Shadow AI: what your team is already using and how to surface it
Banning AI doesn't stop AI. It just pushes it onto personal devices and personal accounts where you have zero visibility.
On this page▾
- If your policy is 'no AI', 60-80% of knowledge workers are still using it — on phones, on home laptops, on personal accounts.
- Amnesty + enterprise alternative beats enforcement. Always.
- The risk isn't 'AI exists', it's 'we can't tell what data left the building'.
- Measure success by drop in personal-account usage, not by 'caught' incidents.
Shadow AI is the natural consequence of a fast tool plus a slow procurement process. Engineers won't wait three months for a Copilot license — they'll paste code into a personal account today. Marketers won't wait either. Your job isn't to catch them; it's to make the sanctioned path faster than the shadow one.
Why shadow AI happens
- The official tool is slower, worse, or not yet approved.
- Approval requires a security review that takes longer than the project.
- Nobody told the user there was an approved alternative.
- The user knows but doesn't trust the approved tool's logging.
The 30-day amnesty program
- 1Week 1Announce: 'For the next 30 days, declare what AI you use — no consequences. After day 30, undeclared use is treated as a policy violation.'
- 2Week 2All-hands + manager 1:1s. Each team lists tools, use cases, what data they put in. Stand up a Slack channel for questions.
- 3Week 3Procurement fast-track for the top 5 tools surfaced. SOC 2 + DPA review in 5 working days, not 30.
- 4Week 4Publish the approved list, the policy, and the 'how to switch' guide. Run a clinic to migrate accounts.
Provide the enterprise alternative
Every shadow tool that survives the amnesty should have a sanctioned alternative within 60 days. If you can't fund alternatives, you don't have an AI policy — you have an AI ban, and bans don't work.
What to monitor (and what not to)
- Outbound traffic to known AI domains from corp devices
- DLP rules for source code + customer data
- Enterprise account usage trends
- Quarterly self-attest in the AI register
- Keystroke loggers
- Slack/email content scans for 'prompt' or 'ChatGPT'
- Personal device traffic
- Individual prompt history without cause
Read next
All playbooksA workable AI policy isn't a ban and isn't a free-for-all. Here's the four-tier model — green, yellow, red, prohibited — with the language, examples, and…
Vendors will sell you 'ethical AI'. The accountability stays with you. Here are the five HR-specific decisions — sourcing, screening, scoring, surveillance…
An honest map of where AI is delivering value in HR today, where vendors are overselling, the risk and regulation landscape, and a 30/60/90 pilot framework.