Shadow AI is staff using AI tools the company has not approved, such as pasting data into public chatbots or connecting browser extensions to email. Bans rarely work, because people keep using what saves them time. The safer response is to find what is in use, sort it by data risk and offer an approved route that is easier.
Why shadow AI happens
People reach for unapproved tools because they have work to finish and the approved option is missing, slow to request or weaker than what they can open in a browser tab. In most cases nobody is trying to break a rule. A sales manager summarising call notes or a developer asking a chatbot to explain an error is solving a real problem.
That matters for how you respond. If the policy treats every use as misconduct, people stop mentioning it and the company loses sight of where its data goes. If the policy treats it as a signal of unmet need, it tells you which approved tools to provide first.
What can go wrong
The risk is not the tool itself but what is sent to it and what it is allowed to do.
- Data leaving your control. Customer records, contracts, source code or personal data pasted into a consumer service may be stored, reviewed or used for training under terms the company never agreed to.
- Tools that act, not just answer. Browser extensions, meeting bots and agent-style tools can read inboxes, calendars and shared drives, and some can send messages or change records.
- Unchecked output in real work. Drafted contract clauses, code or customer replies can carry errors into production with no record of where they came from.
- Regulatory exposure. Personal data sent to an unapproved processor, or outside the regions you have committed to, can breach data protection duties and customer contracts. Our post on data residency for AI agents covers the location side.
Find out what is already in use
Start with evidence, not a survey alone. Network and identity logs show which AI services staff sign in to. Browser management shows installed extensions. Expense claims show paid subscriptions. OAuth consent records show which apps have been granted access to mail, files and calendars, which is often the riskiest list.
Then ask teams directly, with a clear promise that the aim is to provide better tools rather than to punish anyone. The answers explain why each tool is used, which the logs cannot.
Record what you find in the same place you track internal agents. Our guide to keeping track of every AI agent describes a registry that works for both.
Sort tools by data risk, not by name
A list of banned brand names goes out of date quickly and says nothing about what is allowed. Tie the rules to your existing data classification instead.
Low risk
Public information and general drafting, such as rewording a public web page or brainstorming headings. Allow it with guidance, and keep watching.
Medium risk
Internal but not sensitive material, such as meeting notes without customer details. Allow it only in tools with business terms that exclude training on your data and let you control retention.
High risk
Customer data, personal data, source code, financial records, credentials, and any tool that can take actions in company systems. Allow it only in approved, configured tools with logging, access control and a named owner. The approach in stopping agents reaching restricted data applies here too.
Make the approved route the easy one
People follow the path of least effort. Give them an approved assistant with sign-in through the company identity provider, business data terms, retention you control and logging that security can review. Publish it in the places people already look, and make requests for new tools quick to raise and quick to answer.
Where staff need AI inside a specific workflow, such as drafting support replies from the ticket history, build it into that system with proper permissions. That removes the reason to copy data into a separate tab. This is the kind of work our enterprise AI product engineering team does.
Write a policy people can follow
Keep the policy short and practical. It should say which tools are approved for which kinds of data, what must never be pasted anywhere, how to request a new tool, who to tell if something sensitive was shared by mistake, and that output used in real work must be checked by a person. Give examples from real roles, not abstract categories.
Review it whenever the approved tool list changes, and tell staff what changed and why.
Keep watching after the rollout
Shadow AI does not end once an approved tool exists. New services appear, extensions update their permissions and teams find new uses. Keep the discovery checks running, review OAuth grants on a schedule, and treat a spike in use of an unapproved service as a prompt to ask what need it is meeting.
If you want help putting an approved route in place or connecting AI safely to your systems, talk to us.
Frequently asked questions
What is shadow AI?
Shadow AI is the use of AI tools at work without approval from IT, security or data owners. Common examples include pasting documents into public chatbots, installing AI browser extensions, adding meeting bots to calls and connecting AI apps to email or shared drives. It usually comes from people trying to work faster, not from bad intent, which is why the response should focus on safer options as well as controls.
Should we ban public AI chatbots at work?
A blanket ban tends to move use onto personal devices and accounts, where the company has no visibility at all. It works better to block the clearly unsafe cases, such as sensitive data in consumer tools, while providing an approved assistant with business data terms and company sign-in. Pair that with a short policy that explains what is allowed, so people are not left guessing.
How do we find out which AI tools staff are using?
Combine technical evidence with conversations. Identity and network logs show sign-ins to AI services, browser management lists installed extensions, expense claims reveal paid subscriptions and OAuth consent records show which apps can reach mail, files and calendars. Then ask teams what they use and why, with a clear message that the goal is to provide better tools, so the answers are honest and complete.
What should staff do if they shared sensitive data by mistake?
They should report it straight away to a named contact, without fear of blame, so the company can act quickly. That contact checks what was shared and with which service, deletes the content or conversation where the service allows it, revokes any access the tool was granted and decides whether data protection or contractual notifications apply. Quick, blame-free reporting matters more than a perfect record.



