Shadow AI: bring useful work under company control

Consider a salesperson pasting a client contract into a personal AI account to prepare a response. The work may improve while the company loses visibility over where the document went.

That is a scenario to investigate with the team. Ask what was sent, which account was used, whether the use was authorized and what the provider retains.

Establish what is happening

An inventory should distinguish an approved enterprise service from an unmanaged personal account. Product names alone do not tell you the contract, retention settings or access controls.

What is happening in practice is this. Your employees are intelligent people with hard problems and access to extraordinary tools. ChatGPT. Claude. Gemini. Perplexity. They use them daily. They upload documents, paste internal data, write prompts with client names and financial figures and HR details. They get results. They keep going.

From a productivity standpoint, it works.

From a compliance standpoint, it is a live wire.

GDPR and NIS2 do not impose a blanket contradiction. GDPR Article 32 requires security appropriate to risk. For covered entities, NIS2 Articles 21 and 23 address proportionate cybersecurity measures and significant-incident reporting. Monitoring still needs a lawful, proportionate design that protects employees’ rights.

The regulatory tension is real but it is not even the most immediate problem. The most immediate problem is that you are operating with a blind spot at the core of your business. Confidential client information is leaving your systems through channels you did not build, did not authorize, and cannot audit. The people doing it have no bad intentions. They are just trying to do their jobs faster, with the best tools available to them.

The fact that it is well-intentioned does not change the exposure.

A workforce announcement elsewhere cannot establish the return or safety of your own AI use. Your evidence is the workflow, its outputs and the handling of company data.

But there is another version of the AI story that is happening right now, in a quieter way, in companies that have not made any announcement. It is the company where AI is everywhere, unstructured, undocumented, without any central visibility. Where your sales team is drafting proposals with Copilot, your finance director is running analysis with tools that have never seen the inside of a security review, and your operations manager is summarizing board materials with Claude on a personal account.

Everyone is more productive. Nobody has any idea what data went where.

When the question arrives, and it will arrive, from a client doing a security review, from a regulator, from a prospective acquirer running due diligence, the answer we did not really track that is not a good one.

You have two options when your employees are using AI on company data.

Provide an approved route for useful work

Restrict high-risk uses while offering an approved way to do the underlying job. Some data or actions should remain prohibited even inside an otherwise useful tool.

A policy needs practical enforcement: managed accounts, access boundaries and a route for exceptions.

A motivated employee with a personal laptop and a browser extension will find a way. The cost of prohibition is also steep in a different sense: you lose the productivity gains that your competitors are capturing, and you create resentment among your best people who have seen what these tools can do and know you are asking them to go back.

The companies that end up ahead are the ones that channeled it instead.

Channeling means building a corporate AI environment. Your data, under your control, with your permissions structure, your audit trail, your policies. Not consumer ChatGPT. A system your legal and security teams can actually account for when someone asks. A system where the same capabilities your employees are already using are available within a structure you can manage.

This is a governance decision.

Start with one approved use. Confirm the data classification and supplier terms, set retention and permissions, test outputs, then name the person responsible for operating it.

The implementation scope depends on connectors, data sensitivity, access inheritance and the operating team. Agree it before promising a launch date.

NIS2 required national transposition by 17 October 2024, with measures applied from 18 October. Coverage depends on sector, size, exceptions and national law. Confirm applicability for your entity; the directive is not a rule for every mid-market company.

The practical question is not whether to act. The practical question is what acting looks like in a company your size, with your specific mix of clients, data types, and regulatory exposure.

Your employees are already using these tools. The question is whether it is happening on your terms or theirs.