Shadow AI: your team is already vibecoding without telling you
Someone in your company has already built an internal app with AI. No ticket, no budget, no heads-up to IT — if you even have an IT team. A sales dashboard wired to a CRM export. A bonus calculator thrown together on a Sunday night. An HR form built on Lovable in two hours, with real employee names inside.
This is shadow IT in the AI era: nobody needs to know how to code anymore to build a tool. They just need to describe it. And the right response is not a ban. It's four decisions:
- Inventory without punishing: run a declaration amnesty so you find out what actually exists.
- Define three zones: a free sandbox with fake data, a vetted zone for internal data, and an off-limits zone for customer, payroll, and health data.
- Require two things per app: a code export and a named owner.
- Review everything that touches personal data. No exceptions.
The rest of this article covers why this is a useful signal rather than a betrayal, and how to put the framework in place starting this week.
What's actually happening on your team
Conversational tools like ChatGPT and Lovable changed who gets to make software. Your marketing lead can describe a lead-scoring tool and have it running by the afternoon. Your finance controller can get a script generated that merges three Excel exports.
These apps stay invisible for three simple reasons:
- They ask nothing of anyone: no server, no license purchase, no ticket. The platform hosts everything.
- They often live on personal accounts, created outside any company contract.
- They work. Nobody reports a tool that saves them two hours a week.
The sensitive part isn't that these apps exist. It's that they often run on real company data — customer files, salaries, sales history — pasted into tools nobody has evaluated.
It's a signal, not a betrayal
When someone builds their own workaround, it means a need went unserved: the official tool doesn't exist, or the request has been sitting in a backlog for six months. Shadow IT is the most honest map you'll ever get of what your teams actually need.
Responding with a blanket ban produces the opposite of what you want. The apps don't disappear — they migrate to personal accounts, personal phones, private chats. You lose visibility, and the people involved lose the option of asking for help when something goes wrong.
The energy your team puts into building its own tools is an asset. The job is to channel it, not to shut it off.
The framework: four simple rules
1. The declaration amnesty
Announce a 30-day window: everyone declares the tools and apps they've built or use, zero sanctions, no matter what surfaces. The form fits in five questions:
- What does the app do?
- What data does it touch?
- Who uses it?
- Where is it hosted?
- Who built it?
The goal is a map, not a tribunal. If declaring is risky for the person declaring, you'll only ever see the tip of the iceberg.
2. Three zones everyone knows
| Zone | Data | Rule |
|---|---|---|
| Sandbox | Fake or public | Total freedom, no review |
| Vetted | Internal, non-sensitive | Declaration plus a quick review |
| Off-limits | Customers, payroll, health | Company-approved tools only |
The sandbox matters more than it looks. It's what makes the whole framework credible: people get a place to experiment freely, so the temptation to route around the other two zones drops.
3. One code export and one owner per app
Two requirements, no more:
- A code export. If the platform can't hand over the code, the app is a demo, not a work tool. The day it changes its pricing or shuts down, you need to be able to take over.
- A named owner. An app without an owner goes orphan the first time someone leaves. A name in a spreadsheet is enough.
4. Personal data always goes through review
One absolute rule, easy to remember: the moment an app touches personal data — customers, employees, applicants — it gets reviewed before anyone relies on it. You already carry obligations on that data; this just makes them operational for AI-built apps.
The EU AI Act makes this concrete
The EU AI Act, which has been coming into force in stages, classifies certain AI systems deployed in finance, healthcare, or critical infrastructure as "high-risk." For those systems, it requires audits and design documentation, among other obligations.
In practice: a clandestine business app in one of those areas can tip your company into non-compliance. Not because the app is badly built — but because nobody can document how it was designed, on what data, or who answers for it.
Here's the good news: the framework above produces exactly the paperwork the regulation asks for. The inventory says what exists. The named owner says who answers for it. The code export and the review document the design. You're not doing compliance on top of the framework — the framework is the compliance.
Where to start this week
- Day 1: announce the amnesty. One simple message: "We know AI tools are circulating. We want to know about them, not punish anyone. You have 30 days."
- Day 2: publish the three zones on one internal page. One page — not a 40-page PDF.
- Day 3: set up the declaration form (five questions, five minutes to fill in).
- Weeks 2-3: collect declarations. Sort each app into its zone. Publicly thank the first people who come forward.
- Week 4: for every app in the vetted or off-limits zone, assign an owner, request the code export, and schedule a review if personal data is involved.
- After that: keep the inventory alive. A one-hour quarterly review is enough for a small company.
Shadow IT in the AI era isn't a crisis to smother. It's your team showing you, tool by tool, what it needs. Give it a frame, and a blind spot becomes an advantage.