Skip to content

Use Case Definition

What makes a good AI use case definition from a security and governance perspective — and how to translate one into a risk profile, control set, and operating model.

Part of From Strategy to Production


Why Use Case Definition Matters

Every AI system starts as a use case. "We want AI to help with customer queries." "We want AI to analyse contracts." "We want AI to detect fraud."

These are not use cases. They're aspirations. A use case — one that security, governance, and risk functions can actually work with — is something far more specific. It defines not just what the AI does, but what data it touches, who sees its outputs, what decisions it influences, what happens when it's wrong, and who is accountable.

The difference matters because the framework's entire control model flows from the use case. The risk tier determines which controls apply. The risk tier is determined by the use case. A vague use case produces an uncertain risk tier, which produces either too many controls (killing the business case) or too few (creating unmanaged risk).

A well-defined use case is the single most valuable input to the entire framework.


What Makes a Good Use Case Definition

The Minimum Viable Use Case

A use case definition that security and governance can work with answers ten questions. These aren't optional fields on a form — each one directly determines a control requirement.

Use Case Definition Model

# Question What It Determines Framework Connection
1 What does it do? Scope boundaries, guardrail topic rules Controls — guardrail configuration
2 What decisions does it make or influence? Decision authority dimension → risk tier Risk Tiers — authority scoring
3 What data does it access? Data sensitivity dimension → risk tier, DLP rules Data Protection
4 Who are the users? Audience dimension → risk tier, access controls Identity & Access
5 What happens when it's wrong? Reversibility dimension → risk tier, PACE plan PACE Resilience
6 What's the expected volume? Operational sizing — HITL staffing, Judge compute, logging storage Governance Model
7 What regulatory context applies? Regulatory dimension → risk tier, compliance controls Regulatory Extensions
8 What tools or actions can it take? Agentic scope → agentic controls, execution boundaries Agentic Controls
9 Where does it sit in the business process? Integration points, upstream/downstream dependencies Threat Model Template
10 Who is accountable for its outputs? Governance ownership, HITL escalation, incident response Governance Operating Model

Why Each Question Matters

Q1: What does it do?

Not "what could it do" or "what might it do eventually." What does it do in this deployment? Scope creep is the primary reason use cases drift from one risk tier to another without anyone noticing.

Bad Good
"AI customer assistant" "Answers customer questions about account balances, transaction history, and product information via chat. Cannot transfer funds, change settings, or provide personalised financial advice."
"Contract analysis tool" "Extracts key dates, parties, and obligation clauses from supplier contracts. Outputs a structured summary for procurement review. Does not recommend accept/reject decisions."
"Fraud detection" "Scores incoming card transactions against fraud patterns. Transactions scoring above 0.85 are held for human review. Transactions below 0.85 proceed automatically. Does not block transactions independently."

The good definitions include explicit negative scope — what the system cannot do. This is critical for guardrail configuration. If the use case says "does not provide financial advice," then the guardrails enforce that boundary. Without that boundary in the definition, nobody knows to configure it.

Q2: What decisions does it make or influence?

The most important question for risk classification. The framework distinguishes between four levels of decision authority:

Level Description Example Typical Tier Impact
Informational Provides information; no decision influence Search results, summaries, FAQ answers LOW–MEDIUM
Advisory Recommends a decision; human decides "Consider reviewing this transaction" MEDIUM–HIGH
Influential Output significantly shapes human decision Risk score that determines investigation priority HIGH
Autonomous AI makes the decision; human monitors Auto-approved claims under £500 HIGH–CRITICAL

Many use cases claim to be "advisory" when they're functionally "influential." If the human reviewer accepts the AI recommendation 95% of the time without independent analysis, the system is influential regardless of what the process document says. Be honest about actual decision authority, not theoretical.

Q3: What data does it access?

Not just what data it needs — what data it can access. If the system has a database connection with broader permissions than the use case requires, the risk profile reflects the access, not the intent.

