The AI inventory you think you have

Al­most every or­ga­ni­za­tion I talk with now has an AI in­ven­to­ry. Far fewer could tell you how they know it is com­plete. An in­ven­to­ry is not a list of AI in use. It is a list of AI that some­one told you about, and those are not the same thing.

What I’m seeing

I have built an AI use-case in­take process and an in­ven­to­ry of more than 50 use cases mapped to NIST AI RMF and ISO/IEC 42001. The most use­ful les­son from that work was not about the map­ping. It was about the front door. An in­ven­to­ry fills up through in­take, and in­take only catch­es what peo­ple think to sub­mit.

That leaves pre­dictable blind spots:

• AI that arrived inside existing software. A vendor adds a summarization feature or an AI assistant to a platform you already license. No one filed a request, because no one bought anything new.
• Use that grew past its approval. A tool approved for drafting internal notes is now drafting customer communications. The inventory still shows the original purpose.
• Experiments that quietly became production. A pilot ran, people liked it, and it kept running. The inventory still says “pilot,” or it never listed the pilot at all.
• Entries that are present but hollow. The use case is listed, but the owner left the company, the risk rating is a placeholder, and the last review date is blank.

None of this means the pro­gram failed. It means the in­ven­to­ry re­flects a mo­ment, and the or­ga­ni­za­tion kept mov­ing after that mo­ment.

Why it matters

The in­ven­to­ry is the foun­da­tion for al­most every­thing else in AI gov­er­nance. Risk rat­ings, mon­i­tor­ing, ven­dor re­views, pol­i­cy at­tes­ta­tions, and board re­port­ing all as­sume the in­ven­to­ry is right. If a use case is miss­ing, every down­stream con­trol miss­es it too.

The frame­works make this de­pen­den­cy clear. NIST AI RMF starts with un­der­stand­ing con­text and map­ping AI sys­tems be­fore you can mea­sure or man­age their risk. ISO/IEC 42001 ex­pects an or­ga­ni­za­tion to de­fine the scope of its AI man­age­ment sys­tem and know which AI sys­tems fall in­side it. Nei­ther works from a par­tial list.

There is also a re­port­ing risk. Imag­ine a leader telling the board, “we have 40 AI use cases under gov­er­nance.” The board hears “we have 40 AI use cases.” If the real num­ber is mean­ing­ful­ly high­er, the gap is not just an op­er­a­tional issue. It is a state­ment that turned out to be in­ac­cu­rate.

So the ques­tion is not “do we have an in­ven­to­ry?” It is “what sources, other than in­take, con­firm the in­ven­to­ry is com­plete?”

What to do this week

1. Reconcile the inventory against one independent source. Pick a source that does not depend on self-reporting: your vendor or contract list, software license records, network or browser traffic for known AI services, or expense reports. Compare it to the inventory and list what appears in one but not the other.
2. Ask your top vendors a direct question. For the handful of platforms your teams use most, ask the vendor in writing which AI features are currently enabled in your environment and which are on by default. Add anything you did not already know about to intake.
3. Run a currency check on ten entries. Pull ten inventory records at random. For each, confirm the owner is still in the role, the stated purpose matches actual use, and the last review date is within your policy. Note how many pass all three checks. That ratio is a more honest measure of inventory quality than the total count.

A question for your next meeting

“Other than peo­ple sub­mit­ting re­quests, how do we find AI that is al­ready in use here?”

If the only an­swer is in­take, you have a list of what was de­clared, and the next step is build­ing a way to find what was not.

Until next time,

Nath

---

The Ev­i­dence Brief shares prac­ti­cal per­spec­tives on tech­nol­o­gy risk, audit, and AI gov­er­nance. It is gen­er­al in­for­ma­tion, not legal or reg­u­la­to­ry ad­vice.