← All posts

ai:inventory

Shadow AI: how to find the AI tools your staff already use.

Shadow AI is any AI tool or AI feature staff use for work without an approval on record: consumer chatbots on personal accounts, subscriptions bought on a team card, API keys added to a script, and AI features switched on inside software you already license. Any of these can be harmless. What makes them shadow AI is that nobody has matched the tool against the data going in, the terms that apply and the review the output needs.

This post is the how-to: six checks on systems you already operate, five fields to record per tool, and a decision route for each entry. It is written for the IT, security and risk leads who get handed the task once a policy or framework says start with an inventory.

What shadow AI covers.

Four patterns carry the name. Consumer chatbots, image generators and transcription sites used on personal accounts for work tasks. Subscriptions bought on a team card under the finance approval threshold. API keys to model providers, added by engineering, data or analytics teams inside scripts and notebooks. And AI features switched on inside software you already license, which arrived through a vendor update rather than a purchase.

In the fourth pattern, nothing was installed. Office suites, CRMs, support desks and code hosts all ship assistants and summarizers now, and a default or a checkbox somewhere decided the tool is in use.

The scale is measured: Microsoft and LinkedIn's 2024 Work Trend Index (published 8 May 2024, based on a survey of 31,000 people across 31 countries) found that 78 percent of AI users bring their own AI tools to work. The inventory shows which tools sit behind that figure in your own organization.

Why the inventory comes first.

A governance framework starts by asking what you already run. You cannot classify data you do not know is leaving, approve a model you have never seen, or attach oversight to a workflow with no owner. Our enterprise AI governance framework begins there: classification, approval and oversight all assume a list.

NIST's AI Risk Management Framework puts the inventory in its Govern function. GOVERN 1.6 is the outcome that mechanisms to inventory AI systems are in place and resourced according to organizational risk priorities, and the AI RMF Playbook adds that partial inventories may not provide the value of a full one. An inventory listing what IT bought while missing what staff brought in is exactly that case.

Where to look for shadow AI tools.

Each of the six sources misses something. The inventory is the merge of all six, checked against each other. Work through them fastest first:

If you would rather not run them by hand, our AI readiness assessment audits how your teams work, the data AI would touch and the tools already in use: two weeks inside your workflows, then a written roadmap of tools, workflows and rollout order.

  • Sign-in and identity logs. Pull the sign-in logs your identity provider still holds (Microsoft Entra ID keeps 30 days on P1 and P2, seven on Free, unless routed to storage) and match application names against a published list of AI tools. What it misses: every tool used without the company sign-in, personal accounts included.
  • Cloud-app discovery. For a Microsoft 365 estate, Microsoft Defender for Cloud Apps keeps a Discovered apps page in the Microsoft Defender Portal, filterable by risk score, app category, usage and other criteria; some Microsoft Purview solutions group AI apps into Copilot experiences and agents, enterprise AI apps connected through Entra registration and other methods, and other AI apps detected through browser activity and categorized as Generative AI in the Defender for Cloud Apps catalog (both pages checked 16 November 2026). What they miss: traffic that never crosses the managed browser or monitored path, including phones on mobile data and home devices.
  • Expenses and card statements. Search twelve months of descriptions for tool names and the billing names of known model providers; subscriptions priced under an approval threshold skip review by design. What it misses: free tiers, and paid tools on personal cards that never reach an expense claim.
  • Browser extensions. An extension that summarizes, transcribes or rewrites pages handles the same text staff would paste into a chatbot. Pull the installed-extension list from your browser management tooling where browsers are managed, and ask about the rest in the survey. What it misses: extensions in personal profiles and unmanaged browsers, where no policy reaches.
  • AI features inside licensed software. Go past the admin console. Diff your license add-ons against the base SKUs, since AI features arrive both as paid add-ons approved once and as defaults switched on in a vendor update. List the third-party apps and agents connected inside the suite, which teams can add without central IT. Read the release notes of your five biggest vendors each quarter. What it misses: per-user toggles, and AI in products nobody told IT it owned.
  • A staff survey with amnesty. One page, two questions: which tools do you paste work content into, and what would you use if we provisioned it. Say plainly that nothing bad happens to answers, and put a deadline on it. What it misses: answers shaped by fear or forgetfulness; a tool named only in the survey still gets an entry.

