Purpose and use
A design worksheet for deciding when AI should answer directly, provide a hint, ask for a first attempt, require verification, or route the task to a human.
Decision supported: What level of AI assistance supports the task without weakening safety, accessibility, authority, or the human capability the organization needs to retain?
Use the design questions to choose an assistance level, name the reviewer, and specify how independent capability will be checked.
Who should use it
- Product owner
- UX or service designer
- Engineering lead
- Learning or capability owner
- Accessibility specialist
- Domain reviewer
- Security or risk owner
- Operational decision owner
1. Task consequence
Record:
- What decision or action can the AI output influence?
- Who or what is affected if it is wrong, late, incomplete, or unauthorized?
- Is the outcome reversible?
- How quickly can an error be detected and repaired?
- Does the task involve financial, legal, safety, privacy, eligibility, access, or service consequences?
- Which cases must refuse or escalate rather than answer?
Suggested decision:
- Low-consequence reference or clerical support
- Moderate-consequence professional assistance
- High-consequence reviewed recommendation
- Prohibited for AI proposal or execution
2. User expertise
Record:
- Novice, developing, experienced, or expert user
- Domain knowledge required
- Ability to detect unsupported AI output
- Familiarity with sources and policy
- Need for training or calibration
- Risk of over-reliance or automation bias
Decide whether the interface should explain more, require a first attempt, provide direct assistance, or route to a qualified reviewer.
3. Learning and transfer goal
Select the primary goal:
- Complete the task quickly
- Teach a concept or procedure
- Build independent judgment
- Preserve a rarely used critical skill
- Support reflection and review
- Produce a governed operational result
If learning or independent judgment matters, define how the system will test transfer beyond the assisted case.
4. Time and safety constraints
Record:
- Response deadline
- Safety or incident urgency
- Service-level expectation
- Cost of delay
- Availability of a human expert
- Whether direct guidance is necessary to prevent harm
Do not introduce staged assistance when delay is itself unsafe or operationally irresponsible.
5. Accessibility accommodations
Record:
- Keyboard and assistive-technology requirements
- Cognitive-load and memory demands
- Language or literacy needs
- Visual, auditory, motor, or speech accommodations
- Need for reduced steps, direct answers, or alternative formats
- Whether a required first attempt creates an unnecessary barrier
Accessibility is not optional friction. The assistance design must remain perceivable, operable, understandable, and robust.
6. Select an assistance level
Level 0 - Observe
The system provides a workspace, captures context, or waits without generating an answer.
Use when: the human should perform the full reasoning or when evidence is not yet sufficient.
Level 1 - Ask
The system asks the user to state the goal, assumptions, or first interpretation.
Use when: a first attempt supports learning, judgment, or clearer intent.
Level 2 - Hint
The system identifies a relevant concept, source, or area of concern.
Use when: the user should continue the reasoning but may need orientation.
Level 3 - Scaffold
The system provides a checklist, template, staged questions, or partial structure.
Use when: the process matters and the user should retain ownership of the conclusion.
Level 4 - Draft
The system creates proposed work with sources, limitations, and required verification.
Use when: acceleration is valuable and a qualified person can review the result.
Level 5 - Execute under authority
The system performs a bounded action after explicit approval and with audit evidence.
Use when: the action, authority, rollback, and failure behavior are defined and tested.
7. Reflection and verification
Choose at least one appropriate control:
- User identifies the supporting source
- User explains what changed from the AI draft
- User states the remaining limitation
- User compares alternatives
- User confirms the downstream consequence
- User performs a second independent check
- Another reviewer approves
- The system runs a deterministic test or reconciliation
- The system refuses when evidence is insufficient
8. Independent-performance check
When human capability matters, define one or more measures:
- Similar task completed without AI
- Transfer to a new case
- Ability to identify an incorrect AI response
- Confidence compared with correctness
- Persistence after a difficult case
- User-authored reasoning quality
- Need for escalation
- Retention after time has passed
Do not use immediate time-on-task as the only outcome.
9. Human-review ownership
Name:
- Product or workflow owner
- Domain reviewer
- Approver
- Executor
- Escalation owner
- Accessibility owner
- Learning or capability owner
- Evidence owner
Define who may change the assistance level and who approves a higher-authority execution path.
10. Measurement plan
Track only what is necessary and approved. Candidate measures include:
- Assisted and independent task quality
- Unsupported-claim rate
- Reviewer agreement
- Override and refusal rate
- Time and cost
- Accessibility defects
- Blind acceptance indicators
- User-reported usefulness
- Escalation patterns
- Operational incidents
Do not add surveillance merely because it is technically possible.
Decision record
Conclude with:
- Selected assistance level
- Reason
- User group
- Consequence class
- Required sources
- Required review
- Accessibility accommodations
- Verification step
- Refusal/escalation criteria
- Metrics
- Review date