How Risky Is This AI Use Case?
Not every AI idea needs the same scrutiny. A tool that drafts internal notes and a system that decides what a customer is told are different propositions. Eight questions place a proposed use in one of three bands: proceed, proceed with controls, or redesign first.
Score one proposed use at a time, as it would run in production, not as the pilot version. A higher score means higher risk.
AI Risk Assessment Quiz
Score one proposed use as it would run in production. Higher is riskier.
A triage tool, not a formal risk assessment or legal advice. Uses involving personal information, regulated activities or decisions about people may need a privacy impact assessment or sector-specific review. Nothing entered is recorded or transmitted.
What the Bands Mean
The score is a share of the maximum possible risk. The band tells you what to do with the idea, not whether it is a good one.
Proceed
Low-consequence, internal, reversible and reviewed. Normal project discipline is enough. Start it, measure it, and keep the review path until accuracy is known.
Proceed with controls
The use is workable, but one or more factors need a specific control before it goes live: a data handling decision, a mandatory review step, a rollback plan, or an exit clause with the vendor.
Redesign first
As proposed, the use combines sensitive data, external consequence and weak oversight. The usual fix is not to abandon the idea but to change its shape: recommend instead of act, narrow the data, or add a person before the output leaves.
How the Score Works
Eight questions, each scored zero to three where three is the highest risk. The total is shown as a share of 24, and the band thresholds are fixed.
Score each factor
Data sensitivity, who sees the output, regulatory exposure, reversibility, human review, explainability need, vendor lock-in and scale. Zero is lowest risk, three is highest.
Add them up
All factors are weighted equally. A single high-risk factor does not fail the use on its own, but three or four of them will.
Compare to the bands
Under one third of the maximum is proceed. One third to under two thirds is proceed with controls. Two thirds and above is redesign first.
Read the actions
Each band lists the controls that usually bring a use case down a band. The point is to change the design, then score it again.
The Factors That Move the Score
Four of the eight factors account for most redesign-first results. Each has a design change that reduces it.
Data sensitivity
Personal information, health records, financial details and anything covered by a client confidentiality clause raise the score sharply. The Privacy Act 1988 applies to personal information whether or not an AI system is involved. The design fix is to narrow the data: give the system only the fields the task needs.
- Personal information handling obligations do not change because a model is involved
- Health and financial data usually need a documented handling decision
- Redact or exclude fields the task does not need before any data moves
- Check where the vendor processes and stores data
Customer-facing output
Output that reaches a customer, a patient, a regulator or the public carries consequence that internal drafts do not. An error inside the business is a correction. An error outside it is a complaint, a refund or a notification. The fix is a review step, or a redesign so the system recommends to a person rather than replies directly.
- Internal use with a person in between is a different risk class
- Consumer law on misleading conduct applies to AI-written content
- Start external uses in recommend-only mode
- Decide whether customers are told they are dealing with AI
Reversibility
Ask what it costs to undo a wrong output. A draft can be deleted. A sent email cannot. A payment, a record change or a decision about a person may be very hard to reverse. Uses where errors are cheap to reverse can tolerate lower accuracy and lighter review than uses where they are not.
- Rank outputs by cost to undo, not by how often they might be wrong
- Add a hold step before anything irreversible
- Log every action so it can be traced and reversed
- Where undo is impossible, review is mandatory
Vendor lock-in
A use that depends on one vendor’s model, one proprietary format and one contract with no exit terms is a business continuity risk regardless of how well it works. The fix is contractual and technical: data export rights, a documented switch path, and a preference for standard formats.
- Confirm you can export your data and prompts in a usable form
- Ask what happens if the vendor deprecates the model you rely on
- Prefer designs where the model can be swapped
- Get exit terms in writing before, not after, go-live
Next Steps
AI Governance Checklist
The decisions that stop every use case renegotiating risk from scratch.
Open the checklist →AI Vendor Due Diligence Checklist
Thirty questions to put to a supplier before they touch your data.
Open the checklist →AI Pilot Scoping Checklist
Once the use case is proceed or proceed with controls, scope the pilot.
Scope the pilot →Frequently Asked Questions
Is a high score a reason to drop the idea?
Usually not. A high score means the use case, as currently designed, combines factors that need attention. Most high-scoring ideas can be brought down a band by changing their shape: narrowing the data they see, putting a person between the output and the outside world, adding a hold before anything irreversible, or negotiating exit terms with the vendor. Score it again after the redesign. If it is still redesign first, that is worth knowing before money is spent.
Why are all eight factors weighted equally?
Because a weighting scheme would be a claim about your business that we cannot make from here. Equal weights are transparent and easy to argue with, which is the point: if you believe regulatory exposure should count double in your sector, you can see exactly how that would change the result. The bands are set so that no single factor can push a use case into redesign first on its own, but two or three high-risk factors together will.
Does this replace a privacy impact assessment or a formal risk assessment?
No. It is a triage tool to decide how much scrutiny a use case needs before it goes further. Uses involving personal information may need a privacy impact assessment, and regulated sectors have their own requirements. The Voluntary AI Safety Standard published by the Australian Government and the NIST AI Risk Management Framework are both useful starting points for a fuller assessment. This quiz tells you which ideas warrant that effort.
What counts as human review?
A named person with time and authority who checks the output before it has effect, and who can correct or stop it. A person who is technically able to review but has no time allocated is not review, and neither is a notification nobody reads. The question in the quiz asks about the review as it would operate in production, because pilots usually have more review than the live system ends up with.
How should I score a use case that is partly internal and partly external?
Score the riskiest path the output can take. If a system drafts internal notes that are sometimes forwarded to customers unedited, it is customer-facing. If it usually recommends but can act directly in some cases, score it as acting. Designing out the riskier path, so the system cannot take it, is often the cheapest way to lower the score.
Is my answer recorded?
No. The quiz runs in your browser and nothing is transmitted. The questions ask about the shape of a proposed use, not about any confidential detail. Screenshot the result if you want it for a discussion, and use the tool as often as you like without contacting us.
Sources and further reading
- The Privacy Act 1988 (Office of the Australian Information Commissioner)
- Voluntary AI Safety Standard (Department of Industry, Science and Resources)
- NIST AI Risk Management Framework (US National Institute of Standards and Technology)
Landed in Proceed with Controls or Redesign First?
Send us the use case and your score. We will tell you which controls would bring it down a band, what they cost, and whether the redesign is worth it.