Shadow AI audit: a 14-day plan for growing companies

Shadow AI is the use of AI tools outside an organisation's approved systems and processes. It can be as simple as an employee pasting meeting notes into a personal chatbot, installing an AI browser extension or connecting an assistant to a company account without a security review.

The aim of an audit is to understand that use, decide what is safe and give people a workable route to approved tools. It should not become a hunt for employees to blame. The UK's National Cyber Security Centre describes shadow AI as a form of shadow IT and recommends reducing the risk through a positive security culture and secure AI alternatives.

This 14-day plan is designed for a company with roughly 50 to 500 employees. A smaller company can run it across the whole business. A larger one can start with a department that handles sensitive data or already uses AI heavily.

Decide what the audit must answer

Before day one, appoint one owner from IT, security or operations and one business sponsor who can remove obstacles. Write down the questions the audit needs to answer:

• Which AI tools and features are people using for work?
• What task does each tool help them complete?
• What company, customer or personal data enters the tool?
• Does the tool connect to email, files, calendars, code or other systems?
• Which uses can be approved, which need controls and which must stop?
• What approved alternative will employees use instead?

Include standalone chatbots, meeting assistants, browser extensions, writing and translation tools, coding assistants and AI features added to existing software. The inventory is about use, not branding. One product may contain several AI functions with different access and data flows.

Tell employees what you are reviewing, why it matters and how the findings will be used. If people expect punishment, they are less likely to tell you about the workarounds the audit needs to find.

Days 1–3: discover how AI is actually used

Start with a short survey and targeted conversations. Ask people which task they were trying to complete before asking which tool they chose. That question reveals the business need behind the workaround.

For every use, record:

• tool and feature name
• team and business task
• account type, such as company or personal
• data types entered, uploaded or generated
• connections to company systems
• frequency and business importance
• a person who can explain or own the use

Support the survey with information your organisation already has a legitimate reason to review, such as approved-app records, procurement data and identity or cloud administration logs. Agree the scope with the relevant privacy or legal owner. Avoid invasive monitoring and do not assume that a detected domain proves someone shared sensitive data.

By the end of day three, you need a useful first inventory, not a perfect list. Mark uncertain entries for follow-up instead of filling gaps with guesses.

Days 4–6: assess the use, not just the tool

Review each entry against the consequence of getting it wrong. A public brainstorming prompt and an assistant connected to a shared mailbox should not receive the same decision.

Ask five questions:

1. What data is involved? Separate public information from internal, confidential, personal and regulated data.
2. Who controls the account? Check ownership, access removal, multifactor authentication and whether the company can recover the account.
3. What happens to the data? Review retention, model-training settings, sharing and deletion terms instead of assuming every AI service works the same way.
4. What can the tool reach or do? A connector that can read files, send messages or take actions creates more exposure than an isolated prompt window.
5. Who checks the output? Identify where an inaccurate answer, invented source or insecure code could reach a customer or business process.

Use a simple decision such as approve, approve with conditions, investigate or stop. Record the reason and owner beside it. This is internal decision support, not a security certification of the product.

Days 7–10: replace risky workarounds with a safe path

A ban without an alternative leaves the original business need in place. For every use marked investigate or stop, decide how the employee should complete the task instead.

That may mean:

• enabling an approved feature with safer settings
• providing a company-managed account
• removing a connection that grants unnecessary access
• defining which data may and may not enter the tool
• adding human review before an output is published or acted on
• choosing a different tool for the task

Publish a short approved-use guide. Name the allowed tools and tasks, the data that is off limits, who can approve an exception and where employees should ask about a new service. Examples are easier to apply than a policy that only says staff must use AI responsibly.

Use the findings to shape security awareness training. If several teams uploaded meeting transcripts, train around that real decision. Generic AI warnings will be less useful.

Days 11–14: close the loop

Turn the inventory into assigned work. Each unresolved item needs an owner, a next action and a date. High-impact connections and sensitive-data use should come before minor convenience tools.

Use the final four days to:

• remove or restrict clearly unacceptable access
• move agreed work to managed accounts
• document approved uses and exceptions
• brief managers and employees on the new route
• assign technical, contractual or privacy follow-up
• set a date to review the inventory again

Do not report “shadow AI eliminated” as the result. New tools and features appear continuously. Report what you found, what you decided, which risks remain and how employees can now get an AI use reviewed.

The four outputs to keep

At the end of the audit, retain four practical records:

1. An AI-use inventory with owners, data types and connections.
2. A decision register showing what is approved, conditional, under review or stopped, with the reason.
3. An employee guide to approved tools, prohibited data and the request process.
4. A remediation list with owners and due dates for remaining work.

Review them when a major AI feature is introduced, after an incident or on a regular schedule that fits the pace of adoption in your company. The first audit does not need to solve every AI governance question. It needs to replace unknown use with visible decisions and give employees a safer way to do the work that led them to AI in the first place.