## TL;DR

An AI inventory is the system of record for every model, agent and AI-built app you run: its owner, its purpose, the data it can reach, what it may do and its risk tier. Examiners now ask for one. Banks start from OCC Bulletin 2026-54, and insurers start from the NAIC supplement. Spreadsheet inventories drift three ways: vendors switch on AI features, teams build apps outside IT, and agents pick up permissions. The fix is an inventory the platform writes as the AI runs. There is a free template at the end.

## Introduction

If an examiner asked for your AI inventory tomorrow, would you send them a system of record, or a spreadsheet someone updated last night?

Most companies we talk to know the routine. Once a quarter, a survey goes out to the business units. People list the AI they remember using, someone merges the tabs, and the result lands in a shared drive. It is out of date the day it is finished.

That was fine while nobody outside the company read it. Two regulators now start from that document.

For banks, [OCC Bulletin 2026-54](https://occ.gov/news-issuances/bulletins/2026/bulletin-2026-54.html), published on November 4, 2026, gives examiners a framework that tiers every AI use case. [King & Spalding's analysis](https://www.kslaw.com/insights/articles/the-occ-issues-examiner-framework-for-bank-use-of-ai) reads it as a call for a current inventory, and that includes the AI features vendors have switched on inside software the bank already owns.

Insurers face a parallel requirement. The NAIC's AI Risk Evaluation Supplement, formerly the AI Systems Evaluation Tool, opens with an inventory in Exhibit A. The draft supplement scopes that inventory with a materiality threshold each insurer sets and has to justify, and [the examination evidence](https://actuary.info/insights/naic-ai-risk-evaluation-supplement-examination-evidence-2026) now reaches past the inventory sheets themselves.

An inventory an examiner may read has to hold up on its own. Most of the facts have to come from the systems that run the AI, and whatever people remember can only fill the gaps.

## What is an AI inventory?

An AI inventory is a current record of every AI model, agent and AI-built application an organization runs or buys. Each entry carries a named owner, the purpose, the data it can reach, the actions it is allowed to take and its risk tier, and the record has to stay accurate enough to show an examiner.

Two clarifications matter. First, this is not AI for inventory management, where AI counts stock in a warehouse. Second, the neighboring terms are narrower. A [model inventory](/glossary) is the older model-risk list and covers models only. An AI use case inventory lists purposes, and it is the term US federal agencies use. An AI inventory covers both, and then adds the things those lists tend to miss: agents, AI-built apps and the AI features bundled into vendor software.

## The fields an examiner-ready inventory contains

One row per AI system. That includes the systems you did not build and the features you never asked for.

Here is the field list we would start from, along with how each field lines up with the OCC framework, the NAIC supplement and [the risk areas in the CISA guidance on agentic AI services](https://www.cisa.gov/resources-tools/resources/careful-adoption-agentic-ai-services), and where the answer usually comes from today.

