Every executive team we talk to estimates the same thing: "We're probably using four or five AI tools." Then the audit happens. By 5 PM on day one, the estimate is usually wrong by a factor of between five and ten. This page is the story of one afternoon in a real company — what we looked for, what we found, and the three findings that changed everything.

This matters because hidden AI use is your biggest EU AI Act exposure. You can't classify what you don't know exists. You can't document it, set literacy standards for it, or handle incidents involving it. And if an undeclared tool processes personal data or touches a regulated process, the deployer obligations still land on you — you're just doing them blind. For the wider context, start with our full guide to the EU AI Act for UK and Irish businesses.

What happened

The firm: 85 staff, B2B SaaS, UK and Dublin offices.

The MD's estimate going in: four or five tools.

Official register before the audit: nine systems.

The result, by 5 PM: 40 systems documented, 31 of them undeclared.

Finding that changed the roadmap: two systems put the company on the wrong side of EU AI Act deployer obligations.

Elapsed time: six working hours.

How a shadow AI audit actually works

It's not magic. It's methodical. The afternoon breaks into five parallel workstreams, each one hunting for a different category of AI use.

Identity and access logs. The IT lead pulls SSO and OAuth app-grant records — every third-party service anyone has authenticated to, with timestamps and access levels. This catches enterprise tools, personal account logins to SaaS, and browser extensions that request permissions.

Expense and procurement. Finance runs a trawl of the last 18 months of procurement cards, subscription platforms (like Expensify), and invoices. Every vendor that touches data, every SaaS tool approved or under budget oversight, surfaces here. This is where the officially-approved tools live and where the quietly-purchased ones are often hiding.

Browser extensions and OAuth grants. Employees check their own browser extension lists and API-authorisation pages (Google Workspace app permissions, Microsoft integrations, Slack apps). It takes 15 minutes per person and it's where most of the personal ChatGPT accounts, meeting notetakers, and one-person tools surface.

Anonymous staff survey. Six simple questions: what AI tools do you use in your daily work? Have you used any tools without declaring them? Has anyone pasted company data into a free or personal account? The anonymity is critical — the survey only works if people believe they won't be punished for honesty. In this firm, it surfaced two major tools nobody else knew about.

The known unknowns interview. IT and the leadership team list what they know they don't know. Are there any tools they suspect but can't verify? Any legacy systems with AI-like features? Any team that tends to adopt tools quietly? This catches the self-hosted experiments and the team-level shadow buys. For context on what makes undeclared use risky, see our guide to shadow AI compliance exposure.

What the 40 tools were

This is where the register surprises most firms. The tools aren't just cutting-edge AI companies — they're AI embedded in software you already pay for, plus the personal accounts your team is using in parallel.

  1. AI coding assistants — GitHub Copilot, Claude for VS Code, Codeium: 9 across engineering and ops teams.
  2. Personal ChatGPT and Claude accounts — undeclared individual use across multiple teams: 8 accounts with varying approval states.
  3. AI features inside existing SaaS — Slack's summarization feature, HubSpot's AI email writer, native Copilot in Microsoft 365: 7 tools nobody knew were toggled on.
  4. AI meeting notetakers — Fireflies, Otter, Grain recording and summarising sales and internal calls: 5 subscriptions.
  5. Marketing and content tools — generative copywriting, image generation, SEO optimization: 4 active.
  6. Data wrangling utilities — AI-powered Excel macros, SQL helper tools, data-transformation assistants: 4.
  7. Miscellaneous — customer feedback analysis, interview transcription, internal search: 3.

