Shadow AI: govern the tools your employees build with ChatGPT or Claude (2026)
Your employees use ChatGPT or Claude, often with their own account, and some already turn them into real tools: a calculator, a tracker, a small client portal. Banning them pushes them further into the shadows, and ignoring them leaves your data anywhere. Here is how to govern them while keeping what works.
By Aloïs Oddon, co-founder of Nextmap. Facts checked on November 18, 2026.
Shadow AI is the use of AI tools outside the company’s view: an employee pasting a contract into their personal ChatGPT account, a salesperson having Claude proofread their quotes. The numbers show the scale of it. In 2024, Microsoft and LinkedIn’s Work Trend Index found that 75% of knowledge workers used AI at work, and that 78% of them brought their own tools, 80% at small and medium-sized companies1.
Since Claude, ChatGPT and Cursor started building tools, shadow AI also produces real tools, which colleagues, and sometimes clients, use every day. A quote calculator shared through a public link, a stock tracker on an employee’s personal subscription, a small client portal made over a weekend. These tools are useful, and that is exactly why they deserve attention.
The short version. Banning AI pushes its use into the shadows. Choose the allowed AIs instead, and give a home to the tools they build. On while, each tool is kept for employees through the company account, its credentials and data stay in France, and the history shows who did what.
What shadow AI builds in a small business
AI-built tools look nothing like the software IT is used to tracking. They are born in a conversation, in an hour, and shared in one click. Here are the ones you often come across, and what they leave open.
- A Claude artifact shared through a public link. Anyone with the link can use it, and for most of them, there is no need to even be signed in to Claude2.
- A tool built with Lovable. By default, anyone with the link to a published tool can use it, and tools are served from around the world5.
- A tool on an employee’s personal account. It lives on their subscription, with their settings, and it leaves with them the day they leave the company.
- A tool that reads company software. For it to read the invoicing system or the CRM, its credentials were often pasted into the code or the conversation.
Why a ban is not enough
The obvious answer is to ban everything, and it holds up poorly. People who got used to AI carry on with their own account, and an IT team that turns down every request becomes the obstacle people work around. The risk lies mostly in what nobody sees: tools with no record of who uses them, data hosted who knows where, credentials copied by hand.
The other answer, letting it happen, comes to the same thing. The right one is to say yes to the teams who build, while giving them a framework they want to follow.
Govern shadow AI with five rules
These rules fit on one page, and they apply whatever AI your teams prefer.
- Take inventory, without blame. Ask each team which tools it has built with AI, and who uses them. An inventory without blame will teach you far more than an investigation.
- Choose the allowed AIs. The team plans of Claude, ChatGPT or Cursor give admins control, as the table below shows.
- Give the tools a shared home. One place, known to all, where AI-built tools live behind the company sign-in.
- Get credentials out of the code. The credentials of company software should live neither in the code nor in the conversation.
- Decide who approves, and keep a record. A tool kept for your colleagues can be shared freely. Sharing it with a client or with everyone goes through a designated person, and every action stays on record.
You, to your AI: Put this quote calculator on while, for the sales team only.
What the AIs control themselves
Each AI offers settings for its team plans. They matter, but they only apply to the tools built with that AI.
| AI | What the admin controls |
|---|---|
| Claude, Team and Enterprise plans | An artifact’s public link stays off until an owner allows external sharing2 |
| ChatGPT Sites, Enterprise workspace | Public publishing of a Site stays off until an admin turns it on3 |
| Cursor Apps, Teams and Enterprise plans | Public is off by default, and admins see every tool4 |
| Lovable, Business and Enterprise plans | Publishing can be limited to workspace members5 |
A small business that lets its teams pick their AI quickly ends up with four admin consoles, and as many lists of tools. A shared home for every AI avoids that sprawl.
What while brings to IT
We built while so IT can say yes to teams without losing control. Your employees keep the AI they prefer, Claude, ChatGPT, Cursor or another, and it puts their tools on while. You find them all in one place, whatever AI built them.

- The company sign-in in front of every tool. On the Pro plan, everyone signs in with the company account: Microsoft Entra, Google Workspace, Okta or any OIDC provider. See keep a tool for your colleagues.
- A person approves what matters. Opening a tool to everyone, or letting it reach a service on the Internet, goes through an approval, and the AI never approves anything itself.
- Credentials stay in while. The AI never sees them, and every call is traced. Credentials left in the code are detected, and each tool’s dependencies are continuously checked against known vulnerabilities.
- The history shows who did what, and it is kept for 13 months.
- The data is backed up every day, and hosted in France by Nextmap.
A client or a supplier can also use a tool with an access link and a password, without an account, which avoids public links. The details of the protections are on our security page, and what while changes for IT is on the IT page.
Try it with one team
Start with a tool that already exists. Ask its author to give this message to their AI: the trial is free, with no sign-up, and the tool has its address in 3 minutes.
Read while.sh/agents and help me get started with while.
In the docs: Dependency vulnerabilities.
Sources
- Microsoft and LinkedIn, Work Trend Index 2024: AI at work is here. Now comes the hard part, accessed November 18, 2026
- Anthropic, Share artifacts, accessed November 18, 2026
- OpenAI, Sites: Publish and share, accessed November 18, 2026
- Cursor, Build Live Apps in Cursor, accessed November 18, 2026
- Lovable, How Lovable hosts your app, accessed November 18, 2026