| Field | What it records | OCC 2026-54 (what examiners look at) | NAIC AI Risk Evaluation Supplement | CISA et al. risk area | Usually comes from |
| --- | --- | --- | --- | --- | --- |
| Name, ID and type | Model, agent, AI-built app or vendor feature | Every use case listed | Exhibit A: model name | All five | Survey today; platform registry |
| Owner and backup | The named person who answers for it | Owners who answer for each use case | Governance (Exhibit B) | Accountability | Survey |
| Purpose and business area | The use case and the decision it touches | The use case and its context | Exhibit A: use, use case, program area | Accountability | Survey |
| Acts or advises; autonomy | Supports a person, augments a decision, or acts alone | Agency and human oversight drive the tier | Support, augment or automate; agentic AI category | Behavior | Survey; platform for agents |
| Data it can reach | Sources, sensitivity, customer data | The data the AI may use | Exhibit D (high-risk data) | Privilege | Platform; data catalog |
| Permissions and identity | Actions allowed, systems touched, and whether it holds its own identity or runs on a person's login | Permitted actions, data and systems | Not covered | Privilege | Platform; identity provider |
| Model, version and configuration | Provider, version, prompts or settings in use | Not covered | Third-party models (Exhibit B) | Design and configuration | Platform |
| Third party | Vendor, contract, and whether the vendor switched it on | Vendor-embedded features listed; accountability stays with the bank | Third-party oversight | Structural | Procurement; vendor release notes |
| Dependencies | Other agents, tools and systems it calls or feeds | Not covered | Not covered | Structural | Platform |
| Risk tier | Cost of failure, reversibility, consumer and financial impact | Tier by consequence and reversibility, not by sophistication | Exhibit A: inherent risk, consumer impact, financial impact | Not covered | Risk team |
| Oversight and shutdown | Who reviews it, who can override or stop it | Human review and override; kill switches for higher tiers | Exhibit B governance | Behavior; accountability | Survey; platform |
| Status and dates | Deployed, last changed, last reviewed | A current inventory | Systems implemented in the past X months | Not covered | Platform |
| Evidence | Link to logs, tests and approvals | Testable controls | Exhibit C (high-risk details) | Accountability | Platform |

The mapping above is our reading of each document, not their exact wording. Two rows carry the most weight with examiners: the risk tier and the shutdown path. [BERI's analysis](https://www.beri.net/article/occ-bulletin-2026-54-super-intelligence-supervisory-framework-bank-ai-agents-risk-tiers-consequence-kill-switch-vendor-contracts) expects examiners to ask for the inventory, the risk tiers and a tested shutdown path, and [the walk-through of the CISA guidance](https://www.kiteworks.com/?p=228769) sets out the privilege, audit and oversight questions behind the last two columns.

Four of those fields are the ones firms leave blank most often, and they carry the most weight. Data reach, permissions, identity and last change are facts about systems. People do not know them reliably, and the answers change without any change request.

## Three ways an inventory drifts

Even a good inventory starts going stale the moment you finish it. In our experience, it happens in three predictable ways.

### Vendor AI features get switched on

Your CRM, ticketing system, HR suite and office tools all ship AI features in routine releases, and some of them are on by default. Nobody files a change request, so nobody adds a row to the survey. Under the OCC framework, the bank stays accountable for that AI even when a vendor supplies it.

How to catch it: read the last 12 months of release notes for your top 20 SaaS vendors, and match every AI feature you find against a row in your inventory.

### Apps get built outside IT

Analysts and business teams build AI applications with low-code tools, notebooks and coding assistants, and they connect them to real data. The apps work, so they stick around. This is the pattern behind our post "Goldman Took the Tool Away. The Apps Are Still Running."

How to catch it: look at OAuth grants, API keys issued to model providers and network traffic to model endpoints. Each one points to an application that may not have an inventory row.

### Agents gain permissions

An agent launches with read-only access. Then someone adds a connector, a write scope, a service account, or runs it on their own login to get something done. The inventory row still says read-only. The [joint warning from CISA, the NSA and allied agencies](https://techinformed.com/?p=38489) calls this privilege risk and asks for agent privileges to be restricted, identities verified and human approval required for high-impact actions.

How to catch it: compare the permissions your inventory lists with the tokens and scopes each agent actually holds.

## What a survey records and what a platform records

The table below sets a quarterly survey next to a platform that writes the record as the AI runs.

| Field | Survey spreadsheet | Platform-produced |
| --- | --- | --- |
| Source | What people remember | What actually runs |
| Freshness | As of the last survey | As of the last run |
| Data and permissions | Self-reported | Read from the runtime and identity layer |
| Changes | Found next quarter | Recorded when they happen |
| Examiner answer | A stated belief | Here is the record |

Be clear about the limit. A platform can only write rows for the AI that runs on it. Vendor features and applications built elsewhere still need the discovery methods above. The goal is to bring them onto the platform, or at least into the record.

