TL;DR: An AI use-case register is the authoritative inventory of where an organisation uses AI: purpose, owner, users and affected people, data, models and vendors, actions and autonomy, exposure to untrusted input, regulatory category, risk tier, controls applied, residual risk and review date. Routing applies a short set of deciding questions at intake to assign a tier, and each tier carries required controls and an approver. Keep it accurate by making registration a prerequisite for access to models, data and deployment, and by triggering re-assessment on material change.
How to approach it
Define the register's fields and purpose, then the routing logic with tiers, then how to keep it accurate in a real organisation. Emphasise speed for low-risk cases so teams use it.
A strong answer
Intake
purpose, owner, data
Routing questions
data, actions, users
Risk tier
and regulatory class
Controls
per tier
Review date
residual risk
Routing questions
Personal or regulated data?
- Privacy review
Acts or only advises?
- Action controls
Untrusted input?
- Injection review
Fields.
| Field group | Examples |
|---|---|
| Identity | Name, description, owner, team, status |
| Purpose and impact | Intended use, decisions affected, users, affected people |
| Data | Categories (personal, confidential, regulated), sources, retention |
| Technology | Models and versions, vendors, data processing terms, hosting |
| Capability | Tools and actions, reversibility, autonomy level |
| Exposure | Untrusted inputs (email, web, uploads, third-party tools) |
| Regulation | EU AI Act tier and role, sector rules |
| Risk | Tier, assessments done, controls, residual risk, approver, review date |
Routing. At intake, deciding questions (impact on people, data sensitivity, actions and autonomy, untrusted input, model source, regulated category, scale) produce a tier:
- Low: internal, assistive, no sensitive data or actions. Self-assessment against approved patterns; fast approval.
- Medium: customer-facing, sensitive data, or limited actions. Threat model, security review, evaluations, logging, owner sign-off.
- High: regulated or consequential uses, autonomous actions on untrusted input. Red team, privacy and legal review, oversight design, impact assessment, senior approval, monitoring.
Keeping it accurate.
- Registration is required to get API keys for approved model providers, access to data platforms, or a deployment slot.
- Discovery for shadow AI: network and expense signals, SaaS reviews, surveys.
- Change triggers: new model, new data source, new tool, new user population, new purpose.
- Periodic review dates per tier.
- Automate fields from existing systems (model registry, cloud accounts, procurement).
Why it matters. It is the basis for NIST AI RMF Map, ISO/IEC 42001 scope and risk assessment, EU AI Act classification, incident response (which systems use the affected model?) and audit.
What interviewers probe next
- "Teams won't register things." Make it fast for low-risk uses, tie it to access to models and data, and show value (approved patterns, faster security review).
- "Who maintains it?" A governance function owns the process; each system owner owns their entry's accuracy.
- "How does it help in an incident?" Query which systems use a compromised model, vendor or data source.
What if?
What if · interview follow-ups
Hold your answer as the constraints change
A company has no idea how many AI tools and features are in use across departments, and the CISO wants a register within a quarter.
Start here
What should each register entry contain?
What would you need to know to decide the review level?
Purpose, owner, users and affected people, data used, models and vendors, actions and autonomy, exposure to untrusted input, regulatory category, risk tier, controls applied, residual risk and review date. These fields let you route reviews and answer regulators and customers.
Carry it forwardEntries capture purpose, data, autonomy, exposure and tier.
Weak answerRecording only tool names and vendors.
Constraint changes
How do you find what already exists?
Where does AI use leave traces?
Combine sources: procurement and expense records, SaaS discovery, network and API gateway logs showing model provider traffic, code searches for model SDKs, and a simple survey. Expect to find more than anyone thought, and prioritise by data sensitivity and exposure.
Carry it forwardDiscover from spend, network, code and surveys.
Weak answerRelying on self-reporting alone.
Constraint changes
How would routing to a tier work?
Which few questions separate low from high risk?
A short set of deciding questions at intake: does it touch personal or confidential data, does it take actions, does it read untrusted input, is it customer-facing, is it in a regulated use? The answers assign a tier, and each tier carries required controls and an approver.
Carry it forwardA few deciding questions assign tiers and controls.
Weak answerSending every entry through the same full review.
Constraint changes
How do you keep the register accurate?
What makes teams register willingly?
Make registration a prerequisite for access to approved models, data and deployment pipelines, so unregistered systems cannot get what they need. Trigger re-assessment on material change, such as new data, tools or users, and expire entries without review.
Carry it forwardTie registration to access; reassess on change.
Weak answerAn annual spreadsheet refresh.
Constraint changes
What if a team refuses to register a small internal tool?
What does the team get from registering?
Make low-risk registration take minutes and give something back, such as access to the approved gateway with enterprise data terms. If the tool sends sensitive data to an unapproved service, it is a policy issue to escalate, but most resistance disappears when registration is fast.
Carry it forwardMake registration cheap and useful.
Weak answerMaking registration a heavy form for every tool.
Now defend it
How would you set up the register in a quarter?
Define the fields, run discovery from spend, network, code and surveys, register the most sensitive uses first, apply deciding questions to assign tiers with required controls and approvers, make registration a prerequisite for models, data and deployment, keep low-risk registration fast, and trigger reassessment on material change.
How it's scored
Weak: "A spreadsheet of AI projects."
Hire: Defines fields and tier routing with required controls.
Strong Hire: Adds enforcement through access gating, shadow AI discovery, change triggers, automation, and incident uses.
Common mistakes
- Same heavy review for every use.
- Register disconnected from actual access and deployment.
- No re-assessment when capabilities change.