You cannot govern AI systems you do not know about. Here is what to record, how to find the systems nobody registered, and how to keep the list current.
Every AI governance program eventually asks the same first question: what AI are we actually running? Most companies cannot answer it. There are features in production, prototypes that quietly became production, vendor tools with AI bolted on, and scripts that call a model API with a key someone pasted in last spring.
An inventory is the least exciting part of AI governance and the part everything else depends on. You cannot assess risk, answer a regulator, or respond to a vendor incident for systems you do not know exist.
What counts as an AI system
Start broad, then narrow. A useful working definition is anything where a machine learning model's output affects what a user sees, what a customer gets, or what an employee decides. That includes:
- Features you built that call a model, whether hosted by a vendor or by you.
- Internal tools, such as a support assistant or a code review bot.
- Vendor products with AI features turned on, such as a CRM that scores leads or an HR tool that screens CVs.
- Classic machine learning: fraud scores, recommendations, forecasts.
People will argue about edge cases. Do not let that hold up the list. Include anything borderline and mark it as such.
What to record
Keep each entry short enough that people will actually fill it in. These fields cover most of what risk teams and regulators ask about:
- Name and owner. A named person, not a team alias. Someone has to answer questions.
- Purpose. One or two sentences on what it does and who uses it.
- Who it affects. Customers, employees, the public. And whether it influences decisions about them.
- Models and vendors. Which models, which providers, which versions if you pin them.
- Data. What goes in, especially personal or confidential data, and where outputs are stored.
- Risk tier. Your internal rating, and any regulatory classification, such as the EU AI Act's high-risk categories.
- Controls. Evaluation suite, monitoring, human review, how to turn it off.
- Status. Prototype, pilot, production, retired.
Resist adding thirty more fields. A short inventory that is accurate beats a long one that is half empty and two quarters out of date.
Finding what is already out there
Asking teams to self-report gets you maybe half the picture. The rest you find by looking at traffic and money.
Network and API traffic. If you have an AI gateway, its logs are the best source you will get. If you do not, egress logs and DNS queries to model provider domains show who is calling what.
API keys and accounts. Check which teams have accounts with model providers and cloud AI services. Personal accounts on company cards are common and worth finding.
Procurement and expenses. Vendor contracts and expense reports show AI products the engineering org has never heard of.
Code search. Search your repositories for model provider SDKs and API hostnames. It finds the prototype that became a cron job.
Vendor questionnaires. Ask existing SaaS vendors whether they have added AI features and whether those are on for your account. Many turned them on by default.
Expect the first pass to be messy. That is normal. The point is to get a baseline you can improve.
Keeping it current
An inventory that is only updated during audits is wrong most of the time. Tie it to things that already happen:
- Make registration part of launch. A system cannot get a production API key, or through launch review, without an inventory entry.
- Issue keys through a central path. If model access goes through a gateway or a central account, every new consumer has to register to get credentials. The inventory then updates itself.
- Review on a schedule. Once a quarter, owners confirm their entries are still right. Unconfirmed entries get flagged.
- Retire things explicitly. When a system is switched off, mark it retired rather than deleting it. You may be asked about it later.
Where to keep it
A spreadsheet is fine to start. What matters is that it has one owner, everyone knows where it is, and it is linked from your launch checklist. As it grows, move it into whatever your company already uses as a service catalog, so AI systems sit next to their services, owners and on-call rotas rather than in a separate governance tool nobody opens.
Once you have it, everything else gets easier. The risk assessment has a list to work through. The incident response team knows who to call. And when the next regulation or vendor breach lands, you can answer "are we affected?" in an afternoon instead of a month.
Share this article
David Childs
Consulting Systems Engineer with over 10 years of experience building scalable infrastructure and helping organizations optimize their technology stack.