Human–AI Workflow Design

Make judgment, authority, uncertainty, and recovery explicit.

01 / flow

Authority chain

Human sets goalsTools supply evidenceAI interpretsSoftware validatesHuman approves if neededSoftware executes

02 / table

Assign each layer a job

Layer	Owns
People	Goals, risk tolerance, exceptions, consequential approval
AI	Interpretation, classification, synthesis, suggestions and uncertainty
Software	Permissions, schemas, transitions, version checks and safe fallback
Tools + data	Bounded retrieval, scoped API access, observable results

03 / table

Suggest vs enforce vs authorize

AI suggests	Software enforces	People authorize
Ambiguous judgments	Exact permissions and limits	Consequential or irreversible actions
Draft recommendations	Validation and version checks	Tradeoffs and exceptions
Possible changes with evidence	Safe defaults and denial	What becomes authoritative

04 / list

Make uncertainty visible

Evidence is not automatically accepted truth
Pending is not approved; unknown is not zero
Show source freshness and provenance
Offer correction, rejection, unchanged, and escalation paths

05 / table

Failure recovery

Failure	Safe response
Provider unavailable	Fail closed or use a deliberately limited fallback
Tool timeout	Retry only if safe; show manual path
Missing permission	Deny action and explain authorization
Weak evidence	Request more or keep question open
Stale state	Version check and re-review
Bad consequential action	Preserve history, reverse or escalate, add regression

06 / note

Remember

Human-in-the-loop is not a person tacked on at the end. Authority and recovery must be designed throughout the workflow.

GOING DEEPER

Debugging matrix and PM gates

First observable failure

Layer	Signal	Inspect
Instructions	The same evidence is misread across cases	Conflicting rules, weak examples or ambiguity
Retrieval	Correct source missing, stale or conflicting	Actual retrieved context, freshness and scope
Tools	Timeout, partial response or wrong data	Inputs, results, bounded retries and dependency
Permissions	Agent can see/do too much or too little	Roles, scopes, approval and isolation
UX	User cannot understand or recover	Visible state, explanations and reversal path
Logic	Interpretation reasonable but product result wrong	Schema conversion, branches, versioning and ownership

PM design and release checklist

When	Check
Before design	Map human authority, sensitive data, uncertainty, read/write scope and reversibility
Before release	Separate evidence from proposals, enforce validation/version checks, test approval and corrections, expose failures and fallback
During iteration	Retest original and neighboring cases, holdouts and repeat runs; measure silent misses, false certainty and human review burden

Recovery cases: When data is weak, keep a question open. When state is stale, block changes and require re-review. When tools fail, retry only if safe or escalate. When an unauthorized action slips through, preserve history, support reversal and add a regression case.

PRACTICAL FRAMEWORK / AI PRODUCT WORK

Reversible autonomy: decide, act, revoke

Is the action consequential or hard to reverse?Yes → approve firstNo → bound scope and verify recoveryMonitor defined evidenceReverse or compensate when contradictedRevoke autonomy when reliability degrades

Gate	PM question
Before action	Is the action low impact, reversible in practice, time-bounded, logged and within an approved permission scope?
After action	What contrary evidence, timeout or changed state triggers compensation or human review?
Authority withdrawal	What failure rate, serious incident or missing evidence forces the system back to recommend-only?

Compensation is not deletion: retain the audit trail, and never call an action reversible without testing the recovery path.