How to Build an AI Inventory (Free AI Register Template)

You cannot govern what you have not counted. Most organisations starting AI governance reach for a policy first. The teams that succeed start one step earlier: with an inventory. A policy written before you know what AI you actually run governs an imaginary organisation.
The pattern we see in Australian organisations is consistent. IT holds a list of service accounts. Marketing holds a list of tools they pay for. Nobody holds a list of what those tools actually do, and nobody owns the register. And as practitioners in governance forums keep warning: asking managers, IT and security to list what they know about will only ever cover what people remember. It says nothing about what is actually turned on.
This article shows you how to build an AI inventory that captures shadow AI and the AI features hiding inside your SaaS stack, and gives you a free AI register template you can start using today.
What is an AI inventory?
An AI inventory (also called an AI system inventory, AI asset register or AI register) is a single, owned, living list of every AI system and AI-enabled feature in use across your organisation: what it is, who owns it, what it is for, what data it can reach, and what access it holds.
For an inventory to stay useful it needs two things: a named owner and a review date. A spreadsheet nobody maintains goes stale within months and starts describing an environment that no longer exists.
What should an AI system register include?
An AI system register needs seven fields per entry: the system or feature name, the vendor, the owner, the purpose, the data it can reach, the access it holds and when that access expires, and the date it was last reviewed.
That is deliberately short. Registers rarely fail for lack of sophistication; they fail because the template is so complex that nobody finishes it. Seven columns you can complete in a fifteen-minute conversation with the system's owner will see more use than a thirty-column template that gets abandoned.
What each field is doing:
- System or feature – the name as people in the business know it. "AI chat widget on the public website" tells everyone what you mean; a project code tells nothing.
- Vendor – who supplies it, including "existing licence" where it is a feature of a platform you already pay for. This is where embedded AI surfaces.
- Owner – one named person accountable for this entry, for example your IT manager. "The team" or "IT" as a whole cannot be chased when a field is missing.
- Purpose – what it is for, in one line. If nobody can write this line, treat that as a finding.
- Data it can reach – be specific. "Entire CRM: accounts, contacts, opportunities, notes" tells an assessor what is at stake; if you write "customer data" it gives no clarity on what is actually exposed.
- Access it holds, and when it expires – the credentials and permissions the system operates with: OAuth tokens, API keys, whose permissions it runs under, admin consoles. Include expiry. If the access has no expiry, record that and treat it as a gap to close.
- Last reviewed – the date the entry's owner last confirmed it is still accurate, for example "12 Nov 2026".
The template
Copy the table below into a spreadsheet. We have filled four rows with worked examples so you can see what a good entry looks like; replace them with your own.
| System or Feature | Vendor | Owner | Purpose | Data it can reach | Access it holds, and when it expires | Last reviewed |
|---|---|---|---|---|---|---|
| AI chat widget on the public website | Conversational marketing platform | [name] | Qualify inbound leads, write to CRM | Entire CRM: accounts, contacts, opportunities, notes | OAuth token into the CRM tenant. No expiry | [date] |
| AI agent in the CRM summarising leads | CRM platform, existing licence | [name] | Summarise and triage inbound leads | Every lead record. Can make outbound web requests | Runs with the permissions of whoever asks it | [date] |
| AI coding agent with platform API access | Code editor plus hosting platform | [name] | Build and deploy application changes | Production database, staging, backups | API token. All operations including delete, account-wide. No expiry | [date] |
| AI recruitment chatbot on careers site | HR technology provider, vendor-hosted | [name] | Collect applications, schedule interviews | Every applicant record: names, contacts, addresses, transcripts | Vendor-hosted. Admin console on the public internet | [date] |
How do you find AI features hidden in SaaS tools?
You find hidden AI by combining four discovery channels. Rely on any one of them alone and it becomes almost impossible to see the whole picture:
- Ask, but do not stop there. Interview the people closest to the work: team managers, IT administrators, security staff, and whoever signs off software spend. This is fast and surfaces intent, but it only covers what people remember, not what is actually turned on.
