AI inventory and ownership

Finding every AI system, describing and classifying it, assigning a risk tier and a named owner, and tracking the models inside.

AI governance

Also called responsible AI, AI risk management

What: How an organisation decides which AI it builds or uses, who owns each system, which risks it accepts and what evidence shows its controls work.
Why: Laws, customers and incidents all demand it, and it is cheapest when it shapes systems early rather than auditing them late.

Watch out

Treating it as ethics statements or paperwork after launch rather than decisions and evidence throughout.

AI system

Also called AI system definition

What: In the EU AI Act, following the OECD: a machine-based system that works with some autonomy and infers from inputs how to produce outputs such as predictions, content, recommendations or decisions.
Why: The definition decides what goes in the inventory and what the law covers. Fixed rules written entirely by people usually fall outside it.

Watch out

Counting only in-house models and missing AI inside bought software and staff tools.

AI inventory

Also called AI register, AI system register

What: A maintained list of every AI system an organisation builds, buys or uses, with purpose, owner, data, vendor, affected people, legal role, risk tier and review dates.
Why: You cannot govern what you cannot see. Every later control, review and audit starts from an inventory entry.
Example: Fields: name, purpose, owner, vendor, data categories, people affected, AI Act role and tier, internal tier, last review, open risks.

Watch out

A fifty-column spreadsheet so heavy that teams stop filling it in.

Shadow AI

Also called unapproved AI, bring your own AI

What: AI that staff use without the organisation knowing or approving: public chatbots, personal API keys in scripts, AI browser extensions.
Why: It is where customer data most often leaks to external models. Find it through surveys, expense claims, network logs and code scans.

Watch out

A blanket ban that drives use further underground instead of offering approved tools.

System classification

Also called AI use case classification

What: Describing an AI system by purpose, users, affected people, decisions supported, data, model and deployment, then labelling its type and legal category.
Why: Classification is the input to risk tiering, and the written reasoning is what reviewers and auditors check.

Watch out

Classifying from the vendor's marketing description instead of how the system is actually used.

Risk tiering

Also called internal risk tier, risk rating

What: Assigning each system an internal tier, often low, medium or high, from its impact on people and how much a human checks its output, alongside its legal tier.
Why: The tier sets how much review, testing and monitoring a system gets, so effort goes where harm is likeliest.

Watch out

Tiering by model size ("it is just a small model") rather than impact.

System owner

Also called accountable owner, business owner, technical owner

What: The named person accountable for an AI system's behaviour, changes and residual risk, usually supported by a technical owner who runs it day to day.
Why: Warnings, incidents and reviews need someone with authority to act. Without a named owner they go nowhere.

Watch out

Writing a team or "IT" as owner, so nobody personally has to act.

Provider and deployer

Also called AI Act roles, importer, distributor

What: Under the EU AI Act, the provider develops a system or model and places it on the market under its name; the deployer uses it under its authority. Importers and distributors have their own duties.
Why: Duties differ sharply by role, so a mapping starts by naming yours. Modifying a system substantially can turn a deployer into a provider.

Watch out

Assuming that buying an AI tool moves all responsibility to the vendor.

Third-party AI

Also called vendor AI, AI supply chain

What: AI that arrives through vendors: API models, AI features inside SaaS products and open models you download.
Why: Most organisations now get most of their AI this way. Vendor questionnaires, contracts and testing in your own context are the controls.

Watch out

Accepting a vendor's certificate without testing the system on your own users and data.

Model registry

Also called MLflow Model Registry, model catalogue

What: A store of model versions with training data reference, code version, parameters, evaluation results, approver and stage (staging, production, archived).
Why: It answers "which model made this decision last March?", which incidents and audits ask. For API models, record provider, model name, version and date checked.

Watch out

A registry only data scientists can read, never linked to the inventory.

Go deeper