Data Category Examples Risk Impact
Public Product info, published policies, public content Minimal
Internal Internal docs, meeting notes, project data Low–Medium
Confidential Strategy docs, financial forecasts, trade secrets Medium–High
PII Names, addresses, contact details High (GDPR, privacy)
Sensitive PII Health data, biometrics, financial records High–Critical
Regulated Credit data, medical records, legal documents Critical (sector regulation)

Q4: Who are the users?

User Type Risk Implication
Internal — technical Lowest risk; users understand AI limitations
Internal — non-technical Low–medium; may overtrust AI outputs
Partners/suppliers Medium; less control over use, contractual implications
Customers — authenticated High; reputational risk, regulatory exposure
Customers — unauthenticated High; abuse potential, no user accountability
General public Highest; maximum blast radius, maximum regulatory exposure

Q5: What happens when it's wrong?

Consequence Example Reversibility
Inconvenience Wrong FAQ answer; user asks again Fully reversible
Wasted effort Bad document summary; human redoes it Time lost, no harm
Financial loss Incorrect pricing shown to customer Recoverable with cost
Reputational damage Offensive response to customer Difficult to reverse
Regulatory breach PII disclosed in response Notification required
Physical harm Wrong medical triage priority Irreversible

Q6: Volume determines operational cost — see Business Alignment for the FTE calculation. A CRITICAL-tier system at 100 interactions/day is operationally different from one at 100,000 interactions/day, even though the control requirements are identical.

Q7: Regulatory context is often the simplest question to answer and the one most often skipped. If the use case operates in financial services, healthcare, legal, HR, or insurance, there are sector-specific requirements that may override the framework's general tiers.

Q8: Tools and actions determine whether agentic controls apply. If the AI can call APIs, write to databases, send emails, or trigger workflows, it's agentic — even if nobody calls it that.

Q9: Business process position determines upstream and downstream dependencies. Where does the AI sit? What feeds it? What consumes its output? This shapes the threat model — attack surfaces are at integration points, not inside the model.

Q10: Accountability is the governance anchor. Someone — named, not "TBD" — owns the outcomes of this AI system. This person is responsible for HITL decisions, incident response, and regulatory engagement. Without named accountability, the governance model has no force.


Translating Use Cases to Risk Profiles

The Six-Dimension Scoring Model

The framework's risk tiers use six dimensions. Each use case answer maps to a dimension score.

Dimension Source Question LOW MEDIUM HIGH CRITICAL
Decision Authority Q2 Informational Advisory Influential Autonomous
Reversibility Q5 Fully reversible Recoverable with effort Difficult to reverse Irreversible
Data Sensitivity Q3 Public Internal/confidential PII, financial Regulated, sensitive PII
Audience Q4 Internal technical Internal non-technical External authenticated External unauthenticated/public
Scale Q6 <100/day 100–10,000/day 10,000–100,000/day >100,000/day
Regulatory Q7 Unregulated Light-touch regulation Sector-regulated High-risk under AI Act or equivalent

Scoring Rules

Rule 1: Highest dimension wins. If a system scores MEDIUM on five dimensions but CRITICAL on one (e.g., it processes regulated health data), the system is CRITICAL. Risk tiers are pessimistic by design.

Rule 2: Adjacent dimensions compound. If a system scores HIGH on three or more dimensions, consider upgrading to CRITICAL even if no individual dimension reaches CRITICAL. Three HIGH scores suggest systemic risk that a single CRITICAL would capture.

Rule 3: Modifiers can push up, not down. Use case modifiers (agentic, customer-facing, regulated) can increase the effective tier but cannot decrease it. A system that's HIGH based on data sensitivity doesn't become MEDIUM because it's internal-only.

Can AI Score Risk Tiers?

Yes — with caveats. The six-dimension model is structured enough to be machine-evaluable. An LLM can extract the relevant information from a well-written use case definition and propose a risk tier.