- Follow the money. Pull the last twelve months of software spend from finance: invoices, credit card statements, expense claims, procurement records. Any line item from an AI vendor is an entry. Any platform line item (CRM, HR platform, helpdesk, document suite) is a candidate for embedded AI features.
- Follow the identity. Review your SSO or identity provider's app catalogue and OAuth grant logs. Every AI tool staff connected with "Sign in with Google" or a Microsoft work account shows up there, whether or not anyone approved it.
- Ask the vendors. For each platform you already run, check the vendor's release notes or ask their account team directly: which AI features exist, are they on by default, and where does the data go. Vendors ship AI features into existing licences constantly; the feature you never enabled may have been enabled for you.
Who should own the AI inventory?
One named person owns the AI inventory. In most organisations that is the GRC lead, the security lead, or the first AI governance hire; in a small organisation it is whoever owns IT risk. If you do not have a dedicated security person, this is a natural fit for a virtual CISO engagement, where an experienced security leader owns the register and the cadence around it.
Ownership means three things: this person chases the answers when an entry is incomplete, runs the review cadence, and has the authority to say "this tool is not in the register, so it is not approved for use". The register is a control only if someone can enforce it.
Contributors can be distributed – each entry's owner keeps their own row accurate – but accountability for the register itself cannot be. "Everyone owns it" is how registers die.
Is an AI inventory required for ISO 42001?
Yes, in practice. ISO/IEC 42001 expects you to know which AI systems you operate before you can govern them: the management system clauses assume you can define the scope of your AI activities and assign accountability for each system, and neither is possible without an inventory. Auditors ask for it early, and building the AIMS on a guessed scope is wasted work.
The same logic applies to the EU AI Act: providers of high-risk AI systems listed in Annex III must register themselves and those systems in the EU database before placing them on the market or putting them into service (Article 49(1)), and deployer obligations apply only to high-risk systems, so deployers must determine which of their systems fall into scope. Both duties presuppose an inventory. If you use AI rather than supply it into the EU, the Act is mostly context for now; the inventory is what you will need when a customer, insurer or regulator asks.
Our ISO 42001 blueprint covers where the inventory sits in a full AI management system.
Keeping it alive
Registers fail quietly because nothing breaks when they drift. Nobody gets an alert when a new SaaS feature goes unlisted, and nobody gets blamed for an out-of-date row. The decay only surfaces later, at an audit or an incident, when the document that was supposed to give you visibility turns out to describe last quarter. Three habits keep it alive:
- Review cadence. Review every entry quarterly, and the whole register at least annually. Put the dates in the calendar now.
- Intake path. Any new AI tool or feature goes through the register before it goes into use. If the path is slower than the workaround, people route around it.
- Exit path. When a tool is retired, the entry is closed and its access revoked. An entry for a cancelled tool whose OAuth token still works is exactly the kind of thing the register exists to catch.
Rollout checklist
- Appoint the register's owner by name.
- Stand up the template above; spreadsheet is fine to start.
- Run the four-channel discovery: interviews, spend, identity logs, vendor checks.
- Record every AI system and AI-enabled feature found, including unapproved ones. The point is visibility.
- Fill all seven fields per entry; where a field is unknown, record it as unknown and assign someone to close it.
- Set the review cadence and the intake and exit paths.
- Present the register to leadership. The findings usually argue for themselves.
Further reading
- AI Incident Response Plan: When Agents Go Rogue – what happens when a system in your register misbehaves
- AI Acceptable Use Policy Template for Australia – the staff-facing policy that sits next to the register
- ISO 42001: A Blueprint for Responsible AI Implementation – the full AI management system standard
- ISO 42001 Readiness Assessment – check where your AI governance stands
Want a hand standing up your AI register?
CyberNinja helps Australian and New Zealand organisations build practical AI governance, starting with visibility: inventories, discovery exercises and ISO 42001 implementation. Get in touch if you would like to talk through your situation.
This article is general guidance, not legal advice, and reflects the position as at November 2026. Confirm your obligations against the current legislation and, where the stakes warrant it, with your legal advisers.
Sources
Need help with AI governance?
We help organisations build practical AI governance: policies, AI inventories, risk assessments and ISO 42001 implementation.