Regulation (EU) 2024/1689 · Artificial Intelligence Act ● Annex III: — days

How to Find Shadow AI: Seven Discovery Methods That Don't Spy on Employees

You cannot classify what you cannot see — and you cannot see it with a spyware rollout nobody consented to. Seven discovery methods, ordered by signal quality, that inventory AI usage while keeping the survey voluntary and the monitoring on systems, not people.

Renomira Team · 10 min read

Every classification program dies at the same step: the inventory is a guess. Teams file a beautiful register of the twelve tools IT approves, while nineteen more run in browser tabs, personal accounts and expense claims. The Act does not require omniscience — but it requires that what you document matches what runs, and enterprise questionnaires assume you know the difference.

An open-plan office full of computers — the estate shadow AI hides in
Photo: Bertrand SOUBEYRAND / CC BY 4.0 via Wikimedia Commons

Shadow-AI discovery has a second constraint in the EU: it cannot become employee surveillance. Discovery that monitors people invites works-council negotiation (or worse, a prohibited-purpose argument); discovery that reads system and spend signals needs none. The seven methods below are ordered from strongest signal to softest.

The seven methods

1. SSO and identity-provider logs — the backbone

If the company runs Okta, Entra ID or Google Workspace, every SAML/OAuth app grant is a record: app name, first seen, last used, who granted it. Filter for AI-category applications and you get adoption dates and organisational usage without reading a single message. This is your backbone — it sees sanctioned SSO-integrated tools, misses personal-account usage.

Signal: high. Privacy: system metadata only.

2. Expense and procurement trails — follow the money

Personal Card SaaS subscriptions, invoices with “AI” in the vendor name, procurement requests that never reached IT. Finance data answers “who pays for what” at the vendor level — the correct granularity — and surfaces tools too niche to appear in any app catalogue.

Signal: high for paid tools; misses free tiers. Privacy: vendor-level, standard spend review.

3. Network and DNS telemetry — what the estate actually talks to

Gateway logs, secure-web-gateway categories, DNS query volumes to known AI endpoints. Aggregated per service — how much traffic hit which AI domain — it reveals usage patterns IT never approved, including teams on shared infrastructure. Read it per tool, not per person: the finding is “AI endpoint X saw 400 sessions weekly,” not a named list.

Signal: high, including some free-tier use. Privacy: aggregate by service, retention-limited, works-council-clean when scoped that way in writing.

4. Browser extension and managed-device inventory

For managed browsers and endpoints: the extension list alone names writing assistants, meeting bots and summarisers. Device management consoles report installed AI clients. This is configuration data about company-managed assets, which is why it reads cleanly — the device is theirs, the inventory is administrative.

Signal: medium-high on managed estates; blind to phones and personal laptops. Privacy: inventory, not content.

5. The anonymous disclosure survey — the human signal

Technical methods miss the genuinely personal: the contractor’s account, the tool used on a phone, the workflow nobody would put in writing. An anonymous survey — which AI tools have you used for work in the last 30 days, for what tasks? — closes the gap with the highest-yield single answer: employees already know what they use, and mostly will say so if the framing is discovery, not discipline.

Design rules that make it work: anonymity that is real (no free-text identifiers), explicit “no individual consequences” framing agreed with HR or the works council, and a published outcome (“we found 14 tools; 3 are going on the approved list”). Run it as a systems inventory and it complements methods 1–4 rather than duplicating them.

Signal: surprisingly high — 14 disclosures in week one is a typical result. Privacy: voluntary and anonymous; the reason it gets answered.

6. Helpdesk and internal-knowledge signals

Tickets asking “can we use X?”, wiki pages documenting unofficial workflows, Slack/Teams channels named after tools. Text search of internal documentation surfaces intent and workarounds. This is organisation-level archaeology: you are reading the company’s own notes about itself.

Signal: medium. Privacy: search over internal docs, same as any knowledge-base audit.

7. The quarterly tool-request review — catching tomorrow’s shadow AI today

Every tool request that bypassed procurement, every renewal that auto-charged, every “temporary” trial older than 90 days. A standing quarterly sweep converts discovery from a project into a hygiene rhythm — and pairs with a fast approval lane, because shadow AI grows in the vacuum left by slow approvals.