What AI can reliably do:

Capability Reliability Why
Extract data categories from use case description High Pattern matching against known categories
Identify user types High Explicit in most use case definitions
Classify decision authority level Medium–High Requires understanding of "influence" vs. "inform"
Identify regulatory context Medium–High Can match domains to known regulatory frameworks
Propose a risk tier based on scoring Medium Mechanical scoring is straightforward
Identify missing information in the definition High Can check against the ten-question template

What AI cannot reliably do:

Capability Reliability Why
Assess actual reversibility of consequences Low Requires domain knowledge and organisational context
Determine real decision authority (vs. claimed) Low Requires understanding of how humans actually use the system
Evaluate data access scope (vs. stated need) Low Requires infrastructure knowledge
Make the final classification decision Not appropriate Accountability must rest with a human

Practical approach: AI as classification assistant.

Use an LLM to: 1. Parse the use case definition against the ten-question template 2. Flag missing or ambiguous answers 3. Score each dimension based on available information 4. Propose a risk tier with reasoning 5. Identify questions the human classifier should investigate

Then a human risk analyst reviews the proposal, investigates the flagged areas, and makes the final classification decision.

This is an example of the framework's own principle in action: AI proposes. Humans decide. The classification itself is a MEDIUM-tier use case — it influences risk decisions, should be reviewed by a human, and the consequences of misclassification are significant but recoverable.

Example AI classification prompt:

You are an AI risk classification assistant. Given the following use case
definition, score each dimension and propose a risk tier.

USE CASE:
{use_case_definition}

DIMENSIONS TO SCORE (LOW / MEDIUM / HIGH / CRITICAL):
1. Decision Authority: Does the AI decide, influence, advise, or inform?
2. Reversibility: What happens if the AI is wrong? How easily is it fixed?
3. Data Sensitivity: What data categories does it access?
4. Audience: Who are the users?
5. Scale: How many interactions per day?
6. Regulatory: What regulations apply?

For each dimension:
- State the score with a one-line justification
- Flag if the use case definition is ambiguous or missing information

Then:
- Propose an overall risk tier (highest dimension wins)
- List any questions that need human investigation before finalising

The Judge layer should evaluate the quality of AI-generated classifications just as it evaluates any other AI output. Use the same sampling rates — 100% review for the first 50 classifications, then reduce as accuracy data accumulates.


What Security and Governance Can Extract From a Use Case

A well-defined use case produces specific, actionable outputs for every governance function:

Security Outputs

Use Case Input Security Output How
Scope boundaries (Q1) Guardrail topic rules Negative scope → deny rules; positive scope → allow rules
Data categories (Q3) DLP rules, PII detection config Data types → specific detection patterns
User types (Q4) Authentication requirements, access controls Internal → SSO; external → MFA; public → rate limiting
Tools/actions (Q8) Agentic controls, sandbox configuration API list → scoped permissions; write actions → approval controls
Integration points (Q9) Attack surface map, trust boundaries Upstream/downstream → threat model
Error consequences (Q5) PACE plan severity Irreversible → fail-closed; reversible → fail-open acceptable

Governance Outputs

Use Case Input Governance Output How
Risk tier (scored) Control requirements Tier → control matrix
Decision authority (Q2) HITL requirements Autonomous → mandatory review; advisory → sampling
Volume (Q6) HITL staffing model FTE = Volume × Sample Rate × Review Time / Working Hours
Regulatory context (Q7) Compliance controls, reporting Sector → specific regulatory requirements
Accountability (Q10) Governance assignments Named owner → 1st line responsibility
Error consequences (Q5) Incident severity classification Consequence type → incident response tier

Operational Outputs

Use Case Input Operational Output How
Volume (Q6) Infrastructure sizing Interactions/day → compute, storage, logging capacity
Risk tier Judge sampling rate LOW: 1–5%; MEDIUM: 5–10%; HIGH: 20–50%; CRITICAL: 100%
Risk tier Log retention period LOW: 90 days; MEDIUM: 1 year; HIGH: 3 years; CRITICAL: 7 years
Risk tier Review SLA LOW: best effort; MEDIUM: 24hr; HIGH: 4hr; CRITICAL: 1hr
Scope boundaries (Q1) Monitoring baselines Expected behaviour → anomaly detection thresholds

