Building an AI inventory: tracking where AI is used in your agency

Reference article · 5 min read

A council resolution, a state executive order, or an audit request arrives asking the agency to produce a list of the artificial intelligence it uses. Most agencies do not have that list. What they have is a contract file with hundreds of software agreements, a few pilots started by individual departments, and several vendor products that added a summarization or scoring feature in a release note nobody read.

An AI use case inventory, sometimes called an AI register, is the record that answers the question. It is a maintained list of where the agency uses AI, with enough detail on each entry that a reader can tell what it does, who is accountable, what data it touches, and whether a person reviews its output. Building the first version is discovery work more than technology work.

What the inventory is for

The inventory serves several audiences at once, which is why the fields matter more than the format. It answers requests from a governing body, an auditor, or a records requester. It gives a governance committee a basis for deciding which uses need review before they expand, shows security and privacy staff where agency data reaches a model provider, and saves procurement from rebuilding the same questions on every purchase.

Federal agencies have been required to publish annual AI use case inventories under executive orders and Office of Management and Budget guidance, and several states have issued similar directives to executive branch agencies. Those documents have been revised more than once, so confirm the current requirement with the issuing office or your counsel rather than a secondhand summary. The NIST AI Risk Management Framework treats this cataloging as part of its govern and map functions, and the structure is worth borrowing even where nothing requires it.

Deciding what counts as AI

The scoping question stalls more inventories than anything else. Statutory definitions vary, and many describe a machine-based system that infers from input how to generate outputs such as predictions, recommendations, or decisions. Adopt one definition in writing at the start, and apply it consistently.

A practical filter: include systems that generate text, images, audio, or code; that classify, score, rank, or prioritize people, cases, or assets; that predict future conditions; or that recognize speech, faces, plates, or objects. Exclude deterministic rules engines and ordinary statistical reporting. List borderline items with a note explaining the call. An overinclusive first pass is easier to defend than one that quietly leaves things out.

Fields that belong in each entry

System name. The name staff actually use, plus the product it lives inside.
Owning department and business owner. A named person, not a unit, since this field keeps the rest current.
Purpose. What decision or task the system supports, and who is on the receiving end of its output.
Data used. Categories and source systems, flagged for personal, criminal justice, health, or student information.
Vendor, integrator, and model provider. The company you contract with is often not the company running the model.
Deployment and data flow. Vendor cloud, agency cloud, or on premises, where data is processed, and whether agency inputs are used to train vendor models.
Human review point. Where a person reviews, approves, or overrides the output, and whether that review is documented.
Risk tier. Your own scale, with rights-affecting and safety-affecting uses at the top.
Status and dates. Pilot, production, or retired, with the go-live date, the next review date, and any testing results on file.

Finding the uses already in the building

Most of what belongs in a first inventory is already licensed. Work several sources, not one survey.

Contract and license records. Flag any licensed product line that has added AI features, which is most of them. Single sign-on logs surface applications IT never procured.
Admin consoles. Productivity suites, help desk tools, and case management systems often ship AI features enabled by default or by one tenant toggle.
Written vendor attestations. Ask each significant vendor, in writing, what AI components the product includes and which subprocessors run them.
Department interviews. Ask about tasks, not technology: who summarizes documents, drafts responses, transcribes interviews, translates notices, or scores applications.
Common hiding places. Transcription, translation, redaction, resume screening, call routing, website chatbots, eligibility scoring, predictive maintenance, and camera analytics.

What to ask vendors and put in the contract

Ask before signing and again at renewal: which features use AI, which model providers and subprocessors are involved, where data is processed, whether agency data is used for training, what testing evidence exists, and how customers are notified before a new AI feature turns on.

In the contract, agencies commonly seek advance notice before new AI functionality is turned on, the ability to disable it, a prohibition on training vendor models with agency data without consent, disclosure of material model or subprocessor changes, cooperation with records requests and audits, and documentation that supports the agency's own inventory reporting. Have counsel review the language, since terms that are routine in one state are unusual in another.

Who keeps it current

Assign the record to one office, typically the CIO's organization, a privacy officer, or an AI governance committee, with a named liaison in each department. A list nobody updates is worse than none, so attach updates to events that already happen: a question on the requisition and RFP intake forms, a checkpoint in the security review, and an attestation at contract renewal. Keep retired entries with an end date so the history survives staff turnover.

Start small. Write the definition, pull the contract list, interview a few departments, assign tiers, and publish an internal version with the gaps marked. The second version is easier.

HQT provides technology advisory and systems engineering services to state and local agencies, including AI governance documentation and inventory development.

Frequently asked questions

What is an AI use case inventory?

It is a maintained list of the places an agency uses artificial intelligence, with details attached to each entry such as the owning department, the purpose, the data involved, the vendor and model provider, a risk tier, and where a person reviews the output. Agencies build one to answer mandates from a governing body or legislature, to support security and privacy review, and to give procurement a consistent reference. It is a governance record, not a technical asset database, so it can live in a spreadsheet or a simple internal application.

What counts as AI for the purposes of an agency inventory?

Definitions vary by statute and by directive, so adopt one written definition at the start and apply it consistently. A practical filter is to include systems that generate content, that classify, score, rank, or prioritize people or cases, that predict future conditions, or that recognize speech, faces, plates, or objects, and to exclude deterministic rules engines and ordinary statistical reporting. Record borderline items with a note explaining the decision rather than leaving them out silently.

How do we find AI features inside software the agency already licenses?

Start with the software contract and license list and flag product lines that have added AI capabilities, then review admin consoles where features are often enabled by default or by a single tenant setting. Ask each significant vendor, in writing, which components use AI and which subprocessors run them. Interview departments about tasks rather than technology, asking who summarizes documents, drafts responses, transcribes interviews, translates notices, or scores applications.

Who should own the AI inventory in a city, county, or state agency?

One office should own the record, commonly the CIO's organization, a privacy officer, or an AI governance committee, with a named liaison in each department responsible for its own entries. Ownership must be a named person rather than a unit, since entries go stale when nobody is accountable for them. The owning office sets the definition, the fields, and the review cadence, while departments supply and confirm the content.

How do we keep the inventory from going out of date?

Attach updates to events that already occur instead of relying on a periodic sweep. Add an AI question to the purchase requisition and RFP intake forms, make the inventory entry a checkpoint in security review, and require an attestation at contract renewal. Keep retired systems in the record with an end date so the agency retains a history when staff change.