Outgrace, home
Technology & AIHow-to guide

How to build an AI use-case inventory across your company in six weeks

A repeatable template, a week-by-week schedule and the risk checks to run on each entry. The schedule is our judgement; the frameworks behind it are public.

Ask a chief executive how many AI systems the company runs and you will hear a number someone has already disputed. Procurement counts contracts, IT counts accounts, and a team may have been using a writing tool since last spring without telling anyone. In our experience a complete first list can be built in six weeks, provided the company agrees its scope on day one and holds to one template. The NIST and European Commission pages we reviewed give no company inventory template, timetable or scoring method, so the schedule and fields below are our judgement, built on the frameworks those two bodies do publish.

Before you start

Begin with the purpose. NIST presents its AI Risk Management Framework as a voluntary tool for building trustworthiness into how AI is designed, used and evaluated. That gives the inventory its job: it exists so that someone can judge each use against that standard. Settle the job before anyone opens a spreadsheet, because an inventory built for auditors looks different from one built for product teams, and most companies will need both.

  • A named executive sponsor who can require owners to reply by a set date.
  • One written definition of an AI use, agreed by legal, IT and the business.
  • A register tool or shared workbook that every reviewer can open and edit.
  • Read access to procurement and expense records, because many AI uses arrive as software subscriptions.
  • A rule that the register records what exists today, not what a team hopes to buy.

Check the calendar before you start. The Commission dates its ninth prohibition under the AI Act to December 2026, and says that prohibition arrived through the AI Omnibus. Its transparency rules take effect in August 2026, according to the same page. If you start in November, six weeks ends in December, so test both dates in week one. If any entry could sit in a regulated category, put that question to counsel rather than settling it in a workshop.

  1. Step 1

    Fix the scope and the fields

    Write one definition of an AI use that legal, IT and the business all accept, and keep it to a single page. Then build a template of about fifteen fields: name, owner, purpose, status (live, pilot or planned), vendor or model, data inputs, people affected, whether a person reviews outputs, countries served, risk tier, controls and next review date. Keep the list short for the first version and add fields only when a reviewer cannot answer a question without one. Usually goes wrong: teams add fie

  2. Step 2

    Search every channel

    Pull the vendor list from procurement, the company card and expense records, single sign-on logs and software repositories. Send a short survey to department heads asking which tools their teams use to write, summarise, score, sort or decide things. Ask specifically about spreadsheets with formulas, add-ins and browser extensions, since these are often missed. Usually goes wrong: a search run only by IT finds the tools IT installed and misses the ones a team bought on a card.

  3. Step 3

    Assign one owner per entry

    Every line in the register needs a named person who runs the use and can answer for it. That is seldom the person who signed the contract. Send each owner the template with the entry half-filled from what you found in week two, and ask for a reply within five working days. Usually goes wrong: the vendor manager completes the form from the supplier's sales deck, and nobody who knows the data has seen it.

  4. Step 4

    Record the data behind each use

    For every use, list each data input, where it comes from, whether it contains personal data, and whether it is confidential or customer information. Note whether any output feeds a decision about a person, such as a shortlist, a price, a credit limit or a performance rating. Record whether a human reads the output before anything happens. Usually goes wrong: the entry says customer data and stops there, so nobody can tell whether a support bot sees payment details.

  5. Step 5

    Assign a provisional risk tier

    Set your own tiers and label them as internal. The European Commission says the AI Act sets a risk-based set of rules, and it describes minimal-risk AI, which it says is the vast majority of systems in use, as having no rules under the Act. Score the use, not the vendor's reputation, and record the reason in one sentence. Ask counsel to check any tier that touches a regulated use. Usually goes wrong: a well-known supplier is treated as low risk while the local use is a hiring filter that ranks c

  6. Step 6

    Sign off, date and publish

    Ask the executive sponsor to approve the register, then set a review date for every entry. Hold planned uses in a separate section until they have an owner and a provisional tier. Publish the list to the people who must use it, and keep a changelog so that changes are visible. Usually goes wrong: the register is treated as finished at week six, and by the next quarter it describes a company that no longer exists in that form.

What usually goes wrong

The biggest waste falls in the first two weeks. Teams argue about definitions and forget to count, so by week three there is a long debate and a short list. The second error is counting planned uses as live, which inflates the total and muddles the risk work that follows. The third is accepting a vendor's assurance as a risk assessment. A supplier can describe its product accurately and still have no idea how your staff will use it. The fourth appears only later: a register with no owner and no review date. It describes last quarter, and nobody is paid to notice.

The NIST AI Risk Management Framework (AI RMF) is intended for voluntary use and to improve the ability to incorporate trustworthiness considerations into the design, development, use, and evaluation of AI products, services, and systems.

NIST, AI Risk Management Framework page, nist.gov
Do we need this if we only buy off-the-shelf tools?

Yes, in our view, though the reasons differ from a builder's. A bought tool still shapes decisions inside your business, and the people who rely on it need to know what it touches. NIST describes its framework as voluntary, so it does not by itself impose an obligation on you. According to the European Commission, the EU AI Act applies risk-based rules to both developers and deployers, so whether it reaches your use depends on your role and the use itself. That question belongs with counsel.

Should we use the EU's four risk levels if we are not in Europe?

As a sorting aid, the labels can help. The AI Act defines risk levels, but according to the European Commission, most AI systems in use in the EU fall into a category the Act does not regulate, so a label tells you little on its own. Keep the tier labels as an internal sorting aid, not a legal classification. Keep a separate column for whether each use is regulated in any country where you operate.

Who should own the register after week six?

In our experience, a risk, compliance or operations function is the right home, with the executive sponsor as the escalation point. The owner should run a review each quarter, check new entries against procurement, and keep the changelog current. Avoid giving the job to IT alone, because IT sees the software but not the decisions made with it.

Does the NIST framework cover critical infrastructure operators?

NIST has published a preliminary paper on a profile for trustworthy AI in critical infrastructure, dated 7 April 2026. According to NIST, the note gives operators guidance on risk practices for AI-enabled capabilities. If you run critical infrastructure, read it before you set your tiers. It is a concept note, not a finished profile, so treat it as a direction of travel rather than a settled requirement.

Is there a template we can copy?

Not from these sources. The three public sources give no company-level inventory template, no six-week timeline and no data-input or risk-scoring method. The fields and steps in this guide are our own starting point. Adapt them to your business, and expect the first version to be longer than it needs to be.

On Monday, ask the chief executive or chair to name the sponsor and set the reply deadline. By Wednesday, send a one-page definition of an AI use to legal, IT and two business heads, and ask for objections by Friday. Pull the procurement and expense list the same week, since it is the cheapest route to half the entries. Then book a call with counsel on the two EU dates, because that call, more than the template, will show whether your list needs more than six weeks.

Newsletter

The Briefing

One email a week: the leaders we profiled, the ideas worth stealing, and nothing else.

Weekly, free, one click to leave. Read our privacy policy.