Who Needs to See Use Case Definitions

The RACI Model

Role Responsibility What They Need From the Use Case
Business Owner Accountable for outcomes Clear scope, success criteria, business value
Risk Analyst Classifies risk tier All ten questions answered; enough detail to score dimensions
Security Architect Designs controls Data flows, integration points, tool access, trust boundaries
AI Engineer Builds the system Technical scope, constraints, data access requirements
Governance (AGO) Sets policy, oversees Risk tier, control requirements, staffing needs
HITL Reviewers Review AI outputs What "good" looks like, what "wrong" looks like, escalation criteria
Legal/Compliance Validates regulatory alignment Regulatory context, data categories, decision types
Internal Audit Independent assurance Use case as documented vs. use case as deployed

Format by Audience

Different stakeholders need different views of the same use case:

Audience Format Content Length
Board / governance committee Executive summary Business problem, risk tier, key controls, cost 1 page
Risk function Risk assessment All ten questions, dimension scores, tier justification 2–3 pages
Security / engineering Technical specification Data flows, integration points, control configuration 3–5 pages + diagrams
Operating team Operating playbook What to monitor, when to escalate, PACE procedures 2–3 pages
Audit Control evidence pack Use case definition + control implementation + testing evidence Full documentation set

Who Should Manage Use Cases

The Use Case Registry

Use cases should be managed in a central registry — an inventory that tracks every AI use case from definition through deployment to retirement.

Registry Field Purpose Populated By
Use case ID Unique identifier System-generated
Name Human-readable label Business owner
Status Draft / Under review / Approved / In production / Retired Governance
Risk tier Classified tier Risk analyst (approved by governance)
Business owner Accountable person Business unit
Technical owner Engineering lead Engineering
Last review date When it was last assessed Governance
Next review date When reassessment is due System-scheduled
Change log What's changed since last review All contributors

Ownership Model

Aspect Owner Responsibility
Use case definition Business owner Defines what the AI does and why; updates when scope changes
Risk classification Risk analyst (2nd line) Scores dimensions, proposes tier, documents justification
Classification approval AI governance body Approves or challenges the proposed tier
Control implementation Engineering team Implements controls specified by the tier
Control verification Security / governance Confirms controls are implemented correctly
Ongoing monitoring Technical operations Operates guardrails, Judge, logging; escalates issues
Periodic review Governance Reassesses use case, confirms tier still appropriate
Change management Business owner + governance When the use case changes, triggers reassessment

Review Triggers

Use cases don't stay static. Review is triggered by:

Trigger Why Action
Scheduled review Annual minimum; quarterly for CRITICAL Full reassessment against ten questions
Scope change Use case does something new Reassess affected dimensions; may change tier
Data access change New data sources added Reassess data sensitivity dimension
User population change Internal → external, or new user groups Reassess audience dimension
Incident AI-related incident on this system Post-incident review; reassess controls
Regulatory change New regulation affects this domain Reassess regulatory dimension
Model change Underlying model swapped or upgraded Reassess; model change doesn't change tier but may affect control configuration
Volume change Significant increase in usage Reassess operational sizing; may need more HITL staff

Automating Use Case Discovery and Definition

When Can We Automate?

Automation of use case definition exists on a spectrum:

Automation Level What It Does Maturity Required Accuracy
Template-assisted Structured form with guidance; human fills in Low Depends on human
AI-assisted drafting LLM helps complete the template from a brief description Low–Medium Medium (needs human review)
AI-assisted classification LLM scores dimensions and proposes risk tier Medium Medium–High (needs human approval)
Semi-automated discovery Scanning tools detect AI usage and auto-populate basic fields Medium–High Medium (misses context)
Fully automated classification System auto-classifies and auto-assigns controls High Not recommended for HIGH/CRITICAL

AI-Assisted Drafting

Given a brief description — even a single sentence — an LLM can draft a structured use case definition by asking the right questions.

Input: "We want to use AI to help our support team respond to customer emails faster."