Some fields stay human: the purpose, the reasoning behind a risk tier, and the owner's sign-off. The platform fills the fields people tend to get wrong, which are data reach, permissions, model version, last change and the actions taken.

One NAIC detail is worth knowing if you are an insurer. The supplement's instructions let a company submit an existing AI inventory in place of Exhibit A, so a working system of record can be exported wholesale. [Coverager's reporting on the playbook](https://coverager.com/the-naic-is-building-an-ai-evaluation-playbook-claims-leaders-should-know-what-it-will-ask/) sets out the questions claims leaders should expect, and [Lumenova's readiness guide](https://www.lumenova.ai/blog/naic-ai-systems-evaluation-tool/) covers what a strong answer looks like and which evidence to have ready.

## Three spot checks to run this week

You do not need a big project to find out where you stand. Each of these takes an afternoon or less:

* Pick five SaaS vendors and check their release notes for AI features that are not in your inventory.
* Pull the OAuth grants and API keys issued to model providers, and count the ones with no inventory row.
* Pick three agents and compare the permissions your inventory lists with the scopes they actually hold.

Whatever you find is your drift rate. It is roughly what an examiner would find too, so it is much better to learn it first.

## Where the record should come from

This is the problem we set out to solve at Shakudo. Anything that runs on [Kaji](/kaji) runs inside your own environment and registers as it runs, with its owner, its data reach, its permissions and an action log. Shakudo's [AI Gateway](/product/ai-gateway) holds the model routing, the permissions and the token budgets, so that side of the estate lands in the same record as the agents. The inventory never leaves your perimeter.

It does not replace the judgment calls. Someone still has to decide the purpose, set the tier and sign off. It does mean the facts in the record come from the system rather than from a survey.

If you want to start today, [request the free inventory template](/contact-us). It carries every field above, the OCC, NAIC and CISA mapping columns and one worked example row.

If you would rather have help, [book a private AI inventory](/contact-us) with us. It runs inside your environment, and your data stays there.

Which row would an examiner find that you did not?

## FAQ

### What is an AI inventory?

An AI inventory is a current record of every AI model, agent and AI-built application an organization runs or buys, with a named owner, its purpose, the data it can reach, the actions it is allowed to take and its risk tier.

### What should an AI inventory include?

Name and type, owner, purpose, autonomy, the data it can reach, its permissions and identity, model and vendor details, dependencies, risk tier, oversight and shutdown, key dates and evidence. That is the field list in the table above, and we have written separately about [who can stop a running agent](/blog/ai-kill-switch).

### What is the difference between an AI inventory and a model inventory?

A model inventory lists models for model-risk review. An AI inventory also covers agents, AI-built applications and vendor AI features, and it records what each one can reach and what it is allowed to do.

### Does OCC Bulletin 2026-54 require an AI inventory?

It is an examiner framework, and it states that it does not set supervisory expectations for banks. Examiners using it tier each use case by consequence and reversibility, which assumes you can list them, vendor features included. We went through the tiers in [our post on the framework](/blog/sr-26-2-occ-2026-54).

### What does the NAIC ask insurers about their AI inventory?

Exhibit A of the AI Risk Evaluation Supplement asks for a model inventory scoped by a materiality threshold: model, use, use case, program area, inherent risk, consumer impact and financial impact. [NAIC's own AI page](https://content.naic.org/insurance-topics/artificial-intelligence) shows the working group still discussing the draft, and the exhibit remains under consideration until the November 16 to 17 meeting.

### How often should an AI inventory be updated?

Whenever something changes: a new model, a vendor feature switched on, a new data source or a new permission. A quarterly survey misses everything between the quarters, which is why platform-produced inventories hold up better.

### How do you find AI that is not in your inventory?

Check vendor release notes, OAuth grants and API keys issued to model providers, network traffic to model endpoints, and low-code or notebook platforms. Then compare each agent's actual permissions with what the inventory lists.