What is shadow AI?

By Dylan Bailey, Certified Claude Expert · Updated 4 August 2026 · 5 min read

Shadow AI isn't a technical problem and it isn't really a compliance problem. It's what happens when a tool is obviously useful, easy to reach, and officially unavailable. People don't stop needing the help; they stop mentioning it.

Why capable people do it

It's worth being clear that this is rarely defiance. The pattern looks like:

• A report is due, and a first draft would take ninety minutes
• A free tool produces a workable draft in five
• There's no approved alternative, and asking would take a day
• Nothing bad appears to happen, so it becomes routine

Each step is individually reasonable. That's exactly why prohibition performs so badly — you're asking someone to work an extra eighty-five minutes to comply with a rule that, from where they're standing, protects nothing visible.

What it actually costs you

• Data on the wrong tier. Free consumer accounts often reserve training rights. Client material may be going somewhere you'd never have agreed to.
• No record. If a client ever asks what was shared with a third party, you cannot answer.
• Work you don't own. Refined instructions and accumulated context live in a personal account. When that person leaves, so does all of it.
• No shared standard. Everyone invents their own approach to checking output — some rigorous, some not at all.
• Nothing compounds. Ten people quietly solving the same problem separately, none of them sharing what worked.

An Australian small-business example

A Canberra consultancy working on government contracts bans AI outright, for understandable reasons. No tool is provided and no policy is written beyond the ban.

Eighteen months later an informal conversation reveals that most of the delivery team uses AI daily — on personal accounts, on personal devices, for drafting and summarising client material.

The ban achieved the opposite of its intent. The firm now has the same usage it feared, with none of the controls it could have had, and no idea what has been shared or with which service.

What actually works

• Provide something. One approved business-tier tool eliminates most shadow use immediately, because the sanctioned option is now also the better one.
• Ask without penalty. Run an amnesty. What people are already doing is the best possible map of what you should support.
• Make the rules short. One page. Three buckets of data. A named person to ask.
• Train rather than warn. People who understand why the sensitive bucket exists apply it under pressure. People who only received a prohibition route around it.
• Make the approved path the easy path. If sanctioned use requires a form and two approvals, you've rebuilt the incentive to go around it.

The reframe

Shadow AI is usually described as a discipline problem. It's more useful to read it as feedback: staff have found something that materially helps, and the organisation hasn't caught up.

Treated that way, it's not a threat to shut down but a signal about where the demand is. The businesses that act on it end up with better controls and better adoption than the ones that keep tightening a rule nobody is following.