AI-generated draft:

Question AI-Populated Answer Confidence Needs Verification
What does it do? Drafts email responses for support team review before sending High Confirm: does it send directly or draft only?
Decision authority? Advisory — suggests response; human sends Medium Verify: will humans always review? Or auto-send for certain categories?
Data accessed? Customer email content (may contain PII), customer account data, support knowledge base High Confirm: what account data specifically?
Users? Internal support team High Confirm: any partner/outsourced support?
What if wrong? Incorrect draft; human catches before sending (if reviewing) Medium Verify: what if human doesn't catch it? What's sent on company letterhead
Volume? Unknown Low Requires: current email volume data
Regulatory? Depends on sector and data types Low Requires: sector identification
Tools/actions? Email draft generation; possibly CRM lookup Medium Confirm: what CRM access? Read-only or write?
Business process? Sits between incoming customer email and outbound response High Confirm: any other systems in the chain?
Accountability? Unknown Low Requires: named business owner

The AI identifies what it can populate (6 of 10 questions), what needs verification (3), and what it can't answer (1). The human effort shifts from writing to reviewing and completing.

Semi-Automated Discovery

For organisations with mature observability, use case discovery can be partially automated:

Data Source What It Reveals Use Case Field Populated
API gateway logs Which AI APIs are being called, by whom, at what volume Users (Q4), Volume (Q6), Tools (Q8)
Cloud service inventory Which AI services are provisioned Scope (Q1) — at a high level
IAM policies What data the AI service can access Data access (Q3)
Network flow data What the AI connects to upstream and downstream Business process position (Q9)
Billing data Usage patterns, cost allocation Volume (Q6)
Feature flag platforms What AI features exist and their status Scope (Q1), Status

What automated discovery cannot populate: - Decision authority (Q2) — requires understanding how humans use the output - What happens when wrong (Q5) — requires business context - Regulatory context (Q7) — requires legal/compliance knowledge - Accountability (Q10) — requires organisational knowledge

The shadow AI problem: Semi-automated discovery is most valuable for finding AI use cases that exist but aren't formally defined. The framework's Visibility Problem article notes that over 50% of enterprise AI adoption may be shadow AI. Automated discovery tools that scan API calls, cloud service provisioning, and browser traffic can identify undocumented AI usage — which can then be formally defined, classified, and governed.

Automated Classification — With Guardrails

For organisations processing many use cases (large enterprises may have hundreds), AI-assisted classification can accelerate the pipeline:

Use Case Submitted → AI Drafts Definition → AI Scores Dimensions
→ AI Proposes Tier → Human Reviews

Safeguards for automated classification:

Safeguard Purpose
AI proposes, human approves Final classification always requires human sign-off
Conservative bias When uncertain, AI should propose the higher tier, not the lower
Mandatory escalation Any dimension scoring HIGH or CRITICAL triggers mandatory human review
Confidence thresholds AI reports confidence per dimension; low confidence triggers investigation
Audit trail AI reasoning is logged; human can see why the tier was proposed
Periodic accuracy review Compare AI classifications to human decisions; calibrate over time

When full automation is appropriate: - Fast Lane qualification can be fully automated. The four criteria (internal, read-only, no regulated data, human-reviewed) are binary and verifiable from system configuration. An automated check against these criteria is more reliable than human self-certification. - Tier downgrades can be monitored automatically. The criteria (6+ months stable, no incidents, no scope change) are measurable from operational data.

When automation is not appropriate: - Initial classification of HIGH or CRITICAL systems. The consequences of misclassification are too significant. - Classification where dimension scores are borderline. Human judgement is needed for edge cases. - Any classification that would result in reduced controls. Reducing controls should always require human approval.


Use Case Anti-Patterns

Anti-Pattern What Goes Wrong How to Fix
Vague scope "AI assistant for the team" — no boundaries, no guardrails, no risk tier Require explicit positive and negative scope
Missing negative scope Use case says what AI does, not what it doesn't Require "the system cannot/does not..." statements
Understated decision authority Claimed "advisory" but functionally autonomous Ask: "What percentage of AI recommendations are changed by humans?"
Data access vs. data need System has broad database access but "only uses" a subset Assess risk based on access, not stated intent; apply least-privilege
No named owner "The team" is accountable Require a named individual as business owner
Static definition Use case defined at launch, never updated Scheduled reviews + change triggers
Copy-paste from vendor Use case definition is the vendor's marketing description Require internal definition in your context
Scope that grows silently Started as FAQ bot, now handles account changes Change detection via monitoring + periodic review
Volume unknown "We're not sure how many interactions" Require estimation before classification; measure after launch
Regulatory context ignored "That's legal's problem" Require legal/compliance input before classification

Use Case Definition Template

For practical use. Complete each section before submitting for risk classification.

