What is shadow AI?
By Nick Jang, founder of MINERVA Studio ·
The short answer
Shadow AI is AI that people use for work without the organization knowing about it or approving it. Think of a personal chatbot account used to summarise client files, an AI feature a vendor quietly switched on in software you already pay for, or a browser extension that reads every page an employee opens. It’s the AI version of shadow IT, and the main risks are data leaving the company, decisions nobody can explain, and tools nobody has assessed.
Why it happens
Shadow AI is mostly good intentions. Someone finds a tool that saves them an hour a day and uses it, because the approved route is slow or doesn’t exist. Nobody is trying to break a rule; there often isn’t one.
It comes in three shapes:
- Personal accounts. Staff use their own chatbot subscriptions, often the free tier, for work.
- Embedded AI. A product you already bought adds an AI assistant in an update. Nobody bought AI, but now you have it, with its own data handling.
- Home-built. A team wires a model API into a spreadsheet or script to automate a task, outside any project or review.
The second kind is the one most organizations miss. It doesn’t show up in a purchase order, because nothing new was bought.
What it puts at risk
| Risk | Example |
|---|---|
| Data leakage | Client details pasted into a consumer tool whose terms allow it to keep or learn from the input. |
| Privacy and legal | Personal information sent to a provider and a country no privacy assessment ever covered. |
| Unchecked decisions | An AI-drafted credit memo or claims note used without anyone checking the reasoning. |
| Model risk | A home-built model that scores customers but isn’t in the model inventory, so nobody validates or monitors it. |
| Security | Browser extensions and plugins with broad access to email, files or internal systems. |
| Explainability | When a customer or examiner asks why, there’s no record of what the tool said or why it was trusted. |
How to find it
- Ask. A short, no-blame survey of teams: which AI tools do you use, for what, and with what data? People usually answer honestly when they’re not in trouble.
- Check network and identity logs for traffic to known AI services and for sign-ins with work email to AI tools.
- Read vendor release notes and contracts. Look for AI features added to products you already use, and whether your data is used to train them.
- Review expense claims for AI subscriptions paid personally and claimed back.
- Look at browser extensions installed on managed machines.
- Scan code repositories for calls to model APIs and for API keys.
Treat what you find as an inventory, not a list of offenders. Each tool goes through the same question: approve it, replace it with an approved alternative, or stop it.
Bringing it into the light
Banning AI outright mostly pushes it further into the shadows. What works better:
- give people a good approved tool, so they don’t need their own;
- publish a short AI acceptable use policy and an approved tools list;
- make approving a new tool fast, days not months, for the low-risk cases;
- add every AI tool and every model you find to one inventory with an owner;
- check vendor AI features at contract renewal, as part of vendor risk assessment.
At a bank, the inventory step is not optional. OSFI’s E-23 asks institutions to identify all models in use, AI included, and a scoring script on someone’s laptop counts. MINERVA keeps that inventory and links each AI tool to its owner and supplier. See the model inventory.
Related questions
What is the difference between shadow AI and shadow IT?
Shadow IT is any technology used without approval. Shadow AI is the AI part of it, with extra risks: data typed into prompts, outputs used in decisions, and models that change without notice.
Is shadow AI illegal?
Not in itself, but it can lead to breaches of privacy law, confidentiality duties or regulatory expectations, for example when personal information is sent to an unassessed provider.
How do you detect shadow AI?
Combine a no-blame survey of teams with network and identity logs, vendor release notes, expense claims, browser extension lists and code scans for model API calls.
Should companies ban AI tools to stop shadow AI?
Outright bans tend to push use further out of sight. Providing approved tools, a clear policy and a quick approval route usually works better.
Does shadow AI matter for OSFI E-23?
Yes. E-23 expects institutions to identify all models in use, including AI built or used outside formal projects, and to manage them by risk.
Sources
- OSFI — Guideline E-23, Model Risk Management (2027)
- OSFI and FCAC — Risk Report: AI Uses and Risks at Federally Regulated Financial Institutions
- Office of the Privacy Commissioner of Canada — Principles for generative AI
New answers are posted on MINERVA on LinkedIn. Follow along, or send the question you’d like answered next to minerva@minerva.studio.
A plain-language explanation, not legal advice. The regulator’s own text is the source; read it, and take advice on how it applies to your institution.