Problems & fixes · AI risk and governance
Shadow AI: staff using unapproved tools and what to do about it
Last updated 25 October 2026 · 2 sources listed · Reading time about 3 minutes
Short answer
Shadow AI arises when approved options are slow, limited or unknown. Prohibition alone seldom works. Combine discovery, a clear policy with examples, an approved alternative that meets real needs, a quick exception route, technical guardrails on sensitive data and training that explains reasons. Treat discovery as information about unmet needs rather than only a disciplinary matter.
Who this is for: A security, IT or compliance lead who suspects employees are using unapproved AI tools with company data.
What to evaluate
Discovery of actual use
You cannot manage what you cannot see.
How to check: Combine network and identity logs, expense data and anonymous staff surveys to understand tools and purposes.
Reasons behind it
Staff turn to unapproved tools to meet real needs.
How to check: Interview teams about tasks and obstacles with the approved option.
Policy clarity
Vague rules are ignored or misread.
How to check: Provide short guidance with examples of allowed and forbidden inputs.
Approved alternatives
Rules without alternatives drive concealment.
How to check: Offer sanctioned tools covering the main tasks and publish how to get access.
Technical guardrails
Controls catch accidents that training misses.
How to check: Consider data loss prevention rules, managed browsers and restrictions on uploading sensitive files.
Response proportionality
Harsh reactions to honest disclosure suppress reporting.
How to check: Define graded responses and a safe route to self-report.
Comparison
| Response | Strength | Limit |
|---|---|---|
| Block known tools at network level | Fast reduction of obvious exposure | Easy to bypass on personal devices; endless list |
| Provide approved tools | Addresses the cause | Needs funding, support and communication |
| Policy and training | Builds understanding | Effect fades without reinforcement |
| Data loss prevention rules | Catches certain sensitive patterns | False positives and incomplete coverage |
| Amnesty and disclosure route | Reveals real use | Needs trust and a consistent response |
Decision worksheet
Rate up to three options against this guide’s own criteria. It runs in your browser and saves only on this device. It is a thinking aid, not a recommendation.
1. Have you checked each criterion?
0 of 6 criteria checked.
2. Compare options (0 = fails or unknown, 3 = meets it well)
| Criterion | |||
|---|---|---|---|
| Discovery of actual use | |||
| Reasons behind it | |||
| Policy clarity | |||
| Approved alternatives | |||
| Technical guardrails | |||
| Response proportionality | |||
| Total (max 18) | 0 | 0 | 0 |
Questions to ask before you commit
- What evidence do we have of which tools staff use and for what?
- Does the approved tool cover the most frequent tasks?
- How quickly can a team get access or an exception?
- What examples of forbidden data have we given staff?
- Which technical controls protect sensitive data in unsanctioned tools?
- How do we treat self-reported incidents?
- What do contractors and temporary staff follow?
- How will we measure progress?
Typical trade-offs
Blocking tools versus enabling and monitoring
- Upside
- Blocking reduces exposure quickly.
- Downside
- Drives hidden use to personal devices.
Broad amnesty versus case-by-case discipline
- Upside
- Encourages disclosure and learning.
- Downside
- May seem lenient for serious breaches; define limits.
Purchasing wide licences versus targeted ones
- Upside
- Reduces demand for alternatives.
- Downside
- Cost and unused seats.
Common mistakes
- Relying on a ban with no alternative.
- Surveying staff in ways that feel like entrapment; use anonymous formats.
- Forgetting embedded AI features in sanctioned software that process data by default.
- Failing to explain why rules exist.
Related questions
How widespread is shadow AI likely to be?
Assume it exists wherever staff have internet access and useful tasks, whether or not you have found it. Anonymous surveys and log review give a picture for your organisation. Treat the result as input for your tool provision and training rather than as evidence of bad intent.
What is the real risk?
Confidential or personal data entered into services with unknown retention, training and access terms, loss of ability to respond to regulator or customer questions, and decisions made on unverified outputs. Severity depends on the data involved and the tool's terms, which is why classification of data and clear rules matter.
Can technical tools stop it?
They can reduce it, with web filtering, managed browsers and data loss prevention controls. They cannot cover personal devices, home networks or every channel through which text can be pasted. Combine them with approved alternatives that are genuinely usable, short practical education, and a culture in which staff feel able to ask before using a tool. Blocking alone usually pushes the behaviour out of sight.
What should we do after discovering a serious case?
Contain the exposure by removing shared content and revoking access where possible, assess what data was involved and whether legal notification duties arise in your jurisdiction, document the incident, and review controls. Handle the employee proportionately and learn from the cause.
Sources and further reading
These pages were reachable when we last checked. They informed definitions and rules referred to above; they do not endorse this guide. Where rules differ by country, check the authority for your location.
- 10 Steps to Cyber Security — UK National Cyber Security Centre, retrieved 27 October 2026
- AI Risk Management Framework — National Institute of Standards and Technology, retrieved 27 October 2026
Related guides
How to choose · AI risk and governance
How to assess an AI vendor during due diligence
Assess an AI vendor on data handling, security assurance, model and subprocessor transparency, contractual protections, testing and monitoring support, and operational resilience. Request evidence rather than assertions: independent attestations, data flow diagrams, retention terms and incident commitments. Tailor depth to the risk of the use case, and revisit when features or subprocessors change.
Comparison · AI risk and governance
Comparing AI governance frameworks: risk framework, management standard and regulation
These instruments do different jobs. A risk management framework offers guidance on identifying and managing risk; a management system standard describes the organisational processes that can be audited and certified; regulation creates legal duties with penalties and applies by jurisdiction. Most organisations use a framework to design practice, a standard to structure it and legal advice to meet obligations.
Requirements · AI risk and governance
Requirements for using AI in decisions about people
Tools that influence decisions about people face the strictest expectations. Typical requirements include a defined lawful purpose, documented impact assessment, testing for discriminatory effects, transparency to affected people, meaningful human oversight with authority to override, routes to contest outcomes, record keeping, and vendor evidence. Laws differ by jurisdiction, so obtain legal advice and consult regulator guidance before use.
Requirements · AI coding tools
What an engineering policy for AI coding tools should cover
A workable policy names approved tools and plans, forbids pasting secrets or customer data, sets repository eligibility, requires normal review and testing for all code regardless of origin, defines licence and provenance handling, covers agentic tool permissions, and assigns ownership for monitoring and updates. Keep it short enough that developers actually read it.
Requirements · Enterprise AI tools
What an enterprise AI deployment needs in place before launch
Minimum conditions include an approved-use policy, a register of tools and use cases with owners, risk tiering, data classification rules, vendor due diligence, access and logging controls, human review expectations for consequential outputs, training, and an incident process. Legal duties differ by jurisdiction and sector, so confirm applicable rules with counsel and regulators.
Problems & fixes · Professional services
Why website projects run over budget and behind schedule
Most overruns come from unclear scope, late content, too many decision makers and untested assumptions about integrations, not from slow developers. Prevent them by fixing scope in writing, assigning content owners, naming one approver, testing hard technical risks early and agreeing how changes are handled. Expect some change, and plan a contingency rather than treating it as failure.
Problems & fixes · Enterprise AI tools
Why enterprise AI rollouts stall and what to fix
Rollouts typically stall for organisational reasons: unclear use cases, weak permissions that make security block expansion, poor answer quality on internal content, absent training, and no measurement of value. Diagnose which of these applies, fix content and access foundations first, concentrate on a few workflows per function, and measure outcomes rather than logins.
All AI risk and governance guides · All Artificial intelligence