Shadow AI: how it creeps into your company

The instinctive picture of shadow AI is a rogue developer building something clever under the radar. In practice, it is almost never that exciting. Shadow AI arrives as a series of small conveniences — none of them dramatic, each of them defensible — that add up to a parallel technology estate the company did not know it had.

Knowing the pattern is the first step to catching it before it becomes a liability.

The four entry points

The personal tab. Someone logs into a consumer AI tool with a personal email, on a personal device, and pastes work content into it. They are not "using" company AI; they are using "their own" AI to help with their job. The data is in flight all the same.

The browser extension. A small extension promises to summarise meetings, draft replies, or rewrite documents inside tools the company already approves — Gmail, Notion, Teams. The extension is the AI; the approved tool is the costume.

The "AI feature" inside an approved tool. The company has approved a SaaS product. The SaaS product adds an AI assistant in a new version, on by default. Nobody re-approves anything; the assistant is on the next day. This is the most common entry point we see now.

The script in someone's drawer. A keen analyst writes a small Python script that calls a model API to clean a spreadsheet. The script lives on their laptop, runs whenever they want, and accesses data via their own credentials. From a network perspective, it looks like the analyst working.

Why people do it

Every shadow AI story has the same backstory: "the official tool was not available, did not work for this case, or was going to take six weeks to get approved." The person is not trying to break a rule. They are trying to do their job. They reach for the nearest tool that solves the problem this morning.

Punishing the behaviour without changing the conditions that caused it is a guarantee that next month's shadow AI will be better hidden. The path forward is faster sanctioned options, lower-friction approval, and a serious feedback loop on what was missing in the official toolkit.

What it costs you

Three categories of cost, roughly in the order companies notice them:

Data exposure. The most concrete cost. Content that was not meant to leave is now in a third-party system you did not contract with. The damage is hard to quantify until something specific surfaces.

Duplicate spend. Departments quietly pay for overlapping tools, often on personal credit cards reimbursed through expenses. The same model is being paid for four times under different names.

Architectural drift. Workflows quietly come to depend on the shadow tool. Removing it later breaks things nobody documented. The longer it is in the dark, the more expensive the cleanup.

What surfaces it

You will not find shadow AI by asking "is anyone using unapproved AI?" — the answer is always "no, not really." You find it by looking at the right signals:

Expense reports. Search for the names of the obvious providers. Small, repeating charges to individual cards are usually shadow tools.

Browser extension audits. If your MDM can list installed extensions, an annual scan is informative. The signature is a long tail of AI-themed extensions installed by small numbers of users.

"Help me with this" requests to IT. When users ask for help integrating something the company has never deployed, you have found a shadow tool. Treat the request as a discovery, not a violation.

Release notes of approved SaaS. The most reliable source. Every quarter, scan the changelogs of the tools you have approved for new AI features that were not in the original review.

The amnesty trick

A short, well-communicated amnesty — "register any AI tool you are using by the end of the month, with no consequences" — surfaces more than any technical audit. People are generally relieved to bring it into the light when the cost of doing so is zero. We have seen single amnesty rounds register fifty to eighty initiatives in companies that thought they had four.

After the amnesty, set up the ongoing process — a quick registration form, a tiered review, a clear answer within days, not weeks. This is exactly what Targe is built to handle, end to end.

Bring it into the light

Shadow AI is not an integrity problem. It is a speed problem dressed up as a security problem. Companies that move shadow AI into sanctioned AI without slowing teams down end up safer and faster. Companies that try to crush it without offering alternatives end up with neither.

If you want help running the amnesty and standing up the ongoing process, we would be glad to walk you through it.