Signal: preventive. Privacy: process data only.

Sequencing: the two-week sweep

Week 1 — signals 1–4 in parallel. Identity grants, spend, DNS aggregates, extensions. One spreadsheet: tool, evidence source, estimated reach, owner function (not users), tier guess. Typically 70% of the estate appears here.

Week 2 — survey (5) plus doc sweep (6), then reconcile. Union the lists, dedupe, and mark contradictions: survey-only tools (personal accounts — the real shadow set), telemetry-only tools (nobody admits, everyone uses). The contradiction column is the risk ranking.

Then the payoff loop: classify each tool (or route it to the ban screen), approve the keepers formally, retire the rest, and record the discovery date. The register’s provenance note — found via SSO + survey + DNS, Q4 sweep — is exactly what a questionnaire means by “how do you know?”

What makes this Article 4 evidence, not surveillance

AI literacy duties assume you know where AI runs; discovery is the predicate. The line regulators and works councils read: aggregate by tool, act on tools, never score people. Methods 1–4 and 6–7 are system, spend and document metadata — standard administrative visibility. Method 5 is anonymous and voluntary. None of the seven needs content interception or individual performance data, and that is precisely why they run without a negotiation — and why the outputs still produce a register that survives audit.

How to check your own position in four minutes

The free exposure scan asks twelve questions about your AI usage and governance and returns an exposure score, your ranked exposures with the exact article that fired, and your first three moves — mapped to the dates that apply. Confirm your email and the full report unlocks.

Frequently asked questions

How do companies find out which AI tools employees actually use?

Seven signals, best first: SSO/identity-provider application grants; expense and procurement records; network/DNS telemetry to AI endpoints; browser-extension and managed-device inventories; an anonymous voluntary usage survey; helpdesk and wiki text search; and a quarterly tool-request review. Used together for two weeks they typically surface the full estate — including the personal-account tools no technical method sees alone.

Is discovering shadow AI considered employee monitoring?

It can be, if done wrong. Reading system, spend and device-inventory metadata is standard administrative visibility; monitoring message content or per-person activity patterns is surveillance and invites works-council procedures (Germany’s § 87(1) No. 14 covers any device objectively suitable to monitor behaviour). The clean posture: aggregate findings by tool, never produce individual usage scores, and pair the sweep with an anonymous voluntary survey for the human signal.

What did typical shadow-AI surveys find?

Common first-survey results at mid-market companies: a dozen-plus unapproved tools in week one — writing assistants and meeting bots at the top, plus niche departmental tools nobody outside the team had heard of. The yield is high because the question is low-threat: employees know what they use and answer honestly when anonymity is real and consequences are explicitly not individual.

Once we find shadow AI, what do we do?

Three passes: classify each tool against Article 5 first (remove anything prohibited), then tier it (if kept: approve formally, document classification, add to the register with discovery provenance), or retire it with a documented decision. Speed the approval lane while you sweep — shadow AI regrows wherever sanctioned onboarding is slower than a credit card.

How does shadow AI discovery connect to AI Act compliance?

Classification, literacy and register duties all presuppose an accurate inventory. A register that lists approved tools while unapproved ones run in production fails the moment a buyer or authority cross-checks. Discovery gives the register provenance — dated sweeps with named evidence sources — which is what turns “we think this is everything” into a statement an auditor can rely on.

Next step

Put this guide to work on your own estate.

Run the free exposure scan to see which of these articles actually apply to your company — 12 questions, four minutes, no card. Confirm your email and the full report unlocks.

  • Free forever plan
  • No credit card
  • EU-hosted · GDPR-native
Run a free scan Sign up free

Build a live AI register in minutes — upgrade only when you need the audit pack.

how to find shadow AI toolsdiscover unsanctioned AI usage companyshadow AI detection methodsfind out which AI tools employees useAI usage discovery without surveillanceshadow AI survey vs technical detectioninventory AI tools at work

Keep reading