Shadow AI: the risks, and how to find and govern it

Shadow AI is AI use your organisation didn’t sanction and doesn’t see. Your employees are almost certainly doing it already: pasting customer data into a chat assistant to draft replies, uploading contracts to a summariser, generating code with a public model. Most of it is well-intentioned, people trying to do their jobs faster. The problem is that none of it went through your security review, and some of it is quietly leaking data outside the company.

Banning it rarely works, because the productivity gain is real and people find workarounds. The better approach is to find it, understand it, and bring the risky parts under control.

Is shadow AI a security risk?

Yes, though not every case is. The risk depends on what goes into the tool, and four things tend to go wrong:

• Data leaves the company. Whatever someone pastes is now held by a vendor you haven’t assessed, under terms you haven’t read. Some services keep inputs or use them to improve their models unless you’re on a business plan that says otherwise.
• You break your own obligations. Customer contracts, sector rules and data protection law (including India’s DPDP Act and Rules) usually say where personal data may go and who may process it. An unapproved AI tool is an unapproved processor.
• Wrong answers get used as fact. An AI-drafted contract clause or customer reply goes out without anyone checking it, and nobody can later say which tool produced it.
• Leaked credentials and code. Developers paste config files and stack traces into assistants, and secrets go with them. We covered why that matters in Keeping secrets out of public code.

A summariser used on public marketing copy is a minor issue. The same tool used on employee records is a serious one. That’s why the playbook below sorts by data, not by tool.

Step 1: Find out what’s actually happening

You can’t govern what you can’t see. Start by making the invisible visible.

Ask people directly. A short, anonymous survey — “which AI tools do you use for work?” — surfaces more than you’d expect, and it signals that the goal is control, not punishment.

Look at the network. Check which AI domains your employees are reaching: chat assistants, code tools, summarisers, image generators. Your network or email security tools already have this data if you go looking.

Watch the expenses. Corporate cards and expense reports often show AI tool subscriptions hiding under vague descriptions.

Step 2: Sort it by risk, not by rule-breaking

Once you have a list, resist the urge to approve or ban everything. Sort each tool by two questions: what data could reach it, and what could go wrong if it did.

A coding assistant on public code with no customer data is low risk. A summariser where staff upload customer contracts or employee records is high risk, because that data is now with a third party you haven’t assessed. The goal of the sorting is to find the few tools where real harm is possible, not to produce a spreadsheet.

Step 3: Redirect, don’t just prohibit

For the high-risk tools, banning them with no alternative usually just drives the behaviour underground. Where you can, offer a sanctioned path: an approved assistant, an internal model gateway, a policy that says which tools are allowed for which data.

This is the part where the work becomes a proper governance programme. An intake form for new AI requests, a standard risk review, and a register of what’s approved — that’s exactly what the AI Risk & Governance Kit provides, and it maps to the requirements in ISO/IEC 42001.

Step 4: Put the policy where people will read it

Write a short acceptable-use policy that answers the questions people actually have: “can I use this tool with customer data?”, “which tools are approved?”, “what do I do if I already sent something I shouldn’t have?”. Keep it to a page. If it’s long enough to be a PDF nobody opens, it isn’t working.

Step 5: Repeat

Shadow AI isn’t a one-time cleanup. New tools appear constantly, and people change jobs and habits. Run the discovery step again each quarter. The good news is that after the first pass, the process gets fast, and the risky behaviour tends to shrink once people know someone is paying attention.

If you’d like this set up properly rather than improvised, our Compliance & ISO/IEC 42001 service includes the governance structure, and the AI Risk & Governance Kit gives you the intake form, risk review and register to run it yourself.