November 18, 2026

Shadow AI: Your Staff Already Paste Client Data Into ChatGPT

Your staff already paste client data into public chatbots. Why bans fail, and a 5-step plan to bring shadow AI into the light.

  • shadow AI
  • generative AI
  • data security
A dark card with the eyebrow SHADOW AI, the headline "Banning ChatGPT doesn't remove ChatGPT" and the subtitle "It only removes what you can see".

In short: Shadow AI is the use of generative AI tools at work without the organization’s approval, often through personal accounts. It is already widespread: according to KPMG and the University of Melbourne, almost half of employees admit using AI against company policy, including uploading sensitive information to public tools. Banning it fails; governing it and offering something better works.

This morning, someone on your team probably pasted a client email into a free chatbot to “make it sound more polite.” Nobody saw it. No log recorded it. And the tool kept the text.

That is not malice. It is someone trying to do their job well, faster. Which is exactly why the problem is so hard to fix.

Sceptra card: "A ban doesn't stop the usage", with a "real usage" window still active behind a "banned tool" window crossed out in red.

What is shadow AI?

Shadow AI is the use, at work, of artificial intelligence tools the organization has not chosen, approved or governed. The term comes from “shadow IT,” the software people install outside IT’s control. The difference: a hidden spreadsheet stores your data, while a public AI tool reads it, processes it and, depending on its terms, may retain it.

In practice that means a personal ChatGPT account used on a client file, a browser extension that “summarizes” your inbox, or a transcription bot quietly joining a confidential meeting.

We touched on this in our launch piece on AI and data confidentiality. Here we go further: measure the problem, understand why bans fail, and lay out a plan.

How common is shadow AI?

Very, and the numbers point the same way regardless of method. Here is the most solid data, taken from the original publications:

Source Method What it found
KPMG and University of Melbourne, 2025 Survey, 48,000+ people, 47 countries Almost half of employees admit using AI against company policy; 57% hide their AI use; only 40% say their workplace has a generative AI policy
Microsoft and LinkedIn, 2024 Work Trend Index Survey, 31,000 people, 31 countries 78% of AI users bring their own AI tools to work
Menlo Security, August 2025 Browser telemetry 68% of employees use free-tier AI tools via personal accounts; 57% of them enter sensitive data
IBM, Cost of a Data Breach 2025 600 breached organizations One in five reported a breach due to shadow AI; high shadow AI use added $670,000 to average breach costs

A note of honesty: these studies use different methods and scopes. Yet they all say the same thing. The usage is there, it is mostly invisible, and the rules arrive late.

Why doesn’t banning ChatGPT at work work?

Because a ban removes the visible tool, not the need. The need stays, and the usage moves to a personal phone, after hours, outside any control.

The best-known example is Samsung. In 2023, after engineers pasted internal source code into ChatGPT, the company banned generative AI tools on its devices and networks. An understandable reaction. But a ban like that does not cover a personal smartphone, or a screenshot sent to an AI app from home.

Three reasons bans fail:

  1. The time saved is real. When a tool saves an hour a day, the rule against it feels like an obstacle, not protection.
  2. Bans make usage invisible. People don’t stop; they stop talking about it. You lose the one thing that matters: knowing what is happening.
  3. There is no alternative. Saying “no” without offering an equivalent tool asks people to do worse work.

That last point echoes what we saw on employee resistance to generative AI: adoption follows usefulness, not the handbook.

What are the real risks of shadow AI?

The core risk is confidential data leaving for a third party you never chose, with no contract and no trace. Everything else follows.

  • Confidentiality and privilege. A client document pasted into a public tool is information handed to a third party. For lawyers, accountants and clinicians it becomes a legal question; see our piece on AI and professional secrecy.
  • Privacy law. France’s data protection authority, the CNIL, advises never sharing confidential information such as personal data with consumer AI services.
  • No audit trail. After an incident, you cannot say who sent what, when, or to which service.
  • Quality. An unchecked answer lands in a client deliverable. Nobody knows it came from an AI.

Simulate your situation in your organization

A few questions, and you leave with practices you can apply today.