Three vignettes from the survey and the logs deserve the spotlight:

The free ChatGPT paste. A product manager, building a feature spec, had pasted three months of customer feedback into a free ChatGPT account without realizing it was stored and could train the model. Customer data, unencrypted, outside the business. The discovery forced an immediate policy update and a conversation with the legal team.
The unvetted vendor grant. A marketing contractor had granted OAuth access to a generative image tool nobody had assessed or negotiated terms with. The tool was reading the company's brand assets and potentially syncing them. OAuth grants are easy to miss because they don't appear on procurement invoices — they just sit in app-permission lists.
The NDA meeting recording. One of the AI notetakers had transcribed a sales call involving a prospect under NDA. The tool was storing and indexing the conversation. Nobody had set retention policies or reviewed the vendor's handling of confidential content.

The three findings that changed everything

Finding something is one thing. Understanding what it means is different. Three discoveries shifted the company's entire roadmap:

Finding 1 — The register covered only one quarter of actual use

The company's official AI register, built the month before from procurement records alone, held nine systems. The full audit found 40. That 9/40 gap matters because a register built from procurement is a floor, not a ceiling. It catches approved vendors, but it misses embedded AI in existing tools, personal accounts, and team-level shadow buys. The inventory is what governance depends on — an incomplete picture is the biggest single compliance risk.

Finding 2 — Two systems triggered deployer obligations the company wasn't aware of

The firm had built a custom feature on top of Claude's API that uses AI to score proprietary data — a deployer on high-risk AI. And a meeting notetaker was recording customer calls and generating summaries — that's limited-risk by classification, but it triggered transparency duties the company had never documented. Both systems had been filed mentally as "just tools we pay for." Both required immediate escalation to the engineering and legal teams. For the classification framework, see the risk categories explained in detail.

Finding 3 — The highest-risk uses weren't in engineering

The firm's worst exposures came from sales and HR, not the engineering team. Personal data pasted into free ChatGPT accounts, recordings of confidential calls, and vendor integrations to systems holding employee records. The highest-risk use of AI in the company lived in the people who handle the most sensitive information — which is exactly where training and policies land first.

What changed in the week after

Finding a problem is pointless if you don't act. Here's what the firm did:

  • Six tools were retired. Redundant vendors, unapproved personal accounts, tools with unacceptable terms. No drama, no punishment — just sunsetted.
  • Eleven tools were formally adopted. Named owners assigned, documentation pulled together, data-handling and retention policies written down. Shadow use became governed use.
  • Literacy training was triggered. Every user of a high-risk system went on the Article 4 literacy plan — scheduled, recorded, owned. The register's compliance obligations became a training calendar.
  • Logging expectations were set. The IT lead and engineering manager aligned on what gets logged, for how long, and who has access to the logs. Incident handling expectations followed.
  • The register moved to quarterly review. Instead of a one-time exercise, the register became a standing 30-minute meeting every quarter. New tools, retirements, risk-class changes, new gaps — the picture stays current.

Notice the framing: none of this was punishment. A personal ChatGPT account didn't become a firing offense. It became a documented decision with a policy wrapper. An unapproved vendor didn't get blacklisted — it got a conversation about terms and data handling. The goal wasn't prohibition. It was visibility and control.

Eighty-five people. Forty AI tools. Six hours. That's what a Shadow AI Audit delivers. Discovery, classification, risk summary, and a prioritised roadmap. It starts with a 20-minute call.

Enquire about the Shadow AI Audit

Could this be your company?

If your AI register was built from procurement records alone, the odds are good. You have a floor, not a ceiling. One focused afternoon tells you what you're actually exposed to.

That's exactly what the Shadow AI Audit is: structured discovery, risk classification, and a remediation priority list, delivered in two weeks. Not consultancy, not a steering committee. A fixed-scope, fixed-fee engagement. We do this in two weeks, fixed fee, or you can run it yourself if you have the right sponsor and the person with the classification expertise to avoid filing everything as "minimal." Either way, the point is the same: you can't govern what you don't see.

Not sure if the Act applies to you yet? Take the AI Act quiz (2 min). If the result surprises you, this page is where you come back to.

Client details anonymised; figures unchanged. Practical guidance, not legal advice — we work alongside your legal counsel.