What to record for each tool.

Five fields turn a discovery exercise into a working inventory; four come from whoever uses the tool:

  • Owner. The named person who answers for the tool. Every entry needs one, and 'the team' does not count.
  • Users. Which teams run it and who outside them sees the output; this sets the blast radius of a wrong answer.
  • Data classes going in. The highest class of data anyone pastes or connects, judged against the classes of your AI data classification policy. When it depends on who is using it that day, record the highest case.
  • Account type. Company account, personal account used for work, or a shared login. This decides whose terms apply and whether a usage log exists at all.
  • Terms that apply. For that exact tier: whether inputs train the vendor's models, how long inputs are retained, and where processing runs. The consumer and enterprise terms of the same product can differ.

Deciding on each tool: approve, limit, replace or stop.

Give each tool one of four verdicts: approve onto the sanctioned list; approve with limits, meaning a data ceiling and a review step; replace by naming the sanctioned tool that does the same job; or stop, with a reason and a date. The tools you approve then move onto the route our AI model approval process describes: tiers by consequence, evidence per tier, a named decider.

Clear the low-risk bulk first: approving a transcription tool for internal meetings builds the habit of asking. Blocking before any approvals can push the next wave onto personal accounts and home networks, where none of the six checks reach.

Keeping the inventory current.

Two habits keep the list alive. The first is a fast request route: a one-page form with a published turnaround, so that asking is quicker than signing up with a card.

The second is a repeat of the checks on a schedule: a quarterly pass over sign-in logs, discovery data and expenses, and a release-notes check per major vendor at its update cycle. Vendors keep switching AI features on inside licenses you already pay for, so an inventory built once goes stale. Tie each entry to a recheck date and an owner who chases it.

Do it yourself, or bring us in.

The inventory needs a spreadsheet and read access to six systems, nothing more. An IT or risk lead with access to the sign-in logs, the expense system and the admin consoles can produce a first version from the six checks above.

Outside help fits when the repeat checks compete with everything else, or when adoption has started but feels scattered. We audit what is in place and tell you what to keep, what to consolidate, and where you are paying for overlap.

Book a 30-minute call

Not ready for a call? Take the 10-question AI readiness check

Frequently asked questions.

Is shadow AI always a risk?

No. Unapproved and harmful are different questions, and a summarizer fed with public text on a company account can be fine once it is recorded and approved. The risk is that nobody has asked: the data class is unknown, the terms are unread, the owner unnamed.

Should we block public chatbots before the inventory?

Blocking first can move the same usage to personal devices and home networks, where none of your logs reach. Block after the inventory, per tool, and publish the sanctioned alternative on the same day; otherwise the block can start the next wave of shadow AI.

Who owns the shadow AI inventory?

One named owner in IT or risk maintains the record; the AI RMF Playbook's suggested actions for GOVERN 1.6 include naming the individual or team responsible for maintaining an inventory. Each entry also names a business owner, who answers for that tool.

Does a staff survey get honest answers?

Honest enough for the purpose, with amnesty and a short form. Say that nothing bad happens to answers, keep it to two questions, and cross-check the tools named against the log checks. The survey is there to find tools, so do not chase names.

How often should we repeat the discovery?

Quarterly for the log and expense checks, and at each major vendor's release cycle for licensed tools. Vendors keep switching AI features on inside existing licenses, so a once-a-year inventory goes stale between checks.

Keep reading.

:

An enterprise AI governance framework you can actually run.

:

What data employees can put into AI tools: a classification guide.

:

How to run an AI model approval process before production.