Building an AI System Inventory: The Register Every Governance Program Starts With

An AI system inventory is one register listing every AI tool your organization uses, who owns it, what data it touches, and what decisions it influences. Build it before anything else, because every framework worth following assumes you already have one. ISO/IEC 42001 asks you to scope your management system. The NIST AI Risk Management Framework asks you to map context. The EU AI Act asks you to classify systems by risk. All three questions are unanswerable if you cannot list the systems.

Most Canadian companies cannot. Not because they are careless, but because AI arrived through the side door: a marketing contractor's copywriting tool, a support lead's ticket summarizer, a developer's coding assistant, a sales rep's meeting notetaker. Nobody filed a request. There was nothing to approve.

Why does the inventory have to come first?

Because governance is a set of decisions about specific systems, and you cannot decide about a system you have not named.

Think about what a regulator, an auditor, or an enterprise buyer actually asks. Which AI systems process personal information? Which ones make or influence decisions about people? Where does the data go? Who signed off? Each of those is a filter applied to a list. Without the list, you are guessing, and a guess written into an audit response is a finding waiting to happen.

There is a second reason, less obvious and more useful. The act of building the inventory is the cheapest risk assessment you will ever run. Every time I have sat with a team through this exercise, something surfaced that nobody in the room knew about. A customer support tool quietly summarizing tickets that contain health information. A resume screener a hiring manager trialled and never turned off. The inventory pays for itself in the first week, before a single control is written.

What goes in the register?

Keep it small enough that people actually fill it in. Twelve fields is usually the ceiling before a register rots.

The decision-influence field is the one people skip, and it is the one that matters most under Canadian law. PIPEDA's proposed successor in Bill C-27 and Quebec's Law 25 both attach specific obligations to automated decisions about individuals. If your register cannot answer that question, you cannot answer theirs. Our Bill C-27 tracker follows where that legislation stands.

How do you find the systems nobody told you about?

Four passes, in this order, cheapest first.

Start with money. Pull twelve months of corporate card and expense data and search for the vendors. Anything with a subscription leaves a financial trail. This single pass typically surfaces half of what you are missing.

Then look at identity. Your single sign-on logs and Google or Microsoft admin console show which third-party applications employees have connected to their work accounts. OAuth grants are a confession.

Then ask, properly. A two-question form beats a policy memo: what AI tools do you use for work, and what do you paste into them? Ask without consequences attached, or you will get silence. The goal is an accurate list, not a disciplinary file.

Then look at the network. Egress logs or a browser extension inventory catch the browser-based tools that never touched a credit card or an SSO grant. This is the pass that finds free-tier usage.

Run all four. Each catches something the others miss, and the free tools are usually the riskiest, because nobody read the terms.

Who should own each entry?

A named individual in the business, not IT and not legal.

This is the part organizations get wrong most often. IT owns the platform, legal owns the interpretation, but neither can tell you whether the sales team should be running customer notes through a summarizer. The person who chose the tool and depends on the output is the person who has to answer for it. IT and legal advise; the business owner decides and signs.

Give that owner three obligations, and keep the list to three: keep the entry accurate, flag material changes in how the tool is used, and confirm the entry every quarter. A vendor inventory that nobody is accountable for stops matching reality within a few months.

How often should it be reviewed?

Quarterly for confirmation, immediately for change.

The quarterly cycle is a light touch: each owner confirms their entries are still accurate, and anything unconfirmed for two cycles gets escalated or retired. The immediate trigger matters more. New system, new data category, new decision the tool influences, vendor changes its model provider, or the tool moves from pilot to production. Any of those means the entry is re-reviewed now, not in eleven weeks.

Tie the register to your assessment workflow so that a new high-tier entry automatically opens a privacy impact assessment rather than sitting in a queue. The gap between "we added a system" and "we assessed it" is where most programs quietly fail.

The mistakes that ruin a register

Five, in rough order of how often I see them.

• Building it in a spreadsheet nobody owns. Spreadsheets are fine to start. They stop being fine when the person who made it changes roles.
• Recording the platform instead of the use. "Microsoft Copilot" is not one entry if three departments use it for three different jobs with three different data sets.
• Leaving out the tools embedded in software you already bought. Your CRM, your help desk, and your HR system all shipped AI features last year. They count.
• Treating pilots as out of scope. Pilots run on real data more often than anyone admits.
• Perfectionism. A register covering eighty percent of your systems, in use and reviewed, beats a complete one that took nine months and arrived dead.

Start with a list of names this week. Add fields as you go. The register earns its value from being current, not from being elegant.

If you want the wider programme context around the inventory, ISO 42001 vs SOC 2 covers which framework the register ends up feeding.

This is general information, not legal advice.

Valdra keeps the AI system inventory, the risk tiering, and the assessments in one place, so a new entry opens the right assessment automatically instead of waiting for someone to notice. You can see how the AI governance side works without talking to anyone first.