Independent, source-aware and testable AI guides
AI delegation framework · A3

What Should You Delegate to AI? A Four-Decision Framework

Decompose work across definition, context, discernment and accountable designation to set practical human-AI boundaries.

Conceptual diagram for What Should You Delegate to AI? A Four-Decision Framework
Short answerDo not assign an entire job to 'AI' or 'human'. Split it into work units: AI may draft and surface patterns while people retain criteria, exceptions and authority over consequential decisions.

The automate-or-manual binary hides the shape of most workflows. Collection and formatting may be reversible; interpretation may require expertise; publication or customer action may require formal authority.

The four-decision label here is an editorial mnemonic, not a global standard. Use NIST and OECD sources for formal risk and governance reference.

Evidence the course or workflow should provide

Define

A person sets the unit of work, purpose and completion boundary.

Observable evidence: Task card.

Describe

AI may organise context; the data owner approves what is permitted.

Observable evidence: Context pack.

Discern

AI can surface signals; a reviewer applies criteria and hard failures.

Observable evidence: Rubric and evidence.

Designate

External authority and accountability remain assigned to people.

Observable evidence: Approval map.

Decompose

Large work becomes reversible units with visible handoffs.

Observable evidence: Workflow map.

Human-AI matrix by consequence

Fill each cell with an actual role; a shared task needs an explicit handoff.

Work unitAI roleHuman roleControl
CollectSummarise/classifyApprove sourcesSource gate
DraftGenerate optionsSet intent and voiceRubric
AnalyseSurface patternsInterpret exceptionsEvidence review
DecideRecommendHold authoritySign-off
ActOnly authorised scopeOwn consequenceAudit/rollback

How to apply this guide

  1. Map the process firstIdentify input, judgement, output and external consequence before selecting tools.
  2. Rate each unitScore reversibility, data sensitivity and error severity separately.
  3. Start with low consequenceSummaries and drafts before decisions and actions.
  4. Test the handoffName who resolves ambiguity and by when.

Failure modes to avoid

  • Whole-job delegationLow- and high-impact units become inseparable.
  • Ceremonial human reviewA click without criteria or time is not a control.
  • Model accountabilityA model cannot hold organisational authority.

Official and primary sources

  1. NIST AI RMF Core: Govern, Map, Measure, Manage
  2. OECD AI Principles
  3. Google Cloud Responsible AI
  4. AWS Responsible AI
Next step

Practise the matrix with AI Fluency Foundations

Compare the course scope and write the human review points for your own process.

View AI Fluency Foundations