Run the simulation

How do you govern shadow AI without banning it?

By making the official path easier than the workaround. Here is the method we recommend, in five steps.

The Map–Rule–Replace–Train–Trace plan

  1. Map. Inventory real usage: which tools, for which tasks, with which data. Ask without threat (“tell us what helps you”), then add network logs. You are looking for needs, not culprits.
  2. Rule. Write a short policy. The CNIL recommends clearly defining permitted and prohibited uses. Sort data into three levels: public, internal, confidential. One page is enough; nobody reads twenty.
  3. Replace. Offer a sanctioned tool as good as the public one, ideally better because it knows your documents. Most organizations skip this step, and it is the only one that durably reduces shadow AI.
  4. Train. Teach what AI does badly: errors, invented sources, bias. In the EU this is also a legal duty: Article 4 of the AI Act has required “AI literacy” measures for staff since 2 February 2025 (see our GDPR and EU AI Act 2026 checklist).
  5. Trace. Log what leaves: which data, to which model, when. The goal is not to watch people, but to be able to answer a client, an auditor or a regulator.

To test your plan, ask one question: will a rushed employee at 6 p.m. pick the official tool or their personal account? If the honest answer is “personal account,” step 3 isn’t done.

What should replace public chatbots?

A sanctioned tool only gets adopted if it is at least as useful as the one it replaces, and it must also keep your data with you. Three concrete criteria:

  • It works on your documents. The decisive edge over a public chatbot is context: your templates, files and procedures. We explained why in context is the missing 90%.
  • You choose where the model runs. A local model for sensitive work, a powerful cloud model for the rest (we compare options in local LLM vs cloud). The CNIL itself considers on-premise deployment generally more secure when personal or sensitive data is involved.
  • Outbound traffic is visible. Anything sent outside goes through a known connection and leaves a record.

That is exactly the architecture we built Sceptra on. The workspace runs on your own infrastructure, with the model of your choice. Work is organized in spaces per client or team: the AI working on Client A’s file sees Client A and nothing else. Data leaves only through connections your administrator explicitly creates, and everything that leaves is logged. Every answer cites its sources. And if you leave, you take your data with you: everything is reversible.

Put differently, steps 3 and 5 of the plan become one and the same: the tool people prefer is also the one that leaves a trail.

Frequently asked questions

What is the difference between shadow IT and shadow AI? Shadow IT is any software used outside IT’s control. Shadow AI is a riskier subset: the tool doesn’t just store your data, it reads it, processes it and may retain it under its terms. A single copy-paste is enough for a confidential document to leave.

Does ChatGPT Enterprise solve shadow AI? Partly. A contracted enterprise plan changes how your data may be used and adds an admin console. But your data is still processed by a third-party provider, and the tool only reduces shadow AI if people actually adopt it. Check the contract, data location and logging.

Should we discipline employees who use unapproved AI? Not first. Punishing without offering an alternative pushes usage underground and blinds you. Start with an inventory and a sanctioned tool. Discipline has its place later, for deliberate breaches of a policy that is clear, known and workable.

How can I detect shadow AI in my organization? Combine three sources: an anonymous usage survey, network and proxy logs to AI domains, and an audit of installed browser extensions. The survey is often the richest: it reveals needs, not just tools.

Is shadow AI illegal? Not in itself. But it can lead to breaches: violated client confidentiality, personal data processed without a legal basis or contract, unmanaged transfers abroad. The organization, as the party responsible for the data, is the one held accountable.

Key takeaways

  • Shadow AI is the majority case and mostly invisible: almost half of employees admit bypassing the rules.
  • Bans remove the visible tool, not the need; usage moves to personal phones.
  • The plan: Map, Rule, Replace, Train, Trace.
  • The decisive step is offering a sanctioned tool as good as the public one.
  • Anything that leaves should go through known connections and be logged.

Want to see what a sanctioned alternative would look like in your organization, on your documents and your infrastructure? Book a 30-minute call with our team, or try the domain-by-domain simulator at sceptra.ai.