Blog · Shadow AI · Controls
The shadow AI toolkit: tools and strategies to find, reduce and govern unapproved AI
16 September 2026 · 4 min read · By LeadAI
Once you know what shadow AI is and why it matters, the next question is practical: what do we do about it?
Start with an honest goal. You cannot fully eliminate shadow AI, and trying to will push it further out of sight. A better aim is to find it, shrink it, and move the useful parts into governed channels. That takes a mix of tools and policy, and neither works alone.
The types of tool, and what each one does
| Type of tool | What it does | Main limit |
|---|---|---|
| Identity and app discovery | Finds AI apps from sign-ins, app permission grants and software inventories. | Misses personal accounts used on personal devices. |
| Network and cloud access controls | Sees traffic to AI services and can allow, warn or block by tool or category. | Weak off the corporate network and on unmanaged devices. |
| Data loss prevention | Detects sensitive data heading into AI tools and warns or blocks. | Needs good data classification, or it produces noise. |
| Browser-level controls | Sees prompts, pastes and uploads in the browser, including personal accounts. | Covers the browser only, not desktop apps or direct API calls. |
| Device and extension management | Controls which apps and browser extensions run on company devices. | Only covers managed devices. |
| AI gateways | Routes model API calls through one point with logging, limits and data filtering. | Only covers traffic you choose to route through it. |
| Spend and procurement checks | Spots AI subscriptions in expenses, cards and purchase requests. | Misses free tools. |
| AI governance records | Holds the inventory, owners, risk reviews, decisions and evidence. | Records and decides, but does not block anything by itself. |
| Approved enterprise AI | Gives staff a sanctioned alternative with proper data protection terms. | Only works if it is good enough that people prefer it. |
No single tool covers everything. Think in three roles: monitoring tools find and block, governance records decide and prove, and approved tools remove the reason to go around the rules. You need all three.
Nine strategies that work
- Offer a better route first. Give people an approved tool that is at least as good as the personal one. This does more than any block.
- Make requests fast. Set a target for answering a new tool request, such as five working days, and publish the approved list so people can check before they ask.
- Run a no-blame amnesty. A 30-day "tell us what you use" window gives you real data. Punishing disclosure hides it.
- Use three data tiers. Public, internal and restricted, with a one-page rule on what may go where. Block the riskiest case, restricted data in unapproved tools, rather than everything.
- Check the less obvious channels. Look for AI features switched on inside software you already pay for, browser extensions, personal API keys and self-built agents.
- Tier your response. For each finding, choose one of four outcomes: approve, approve with conditions, replace or stop. Do not treat every tool the same.
- Train, and explain why. People follow rules they understand. Short, role-based sessions work better than a long policy document.
- Review on a cycle. Compare the declared and discovered lists every quarter, and track unresolved findings and time to decision.
- Fix the contracts. Make sure approved providers offer data processing terms, no use of your data for training unless you agree, and clear retention and deletion.
A sensible order
Weeks 1 to 4: see and set the basics.
- Run the amnesty and pull sign-in, app permission and expense data.
- Publish a first approved list and the one-page data rules.
- Start an inventory (see our guide to building an AI inventory).
Months 2 to 3: reduce the biggest risks.
- Add data loss prevention or browser controls for restricted data.
- Roll out an approved enterprise tool and a simple request process.
- Triage the findings and record each decision and owner.
Month 4 onwards: make it routine.
- Route model API use through a single controlled point.
- Run the quarterly comparison of declared and discovered AI.
- Feed the results into your AI management system and management review.
Match the effort to your size
A small organisation can get most of the benefit without a full monitoring stack: sign-in logs, an expense check, one approved tool and clear rules. Larger organisations with sensitive data or regulated work usually need the access, data and browser controls as well, plus governance records that tie everything together.
Do not forget autonomous tools
Agents and automations that act on their own need more than a block list. Each one needs an owner, limited access, its own credentials and a record of what it does. Our note on giving every AI agent an ID badge explains how.
Where ISO/IEC 42001 fits
ISO/IEC 42001 turns these steps into a repeatable system: a policy, named roles, an inventory, rules for the use of AI, supplier management, training, monitoring and regular review. Tools find the problem. The management system makes sure the fix lasts.
To see where you stand today, take the free AI Readiness Check, or explore how the LeadAI Assure journey turns discovery into an inventory and evidence.
This article gives general information and is not a substitute for ISO/IEC 42001:2023 or professional advice.