How to Risk-Assess an AI Use Case in 30 Minutes, Repeatably

TL;DR: The failure mode of AI risk management at normal companies isn’t recklessness, it’s uniformity. Either everything gets treated as dangerous (and adoption stalls) or everything gets waved through (until the incident). The fix is a repeatable scoring method that sorts uses in about 30 minutes: five factors, scored 1-3, summed into green/amber/red tiers with pre-agreed controls per tier. This page is the method, the scoring rubric, the controls, and the one-page record format. Not legal advice, high-stakes or regulated uses need qualified review on top.

The unit of analysis: the use case, not the tool

The security review answers “is this vendor and product safe to hold our data?” That’s necessary and insufficient, because the same approved tool spans wildly different risk depending on what it’s doing:

So the assessable unit is a use case: who uses which tool to do what, with which data, producing output that goes where. One sentence in that shape is the header of every assessment. If you can’t fill in the sentence, the use isn’t defined enough to assess, which is itself a finding.

The five factors

Each scored 1-3. The factors are chosen so that a non-specialist can score them from knowledge of the workflow alone, no model internals required.

1. Data sensitivity, what goes in

Score the most sensitive data the workflow will realistically touch, not the average. The categories align with the never-paste list in AI and data privacy.

2. Decision impact on people, what it decides

Be honest about the 2/3 boundary: if humans approve the AI’s ranking 95% of the time, the AI is determining outcomes regardless of the org chart. That’s the rubber-stamping problem from Responsible AI in practice.

3. Audience, where the output goes

External audience raises stakes on every error class: a hallucinated fact in a personal draft costs a correction; the same fact in a customer proposal costs trust or worse.

4. Autonomy, how it acts

This factor climbs as tools become agentic, an AI agent running multi-step tasks is a 3 by construction. Autonomy also compounds security exposure: an autonomous tool reading untrusted content is the classic prompt-injection target, which is why amber/red tiers loop back to the security review.

5. Reversibility, what an error costs to undo

Scoring and tiers

Sum the five factors: range 5-15.

Two override rules trump the arithmetic, because averaging can hide a critical factor:

• Any use scoring 3 on decision impact (determines outcomes for people) is at minimum amber, and red if data sensitivity or reversibility is also 3. Hiring screeners land here every time, appropriately, since they’re also the EU AI Act’s most SMB-relevant high-risk category.
• Any use with regulated data (3) and autonomy (3) together is red regardless of total.

Worked examples

(D = data, P = people impact, A = audience, Au = autonomy, R = reversibility.)

Controls per tier

Pre-agreeing the controls is what makes the method fast, the score is the decision:

Green (5-7):

• Covered by the acceptable-use policy; tool must be on the approved list for the data class. That’s it. Log one line in the register.

Amber (8-11):

• Named use-case owner (a person, not a team).
• Data scoped: only the fields the task needs enter the workflow, the minimization habit from AI and data privacy.
• Review designed, not assumed: full review of external outputs, or documented sampling for high-volume flows, per the oversight patterns in Responsible AI.
• Tool’s security review confirmed current for this integration scope.
• One-page assessment filed; re-score on change triggers.

Red (12-15):

• Everything in amber, plus: governance-owner and leadership sign-off; deep-tier security review; a written oversight plan (who can intervene, how, with what authority); legal/professional input where the use touches employment, credit, or regulated data; a staged rollout with logging; and a defined rollback.
• And the question that must be asked out loud: should we do this at all? Some red uses are worth it with controls. Some are a hiring screener you don’t actually need. Declining is an assessment outcome, not a process failure.

The one-page record and the register

Each assessment produces one page, enough to reconstruct the reasoning, short enough to actually get written:

AI USE-CASE RISK ASSESSMENT
Use case: [who] uses [tool] to [do what] with [data], output to [audience]
Requested by: [name]            Assessed by: [governance owner]   Date: [date]
Scores: Data [ ] · People [ ] · Audience [ ] · Autonomy [ ] · Reversibility [ ]
Total: [ ]   Tier: [Green/Amber/Red]   Overrides triggered: [none / which]
Controls required: [list, per tier]
Owner: [name]        Review trigger: [material change / incident / annual]
Decision: [approved / approved with conditions / declined]  Notes: [2-3 lines]

The register is just these records in one place, a spreadsheet or a folder works. It earns its keep three ways: it’s the inventory the EU AI Act sequence starts from, it’s the evidence that diligence happened when anyone asks, and it’s what makes re-assessment a five-minute diff instead of a fresh exercise.

Running it: the process in six steps

1. Requester writes the use-case sentence and drafts scores. Ten minutes with the rubric. Self-scoring makes the process scale; the next step keeps it honest.
2. Governance owner reviews the scoring. Confirms or adjusts, checks overrides, checks the tool is approved for the data class. Fifteen minutes for anything green or amber.
3. Apply the tier’s controls. Pre-agreed, so this is assignment, not negotiation.
4. File the page; update the register.
5. Re-score on triggers. Material change to data, autonomy, audience, or the tool behind the workflow; any incident; or the annual register pass. The most common trigger in practice: a green drafting use quietly gains autonomy (“we just let it send now”), which is exactly the drift the autonomy factor exists to catch.
6. Feed patterns back into policy. If a category keeps arriving amber for the same reason, encode the fix as a standing rule in the acceptable-use policy and stop re-deciding it.

Teams in legal or legal-adjacent work should treat this method as the floor, not the ceiling, privilege and matter-confidentiality add considerations covered in the legal guides.

FAQ

What is an AI use-case risk assessment?

A structured 30-minute evaluation of one specific AI use, who uses which tool to do what, with which data, output going where, scored on data sensitivity, people impact, audience, autonomy, and reversibility, mapped to green/amber/red control tiers, and recorded on one page in a living register.

Why assess use cases instead of tools?

Risk lives in the application. The same tool is near-zero risk summarizing public reports and red-tier risk filtering job applicants. Tool-level vetting (the security review) and use-level assessment answer different questions, vendor safety versus application safety, and approving a tool never implies approving every use of it.

How do we score an AI use case?

Five factors, 1-3 each: data sensitivity, decision impact on people, audience of output, autonomy, and error reversibility. Sum to 5-15: green 5-7, amber 8-11, red 12-15. Two overrides: outcome-determining decisions about people are at least amber (red if paired with sensitive data or irreversibility), and regulated data plus full autonomy is automatically red.

Who should perform the risk assessment?

The requester drafts the use-case sentence and self-scores; the AI governance owner reviews, adjusts, and confirms the tier. This split keeps throughput high enough that people use the front door while preserving a four-eyes check against optimistic scoring. Red-tier uses add leadership sign-off and, where employment or regulated data is involved, qualified professional input.

How often should assessments be revisited?

On triggers, material change to the use (new data class, new autonomy, new audience), change to the underlying tool, or any incident, plus one light annual pass over the register. The one-page record makes most re-scores a five-minute comparison rather than a new assessment.

Rolling this out and want a hand? Webisoft helps companies implement AI safely, get in touch.