Instructions: Every field is required unless marked optional. "TBD" is not an acceptable answer for sections 1–10. If you don't know yet, that's your next action — not submitting the form.

Section 1 — System Identity

Field Value
System name
Business owner (named individual)
Technical owner (named individual)
Date
Version
Status Draft / Under Review / Approved / In Production / Retired

Section 2 — Scope

Description
What it does Specific, bounded description of system behaviour in this deployment
What it does NOT do Explicit negative scope — list capabilities the system is prevented from exercising

Section 3 — Decision Authority

Field Value
Authority level Informational / Advisory / Influential / Autonomous
Specific decisions made or influenced
Human modification rate What percentage of AI outputs are changed by humans before acting on them?

Honesty check: If the human modification rate is below 10%, the system is functionally autonomous regardless of what the process document says. Score accordingly.

Section 4 — Data Access

Data Source Data Categories Sensitivity (Public / Internal / Confidential / PII / Sensitive PII / Regulated) Access Type (Read / Write)

Assess based on access, not intent. If the system can reach sensitive data through its database connection, that's the risk profile — even if the use case only needs a subset.

Section 5 — Users

User Type Description Estimated Count

Section 6 — Error Consequences

Question Answer
If the AI produces incorrect output, what happens?
What is the worst realistic outcome?
How is an error detected?
How is an error corrected?
Is the error reversible? Fully / With effort / With difficulty / Irreversible

Section 7 — Volume

Metric Value
Estimated interactions per day
Peak volume
Growth trajectory (6–12 month projection)

Section 8 — Regulatory Context

Field Value
Sector
Applicable regulations
Regulatory notification required for AI use? Yes / No / Unknown
EU AI Act risk category (if applicable) Unacceptable / High-Risk / Limited / Minimal / Not Applicable

Section 9 — Tools and Actions

Tool / API What It Does Access Level (Read / Write / Execute) Human Approval Required?

If the AI can call APIs, write to databases, send emails, or trigger workflows, it is agentic — even if nobody calls it that. Agentic controls apply.

Section 10 — Business Process Position

Field Value
Upstream inputs (what feeds the AI)
Downstream consumers (what acts on AI output)
Manual fallback (if the AI is unavailable)
Integration dependencies

Section 11 — Accountability

Role Named Individual Contact
Business owner (outcomes)
Technical owner (operations)
HITL reviewers (quality)
Escalation path

Section 12 — Risk Classification

Completed by risk function. Not self-assessed by the project team.

Dimension Score (LOW / MEDIUM / HIGH / CRITICAL) Justification
Decision Authority
Reversibility
Data Sensitivity
Audience
Scale
Regulatory
Overall Tier Highest dimension wins; three or more HIGH scores may escalate to CRITICAL
Field Value
Classified by
Approved by
Classification date
Next scheduled review

The Framework Gap

The framework defines risk tiers and controls but doesn't provide a structured use case definition methodology. The control selection guide has a decision tree, and the model card template captures system metadata, but neither bridges the gap between "we have an idea for AI" and "here's the risk profile and required controls."

This article fills that gap. The ten-question template, the six-dimension scoring model, and the RACI model for use case management connect strategy (what we want to build) to the framework (how we secure it).

Where the framework should evolve: - A formal use case definition template should be added to extensions/templates/ - Risk tier classification should reference this structured input, not assume it exists - AI-assisted classification guidance should be incorporated into the control selection guide - The model card template should be extended to include use case definition fields


AI Runtime Behaviour Security, 2026 (Jonathan Gill).