Step zero: seeing what you have

Lesson 1 of 5 in AI Inventory, Use-Case Intake, and Risk Triage.

Every obligation you will ever operationalise — assessments, documentation, monitoring, incident reporting — attaches to a system. Which means every one of them silently assumes you can produce the list of systems. That list is the AI inventory, and building it is step zero because a governance program without one is regulating a rumor.

The regulators assume it too. An EU deployer cannot meet its deployer duties for systems it does not know it deploys; Colorado’s AI Act expects deployers of high-risk systems to know they have them; a US federal agency must publish its AI use-case inventory. And when the first question in any audit, incident, or lawsuit is "what AI do you operate and who owns it?", the inventory is either your answer or your admission.

Key terms: AI inventory, shadow AI, embedded ai, byo ai, attestation

The hard part is that most of the estate is invisible from headquarters. Think of it as an iceberg: the AI you decided to have floats above the waterline, and everything below it arrived without anyone deciding anything.

The shadow-AI iceberg: what an inventory must reach

  1. Sanctioned AI — Above the waterline — the part you already know

    Systems the organisation deliberately built or bought: the credit model, the demand forecaster, the approved chatbot. Usually 20–40% of what a first serious inventory eventually finds. If your registry only contains these, it is a catalogue of intentions, not an inventory.

  2. Embedded SaaS AI — Just below — AI that arrived inside software you already had

    The embedded-AI problem: vendors ship AI features into products you licensed years ago — lead scoring appears in the CRM, transcript summarisation in the meeting tool, "smart" screening in the HR suite — with no procurement event, no contract change, sometimes enabled by default. Your exposure changed and no process fired.

  3. Employee BYO-AI — Deep water — individual accounts, personal workflows

    Employees using consumer AI tools on work content: drafting contracts in a free chatbot, debugging proprietary code, pasting customer data into a summariser. Invisible to procurement and, on personal accounts, largely invisible to IT. Prohibition without a sanctioned alternative does not stop it — it just removes your visibility.

No single technique reaches the whole iceberg, so working programs run four in parallel and reconcile the results — every mismatch between two sources is a discovery.

Surveys and attestations — ask, formally and repeatedly

Structured questionnaires to business-unit leaders: what AI do you build, buy, or use? Cheap, reaches judgment-dependent cases machines miss, and — made an annual signed attestation — creates accountability: a leader who attested "no AI" while running an unregistered screening tool now owns that answer. Weakness: people cannot report what they do not recognise as AI, which is why surveys alone always undercount.

Network and CASB telemetry — watch the traffic

Cloud-access security broker logs, DNS records, and firewall data reveal which AI services the organisation actually talks to: the sanctioned copilot, and the forty-one other AI SaaS domains nobody has heard of. This is how BYO-AI surfaces. Weakness: blind to on-premise models and personal devices, and it shows that a tool is used, not for what — telemetry starts a conversation, not a registry record.

Procurement and expense records — follow the money

Contract databases and expense reports, searched for AI vendors and telltale line items (API credits, GPU instances, "premium" tiers of known AI tools). Catches shadow purchases that bypassed IT and flags renewals where a vendor added AI features mid-contract. Weakness: free tiers cost nothing and appear nowhere.

Code, pipeline, and platform scans — inspect what runs

Scan repositories for ML frameworks and AI-service SDKs; pull the deployment list from MLOps model registries; inventory cloud AI-service usage from the provider console. The most precise source for in-house AI — and the natural place to keep the technical half of the registry synchronised automatically. Weakness: says nothing about vendor tools or the business context that determines risk.

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.