Shadow AI: how the new shadow IT is becoming a security problem
Employees are using AI tools without IT department oversight. The pattern is familiar - but the risks are different from classic shadow IT.
Shadow IT is something companies have been fighting for the past fifteen years. Employees used Dropbox instead of the corporate file store, Telegram instead of the approved messenger, Google Docs instead of SharePoint. This created data management and access control problems - but the scale of risks was understood.
With AI tools the situation is different. Not because people behave differently - they do the same thing: find a convenient tool and start using it. The difference is that the volume of data flowing through AI tools, and the nature of that data, has fundamentally changed.
How shadow AI differs from shadow IT
When an employee uploaded a file to Dropbox, the data went "somewhere". It was a file, a document, a spreadsheet. Control was lost, but the content stayed fixed.
When an employee pastes a contract excerpt into an LLM chat and asks it to "improve the wording" - several things happen simultaneously. Data is transmitted to a third-party provider. Depending on the terms of service it may be used for model fine-tuning. The query context may contain more information than the employee realises - counterparty names, amounts, terms.
This is not a hypothetical risk. It is already happening in most companies where the IT department has not established a clear policy.
Typical scenarios worth knowing about
Copying client data for processing. A sales manager pastes a CRM export to ask AI to draft emails. The export contains names, contacts, and relationship history.
Uploading internal documentation. An employee uploads a policy or presentation for summarising. The document contains internal strategy, team structure, financial targets.
Developers with code. An engineer pastes code for review or debugging. The code contains business logic, data schemas, sometimes strings with access keys.
Financial data analysis. A finance person pastes a P&L table for a quick trend analysis.
What to do about it
Banning AI tools is the least realistic response. It does not work technically and creates cultural friction that leads to even less transparency.
A sensible approach includes several elements.
An explicit policy with concrete examples. Not "confidential data must not be shared" but a specific list: what can be pasted into public AI tools, what cannot, what needs to be anonymised first.
Corporate tools as an alternative. If the company has LLM access through a corporate contract with appropriate data processing terms - direct employees there. This reduces the incentive to use public services.
Training rather than just prohibition. Most employees do not think about where data goes. Concrete examples of what can happen work better than abstract prohibitions.
DLP monitoring. If data loss prevention tools are in place - configure them to detect patterns of data uploads to AI services.
A few questions for a quick check
- Do you have a current AI tool usage policy that employees are aware of?
- Have you checked which AI tools are actually being used across the company?
- Do you have a corporate contract with an LLM provider that has explicit data processing terms?
- Have you trained employees with concrete examples of what is and is not acceptable?
Shadow AI is not a reason to panic and not a reason to prohibit. It is a signal that policies and tools need to be brought into alignment with how people actually work.