How do we find and record all the AI we use, including AI inside tools we already own?

Start

We need a list of every AI we use

Most organisations do not build AI. They use other people's AI: an assistant in the office suite, a feature in a system they already run, a supplier's chatbot or a free tool a member of staff found. A policy with no list of what is in use cannot be applied, so the inventory comes first. One use is one purpose with one owner, so one tool can have several uses.

• Then: Gather candidates from four sources, not one

Step

Gather candidates from four sources, not one

Ask each manager what their team uses. Read the single sign-on and card-spend reports. Check the release notes of software you own for AI features. Collect supplier notices of new AI features. Judgement: use all four in the first pass, because each misses what another finds. Include contractors and supplier staff who hold your accounts, and AI built into other software (AG-01).

• Then: Ask people, and say nobody is blamed for answering

Step

Ask people, and say nobody is blamed for answering

Free tools on personal accounts will not show in any report, so only asking finds them. Say plainly that nobody is blamed for a report made in good faith (P09 IM-01). Do not send an all-staff email that threatens a penalty, because people hide what they do. Ask what they would need from an approved tool, then approve one: people who have an approved tool stop using the free one.

• Then: Does it infer its outputs from the input it receives?

Decision

Does it infer its outputs from the input it receives?

The Act defines an AI system as a machine-based system that infers from its input how to generate outputs such as predictions, content, recommendations or decisions (Article 3(1)). The pack lists a chat assistant, an image generator, a forecasting feature and a spam filter that learns as AI systems, and a fixed formula as not one. Judgement: if you cannot tell, take the Yes branch and let the supplier or Legal say otherwise. This is a screening answer, not a legal classification.

• Yes, or not known: Do our people use it for our work, on any account or device?
• No: Leave it off, and note why

Outcome

Leave it off, and note why

A fixed formula is not an AI system under the pack's definition, so it does not belong on this list. Judgement: write one line saying why, so the next person who finds it does not ask again. If the supplier later adds a learning or generative feature, treat it as a new candidate.

Decision

Do our people use it for our work, on any account or device?

The policy covers everyone who works for the organisation, including contractors and supplier staff who hold your accounts, whenever they use AI for work, on any device or account. Work done on a personal account counts, and that is the case that shows in no report. If not known, answer Yes.

• Yes, or not known: Could it be a prohibited practice, or take special category data?
• No: Not a work use: leave it off

Outcome

Not a work use: leave it off

Purely personal use outside work is not an entry. If the same tool is later used for work, even once, it becomes a candidate again. What people may use for work is a separate question for the approved list.

Decision

Could it be a prohibited practice, or take special category data?

Ask this before anything else about the use, because these are the two cases the pack stops at once. The prohibited-practice checklist is PP-01 to PP-10 (for example, inferring the emotions of staff at work). Special category data, such as health data, never goes into an AI tool (P15 CL-19). If not known, take the Yes branch.

• Yes, or not known: Stop it now; tell Legal or the DPO the same day
• No: Is it a feature inside software we already own?

Outcome

Stop it now; tell Legal or the DPO the same day

Do not wait for the working days an ordinary unapproved use gets. For a possible prohibited practice, tell Legal. For special category data, report at once through the incident route (P09 IM-01) and tell the Data Protection Officer the same day (AT-3). The pack allows no exception and sets no dated condition (AG-20). Judgement: if you cannot rule either case out, stop the use while Legal or the Data Protection Officer looks. Record the use afterwards as Stopped, with its reason.

Decision

Is it a feature inside software we already own?

AI often arrives as a new feature in a release of software you already run, such as a forecasting module in a warehouse system or a monitoring service with machine-learning detection. It may never have gone through approval, because it was not bought as an AI tool. The pack asks suppliers for notice before they add or change an AI feature (ST-7, P06 clause CL-20), and release notes are one of the ways to find them. If not known, answer Yes.

• Yes, or not known: Record the feature as a use under the system it sits in
• No: Is it already running?

