Task framing

What exactly should AI do—and should it do it at all?

Decide whether AI belongs in the task, then make the expected behavior concrete enough to test with real inputs, a human fallback, and clear stop conditions.

AI VISTA / DECISION01 / 03
WORKING DECISIONA good candidate has variable input, a checkable output, bounded stakes, and a fallback a person can use.
01INPUT
02CHECK
03BOUNDARY
RETURN TO THE REAL TASK

DECISION BRANCH

Do not begin with the tool name.

  1. ASK FIRST

    Is the missing piece a fact, a standard, or model capability?

  2. IF

    The result is verifiable and the failure cost is bounded, expand carefully.

  3. OTHERWISE

    Narrow the action, keep the human handoff, and preserve recovery.

WORKED CASE

Which part of support triage should actually be automated?

A team receives about 900 messages a week and wants AI to identify intent, assign a queue, and send replies. Those three actions carry different consequences but appear in the brief as one task: ‘automate support email.’

WORKING ARTIFACTAn action-boundary sheet mapping every step to its input, consequence, authority, and fallback.
01WEAK SHORTCUT

Treat the entire workflow as suitable because it is repetitive and high-volume, with saved time as the only justification.

02BETTER JUDGMENT

Separate reading, judging, routing, drafting, and sending. Begin with reversible classification suggestions and drafts; keep mixed intent, sensitive data, and refund commitments with a person.

03ACCEPTANCE

Review 50 historical messages and 10 boundary cases. Mixed-intent cases never auto-send, sensitive fields stay out of the model, and reviewers can see the rationale and handoff route.

ENTRY & EXIT SIGNALS

Know when to enter—and when the decision is good enough to leave.

A topic is not an endless knowledge directory. Entry signals tell you whether the problem belongs at this layer. Exit signals decide whether to continue instead of substituting time spent reading for work completed.

AI VISTA / DECISION STATUSTask framing
01ENTER HERE
  • The request still says ‘use AI for efficiency’
  • Actions with different consequences are bundled into one task
  • Acceptable output and human handoff are missing
02LEAVE WHEN
  • The AI action can be named with one precise verb
  • Input, boundary, failure consequence, and owner are explicit
  • A small set of real cases can support continue or stop

Carry inputs, rules, and output fields from the task brief into a map of what the model actually needs to see.

BOUNDARY

What this topic page will not do

This topic scopes an AI-assisted step; it is not a compliance certification for legal, medical, employment, financial, or other high-stakes decisions.

01Keep deterministic rules for hard gates02Do not let a language model approve consequential actions alone03Do not scale without a human owner who can finish the case
When the action, boundary, owner, and fallback are explicit, carry the task brief into the context topic.