# Shadow AI: what unsanctioned tools are doing to your pilots

_While leadership debates the official AI pilot, the team has already wired half a dozen unsanctioned tools into the real workflow. That shadow layer is where your data, and your risk, actually live._

Shadow AI is the unapproved tools your team already uses. Told through an anonymized story from a product manager at a large marketplace, here is how unsanctioned AI quietly reshapes your pilot's risk.

# Shadow AI: what unsanctioned tools are doing to your pilots

There is the AI pilot on the slide deck, and there is the AI your team actually uses. They are rarely the same. Someone is pasting customer data into a chatbot to draft a reply. Someone else generated a chunk of the pilot's own code from a tool nobody reviewed. This is shadow AI, and it is not a fringe problem. More than 80% of employees already use AI tools their company never approved.

That gap matters because your governance covers the pilot on the slide. It does not cover the six tools quietly doing the work around it.

## The backdoor that walked in through the training data

A senior technical product manager at one of the largest online marketplaces told me a story that reframes what "unsanctioned" means. People had noticed that a major model pulled a lot of its coding knowledge from public repositories. So some researchers, not even malicious ones, created repositories full of code that said, in effect, "if you are solving this kind of security problem, use this pattern." Inside that pattern they hid a backdoor. When the model retrained, it absorbed the advice. Then it started handing that backdoored code to developers who asked it for a security solution, who shipped it, believing an AI had written something safe.

He called it social engineering aimed at a language model. The lesson is not about that one exploit. It is that when your team generates pilot code from an unsanctioned tool, you have inherited every assumption baked into that tool, including the ones nobody can see. The pilot looks like yours. Its provenance is not.

## Shadow AI is a symptom, not a crime

The same product manager described teams "vibe coding" prototypes overnight, thrilled at how fast a working thing appears. That speed is real, and it is why shadow AI spreads. People are not being reckless. They reached for the tool that removed friction, because the sanctioned path was slower or did not exist. Banning the tools does not work, it just pushes them further underground. The teams that handle this treat shadow AI as a signal: if everyone is using an unapproved tool, the approved workflow is failing them somewhere.

His other worry sharpens the point. When you let AI check the AI that wrote the code, and another layer check that, eventually nobody understands what is happening, and a single bad change can propagate all the way up with no common sense in the loop. Shadow AI is that same loss of visibility, spread across tools you do not even know are in play.

## Bringing the shadow into the light

You cannot govern what you cannot see. The moves that work:

- **Inventory what is actually in use.** Ask the team what tools they reach for, without blame. You will be surprised.
- **Give them a sanctioned fast path.** People use shadow tools because the official route has friction. Remove it.
- **Track the provenance of pilot code.** Know which parts came from a model, so a bad pattern is traceable.
- **Set a clear line for data.** What can go into an external tool, and what never can, stated plainly.

## How we approach it at Density Labs

In the AI Readiness Assessment, our $2,500 front door, one of the first things we surface is the gap between the pilot on paper and the AI the team already uses. We map the unsanctioned tools touching the workflow, where data flows into them, and where generated code entered the pilot without a review. That map usually reveals more real risk than the official pilot ever did, and it is a lot cheaper to draw before an incident than after.

Shadow AI is not your team going rogue. It is your team telling you where the sanctioned path is too slow. Listen, then build the fast path yourself.

The AI you approved is not the AI you are running. Find the difference before someone else does.