How do you audit shadow AI across a 100-person company?

What is the fastest way to audit shadow AI?

Start with identity-provider app grants, DNS or cloud-discovery logs, browser extensions, company-card transactions, and an anonymous survey. For a 100-person company, these records map useful risk in one week without reading employee prompts. Then classify tools, migrate valuable work to approved accounts, and revoke access only when the risk is clear.

Microsoft and LinkedIn's 2024 Work Trend Index found that 78% of AI users brought their own tools to work, based on 31,000 workers in 31 countries. Treat that as a reason to ask better questions.

Which records should I check first?

Begin with four exports. They overlap, which is useful because no single log sees the whole picture.

Which identity records reveal shadow AI?

1. Identity and OAuth grants. In Google Workspace, review Admin Console > Security > Access and data control > API controls > Manage Third-Party App Access. Google documents this control at support.google.com/a/answer/7281227. Look for AI meeting assistants, writing tools, coding assistants, and anything asking for Gmail, Drive, Calendar, or Slack access. In Microsoft Entra, review Enterprise Applications and audit logs, paying attention to user-consented applications and delegated permissions.

Which network records reveal personal AI accounts?

2. Network and DNS activity. Search your DNS, firewall, or cloud-discovery data for AI domains and categories. Microsoft describes this workflow in its Defender for Cloud Apps shadow IT guide. Cloudflare documents a similar Shadow IT SaaS analytics workflow. This catches someone using a personal Gmail account on a company laptop, where OAuth records may show nothing.

Which browser extensions should I inspect?

3. Browser extensions. Ask whoever manages Chrome, Edge, Jamf, or Intune for installed extensions and their permissions. A summarizer with <all_urls> access can read pages inside a CRM or HR system even when it never receives a formal Drive permission. Google's Chrome Enterprise extension documentation is the right starting point. Flag extensions that can read every site, inject scripts, or have no clear business owner.

Which finance records reveal paid AI tools?

4. Finance records. Ask Finance for 90 days of company-card and expense transactions. Search vendor names such as OpenAI, Anthropic, Cursor, Perplexity, Otter, Fireflies, Midjourney, ElevenLabs, HeyGen, and API providers. The point is not that these vendors are unsafe. The point is finding paid accounts that nobody in IT knows about, plus subscriptions that should be moved into a company workspace.

I have seen teams spend days arguing about the perfect DLP product while a forgotten OAuth grant quietly retains access to a shared calendar. Start with the boring exports. They usually win.

How do I classify the tools I find?

Use a short red, yellow, and green list. Keep it readable enough that a department manager can act on it without a meeting.

Red means revoke or investigate now. Examples include an unknown meeting bot with calendar and mailbox access, a browser extension reading an HR or payroll portal, or an application with broad permissions and no business owner. Save the app name, user, permissions, date discovered, and reason for the decision before revoking access.

Yellow means migrate. These are useful tools running through consumer accounts, personal email addresses, or unmanaged API keys. Ask the user what task the tool handles, then buy the appropriate commercial workspace or replace it with a tool already covered by your contracts. A block without a replacement is theatre.

Green means approved with a named owner. Record the vendor, plan, administrator, data categories allowed, retention terms, and renewal date. NIST's Generative AI Profile, NIST AI 600-1 points organisations toward inventories, acceptable-use rules, and risk management rather than a domain blacklist.

Do not confuse a business plan with permission to paste anything into it. OpenAI's business privacy documentation says business data is not used to train its models by default under its business and API terms. Anthropic and Google Workspace publish similar commercial-data commitments in their documentation. Those statements do not remove the need to restrict customer data, credentials, HR records, and unreleased financial information.

What can a technical audit miss?

A technical audit sees company devices and company networks. It does not reliably see AI use on a personal phone, a home computer, or a local model running on a developer's laptop. It also misses AI features embedded inside software you already approve. Check the AI settings and subprocessors in your main CRM, meeting platform, helpdesk, design tool, and collaboration suite.

There is another blind spot: logs tell you that a group visited an AI service, but not whether they were summarising public research or pasting a customer export. That is why I would not start with prompt inspection or TLS decryption. Those controls can capture credentials and sensitive personal information, and employee monitoring rules differ by country.

If you operate in the European Union, get legal advice before individual-level monitoring. The Article 29 Working Party's Opinion 2/2017 on data processing at work discusses transparency, proportionality, and the weakness of employee consent in an employment relationship. Report aggregate patterns where possible. In Germany, a works council may have co-determination rights when a technical system can monitor employee behaviour.

Should I use an anonymous survey?

Yes. The survey fills the gap between tool telemetry and actual work. Present it as a procurement and workflow exercise, with a written amnesty for past tool use. Say what you will do with the answers. Disable email and IP collection. In a 100-person company, avoid asking for both a narrow department and a seniority level because a three-person team is not anonymous for long.

Ask five questions: which AI tools people used for work in the last 30 days, what tasks they completed, what information those tasks required, which paid tool the company should consider, and whether they built an automation or custom assistant that colleagues now depend on. Include work done on personal devices.

The survey is also a service-quality test. If half the sales team names the same unapproved tool for call summaries, the first action is probably a sanctioned recorder with sensible access controls. If one engineer has a local model and a script nobody else understands, you have an ownership and continuity problem before you have a policy problem.

How can I audit shadow AI in one week?

Day 1: Export OAuth grants, browser extensions, and 90 days of expenses. Record the data source and its collection date.

Day 2: Review seven to 30 days of DNS or cloud-discovery data, depending on what your logging system retains. Send the anonymous survey with an executive message that says the aim is to fund useful work, not punish old mistakes.

Day 3: Join the exports. Put each tool into red, yellow, or green, and give every green or yellow tool a named business owner.

Day 4: Revoke red access after checking whether the app supports a business-critical process. Provision the approved alternative for yellow users. Microsoft documents admin-consent controls in Entra; use them for broad Mail, Drive, and Calendar permissions.

Day 5: Publish a one-page data rule. Green can include public information. Amber can include internal material only in approved commercial accounts. Red should include credentials, customer personal data, HR records, and unreleased financial information unless a specific review allows it.

Days 6 and 7: Meet with the owners of the red and yellow tools, remove abandoned subscriptions, and write down the next review date. The output should be a living register, not a PDF that dies in a shared folder.

What should the final shadow AI register contain?

Keep one row per tool. Record the vendor and product name, owner, users or departments, sign-in method, permissions, data categories, approved plan, retention position, contract or DPA link, decision, and review date. Add a field for embedded AI features so Notion, Slack, Zoom, HubSpot, and similar systems are not treated as ordinary SaaS forever.

I would review the high-risk rows monthly and the full register quarterly. That cadence is a recommendation, not a legal requirement. The important part is that a new OAuth grant or expense does not remain invisible until renewal season.

A shadow AI audit works when employees can move from an unapproved tool to an approved one without losing the workflow that made them adopt it. If your company needs help mapping the tools, permissions, and migration path, contact Deuce Works.