Step

Record the feature as a use under the system it sits in

Give the feature its own row. Name the system it sits in (IF-04, a P07 system) and the supplier (IF-03). The feature is a new use even though the software is old, so it goes through intake like any other (AG-09). The system owner tells the AI governance lead when a supplier adds or changes a feature.

• Then: Is it already running?

Decision

Is it already running?

Most uses found in a first survey are already running. A new use does not start before its intake decision (AG-09). If not known, treat it as running, because people may be using it now and it has to be made safe first.

• Yes, or not known: Record it as running, and make it safe first
• No: Record it as Proposed; it waits for the intake decision

Step

Record it as running, and make it safe first

Ask people to stop putting in anything above Public. Find out what has gone in and report anything above what the tool is approved for as an event (AT-1 to AT-3). Answer the twelve intake questions as the use runs today. A tool not on the approved list is stopped within 3 working days by the pack's default (AG-02), and the people are helped to a tool they may use. Pending a review the AI governance lead may suspend an approved running use, and the tier's approvers decide whether it resumes (AG-10, RS-15). At Tier 2 or 3 the assessment is due within 90 calendar days of the inventory closing, by the pack's default (AG-09).

• Then: Enter one row with the sixteen fields

Step

Record it as Proposed; it waits for the intake decision

A proposed use goes on the list from the start with the status Proposed, because no new use starts before its intake decision (AG-09). The AI use owner answers the twelve intake questions. The AI governance lead screens a complete request within 5 working days by the pack's default (IS-2). Setting the tier is a separate question.

• Then: Enter one row with the sixteen fields

Step

Enter one row with the sixteen fields

Each row carries a unique ID, the purpose, the tool and supplier, the system it sits in, a named owner role (never a team), the role that checks the output, the highest data level, who sees the output, the tier, the status, the approver and date, conditions, the assessment, a pointer to the EU AI Act reasoning, the next review and notes (IF-01 to IF-16). Identical uses may share a row in a small organisation.

• Then: Does any inventory check show a finding?

Decision

Does any inventory check show a finding?

Ten checks read the rows: unique IDs, an owner and a checker, a tool and a supplier in the register, a valid data level with no special category data, a tier not below what the intake answers give, an approval by the tier's approvers and never the use owner, a reason for any hold, decline or stop, an assessment or a due date for Tier 2 and 3 uses, any review past its date listed, and a tool whose approved state allows the data (IV-01 to IV-10). A finding is a gap to close.

• Yes: Fix the cause on that row, then recheck
• No: Is any team, report or system still unread?

Step

Fix the cause on that row, then recheck

Each finding names the check and the row. Correct the row and run the checks again. Where a tier is below what the intake answers give (IV-05), raise it. Where a use has no owner or no checker (IV-02), name them. Where an approver is the use owner (IV-06), the approval is not valid until the executive decision-maker takes that approver's place. Judgement: where the tool's approved state does not allow the data level or personal data (IV-10), take the use back through intake before it goes on.

• Then: Enter one row with the sixteen fields (back to an earlier step)

Decision

Is any team, report or system still unread?

The list is only as complete as the sources behind it. Before you close it, check that every manager has answered, that both reports were read and that release notes were checked for each system with a supplier. If not known, go round again. Judgement: the pack sets no completeness test, so this is the check the flow adds.

• Yes, or not known: Gather candidates from four sources, not one (back to an earlier step)
• No: Close the inventory, date it and set the reviews

Step

Close the inventory, date it and set the reviews

Write down the date the inventory closed, because the assessment dates of running Tier 2 and 3 uses count from it. Set each use's next review within its tier's months, and review the whole inventory at least every 6 months by the pack's default (AG-10). A trigger forces a review within 10 working days (AG-10); the inventory's optional trigger columns are a workbook extension. The triggers are a new purpose, data level or audience, a supplier change or a new AI feature, an incident, a complaint, or a change in law.

• Then: A dated inventory, reported as counts with gaps named

Outcome

A dated inventory, reported as counts with gaps named

Report counts to security and IT leadership with every gap named, for example how many uses have a current approval out of all uses (AIM-01), and never a score or a percentage (AG-22). The inventory is your own record. Keep a copy that can be reached when the main systems are down.