Guide · AI & Privacy · Inventory

An AI Inventory in One Afternoon

One afternoon will not produce a perfect AI register. It can produce a defensible first-pass list of use cases, owners, data, decisions, and legal triggers, enough to know what must be investigated first.

By Lincoln Square Law · 8-minute read

Define the unit: one use case, not one vendor

Start by deciding what a row means. Use one row per business use case. A general-purpose assistant used for marketing copy, contract summarization, recruiting, and customer support should appear as four rows because the data, affected people, decisions, and legal risks differ. This choice prevents a deceptively short inventory that says “Microsoft,” “Google,” or “ChatGPT” while hiding dozens of operational uses.

For the first pass, record: use-case name, department, system or vendor, owner, users, affected people, purpose, input data, output, decision influenced, automation level, human reviewer, geography, deployment status, and contract owner.

NIST’s AI RMF starts with understanding context and governing risk across the lifecycle. These fields provide the minimum facts needed to do that work in a business setting.

Hour one: pull systems from places people already maintain

Do not begin with a blank spreadsheet and a companywide email asking, “Who uses AI?” Pull likely systems from procurement, accounts payable, SSO/application directories, browser-extension lists where available, HR technology, marketing technology, customer-support tools, security products, analytics platforms, and active vendor contracts. Search contract descriptions for “AI,” “machine learning,” “automated,” “prediction,” “scoring,” “recommendation,” “copilot,” “assistant,” “transcription,” and “biometric.” Then add unsanctioned or free tools through targeted interviews. Ask each department what tool helps them draft, summarize, rank, transcribe, predict, route, or decide. People often do not describe a feature as AI even when it fits the legal or risk question. The inventory interview should therefore ask about function, not terminology.

Hour two: map the decision and the people affected

For each candidate system, ask two questions: “What happens because of this output?” and “Who can be affected?” A drafting assistant that produces an internal first draft is different from a tool that ranks job candidates, recommends a patient outreach priority, sets a customer fraud flag, or selects an employee for discipline. Record whether the output is advisory, whether a person routinely reviews it, whether that person can meaningfully disagree, and whether the system automatically takes action.

This is where legal triggers appear. Illinois Human Rights Act section 2-102(L) makes employment uses particularly important because it covers AI used across recruitment, hiring, promotion, training, discipline, discharge, and other employment terms and conditions. If a row touches one of those decisions, flag it for employment-AI review. Do not wait until the inventory is “complete” to escalate a high-impact use.

Hour three: map data classes and special triggers

Add a small set of data flags rather than trying to document every field immediately: personal data, employee/applicant data, health information, financial data, children’s data, confidential business information, source code, customer content, and biometrics. For biometrics, use the statutory categories rather than a vendor’s marketing vocabulary. Illinois BIPA expressly identifies fingerprints, voiceprints, retina/iris scans, and scans of hand or face geometry, along with biometric information based on covered identifiers used to identify a person.

For recruiting, add a recorded-video flag. If the employer asks applicants for recorded video interviews and uses AI analysis for Illinois-based positions, the Artificial Intelligence Video Interview Act has specific notice, explanation, consent, sharing, deletion, and in some circumstances reporting requirements. One checkbox in the inventory can route that row to a much more detailed legal review.

Add an “unknown” value and use it aggressively

A first-pass inventory should expose uncertainty rather than hide it. If nobody knows whether prompts train a shared model, write “unknown.” If the vendor has not explained whether it generates face geometry, write “unknown.” If the team cannot identify who can override a score, write “unknown.” Each unknown becomes a diligence question with an owner and due date.

This is more defensible than guessing. The legal team can distinguish an unresolved factual question from a legal conclusion and prioritize the questions that change obligations. An unknown about whether a recruiting tool analyzes recorded video is more urgent than an unknown about a low-risk drafting assistant’s model parameter count.

Create three risk lanes before the afternoon ends

Use a simple triage, not a final risk score. Red: systems that make or materially influence decisions about employment, eligibility, access, health, safety, or other significant interests, systems processing biometrics or sensitive data, systems with autonomous external action, or systems with unresolved high-impact questions. Amber: customer-facing generation, monitoring, analytics, or internal systems using meaningful confidential data. Green: low-impact assistive uses with limited data and meaningful review. The label tells the team what to investigate first. It is not a legal conclusion.

Tie each lane to a next step. Red rows receive legal, privacy, security, and vendor review before expansion. Amber rows receive targeted diligence and approved-use instructions. Green rows receive baseline acceptable-use and data-handling controls. If the facts change, the lane changes.

Reconcile the inventory with notices and contracts

Before closing the session, choose the highest-risk five rows and compare them to the contract and external notice. Does the contract permit the actual data and use? Does the privacy or employment notice describe the use? Can the vendor support deletion or an accommodation? Are model changes disclosed? A mismatch is often more actionable than another hour spent discovering low-risk tools.

Also record the contract renewal date. AI governance is much easier when legal review occurs before an auto-renewal, when the organization still has leverage to demand data-use restrictions, change notice, audit evidence, deletion support, or an exit path.

Turn the spreadsheet into a recurring control

At the end of the afternoon, assign one inventory owner and one owner per row. Add “last verified,” “next review,” and “material change trigger.” Require procurement to route new AI vendors into the inventory, require system owners to report material feature changes, and give employees a simple channel to disclose new tools or use cases. The inventory should not depend on an annual scavenger hunt.

Use NIST’s Govern, Map, Measure, and Manage functions as a maturity path. The afternoon exercise is mostly Govern and Map: identify ownership, context, affected people, and risks. The red-lane systems then move into Measure and Manage work: testing, assessments, mitigations, monitoring, and documented decisions. That sequencing keeps a small organization from treating “AI governance” as a six-month policy project before it even knows what technology it uses.

What a good first afternoon produces

A useful outcome is not a perfect census. It is a working register with enough coverage to reveal high-impact systems, enough ownership to obtain missing facts, enough legal flags to route sensitive uses, and enough dates to keep the record current. You should leave with a short red list, a question queue, and named owners, not with a false claim that the organization has inventoried every algorithm.

Repeat the exercise monthly at first, then move to a risk-based cadence once procurement and change-management feeds are working. Over time, the inventory becomes the legal and operational index for assessments, vendor diligence, notices, incident response, policy exceptions, and regulatory change. That is the difference between an AI spreadsheet and an AI governance control.

Add one evidence link per row before you stop

A row without evidence becomes stale quickly. Link each use case to at least one authoritative internal artifact: the contract, security review, system owner page, vendor technical documentation, configuration record, or approved procedure. For higher-risk rows, link the privacy notice, assessment, model documentation, or legal review as well. The goal is not to turn the inventory into a document repository. It is to make verification possible without another round of interviews.

Then identify the evidence gaps that block a conclusion. If an owner cannot produce the vendor terms, mark contract status unknown. If nobody can show whether a recruiting feature is enabled, assign a configuration check. If the organization cannot explain where prompts are retained, route the question to the vendor. A short, explicit gap list is one of the most valuable outputs of the afternoon because it converts uncertainty into work.