked against verifiable sources. No AI-generated data should be considered reliable until it has been verified against an original source.
Observability: The company must know which AI tools are being used, for what purpose and with which data. This is not to monitor employees, but to meet the traceability obligations of the AI Act and GDPR.
Reversibility: Any process that incorporates AI must be reversible to its non-AI version in the event of a failure, incident or regulatory change. This means knowledge cannot reside only in the model.

What data should never enter a non-corporate AI tool

The most operational part of the policy is to define clearly which data are absolutely prohibited in unapproved AI tools:

Customers’ personal data: Names, email addresses, telephone numbers, addresses, purchase histories and payment data. Sharing them without a legal basis is a GDPR infringement.
Internal financial data: Profit and loss accounts, forecasts, product margins and pricing strategies. Sensitive competitive information.
Proprietary source code: Algorithms, business logic and integrations with internal APIs. Risk of intellectual property leakage.
Legal and contractual documents: Customer contracts, non-disclosure agreements and commercial terms. Direct legal risk.
Confidential internal communications: Committee minutes, strategic decisions, staff assessments and M&A information.

Simple rule: If you would not share it by email with a stranger, do not paste it into ChatGPT. It works as a quick test for 90% of cases.

How to channel Shadow AI in four steps

The transition from Shadow AI to governed AI follows a practical process:

Audit current usage. Before legislating, find out what is happening. Conduct an anonymous employee survey: which AI tools do you use? For which tasks? With what data? The variety of uses and the depth of adoption will surprise you.
Publish the usage policy. Share the AI Lite Policy. Keep it short (2–3 pages), clear and full of concrete examples of what can and cannot be done. Include a channel for questions.
Provide corporate alternatives. If you ban free ChatGPT but offer no alternative, employees will return to Shadow AI within a week. Deploy a corporate version (Microsoft Copilot, ChatGPT Enterprise, Claude for Work) with a signed DPA, data not used for training and corporate SSO.
Monitor and improve. Maintain a register of approved AI tools. Review usage quarterly, collect employee feedback and adjust the policy. Include AI tools in the inventory required by the AI Act.

Concrete examples: what your employees can do with AI

A useful AI policy does not only say what is prohibited. It also defines what is allowed and encourages it. These are examples of low-risk uses that you can approve from day one:

Writing and proofreading: Use AI to improve the wording of emails, sales proposals or internal documentation, provided that no confidential customer data is included. Example: “improve the wording of this paragraph” is valid. “Improve this email for Acme S.L. with their €500K turnover” is not.
Research and summarising public information: Summarise articles, compare competitors using public data and prepare industry briefings. No internal data involved.
Generic code generation: Ask for help with standard functions, generic SQL queries or automation scripts without revealing the system’s internal architecture.
Brainstorming and creativity: Generate ideas for campaigns, product names and presentation structures. AI as a creative sparring partner without sensitive data.
Training and learning: Use AI as a tutor to learn new skills, understand technical concepts or practise languages.

The principle is simple: if the task does not require internal, confidential or personal data, it is probably a low-risk use that you can approve. When an employee needs to use internal data, they must use the approved corporate tool, not the free version.

Tip: Create an internal channel (Slack, Teams) where employees share how they use AI to be more productive. This normalises usage, creates collective learning and gives you visibility into real adoption.

AI Act and Shadow AI: why the urgency is real

The AI Act introduces traceability and inventory obligations that make Shadow AI especially problematic:

Inventory obligation: Organisations must know which AI systems they use, especially if any fall into the high-risk category (Annex III). If an HR employee is using ChatGPT to screen CVs, that is a high-risk AI system that should be registered, documented and supervised.
Transparency: The AI Act requires people interacting with an AI system to know that they are doing so. If an employee uses AI to generate customer-service replies without the customer knowing, the company is in breach.
Deployment responsibility: Under the AI Act, responsibility lies with whoever deploys the AI system, not with whoever built it. If an employee deploys ChatGPT for a high-risk use without authorisation, the company is still responsible.

Penalties for non-compliance with the AI Act can reach 3% of global turnover. But beyond penalties, a data-leak incident through Shadow AI causes disproportionate reputational damage: the narrative of “a company that does not even know which AI tools its employees use” is devastating to trust.

Shadow AI is an opportunity disguised as a risk

Shadow AI is telling you something valuable: your employees already see AI’s potential and are adopting it on their own initiative. That is a sign of innovation, not a threat. But without governance, this spontaneous innovation becomes a legal, operational and reputational risk vector.

The answer is not to ban it. It is to channel it. An AI Lite Policy implemented in one week, corporate tools with privacy safeguards and an inventory of AI uses allow you to take advantage of Shadow AI’s productivity while eliminating its risks.

If you want to audit Shadow AI in your organisation, design an AI usage policy or deploy governed corporate tools, Impulsa3 can help you create your AI usage policy and implement it with you.