Workflow builder, reviewer, and AI administrator
Use AI for proposals while people retain control
FlowOS AI assists with drafts, explanations, mappings, summaries, simulations, risk signals, and diagnostics. Its output is a proposal. It does not replace validation, publishing authority, or approval for guarded actions.
Before you start
- AI provider configured or fake provider enabled for local testing.
- AI permissions and monthly limits reviewed.
- An inactive workflow draft.
#Choose a bounded AI task
Ask AI to explain a failure, suggest a draft branch, summarize an audit record, or propose field mappings. Provide only the minimum redacted context. Treat confidence and risk indicators as signals for review, not guarantees.
#Review and apply a proposal
- Confirm the intended workflow and inactive version.
- Read the proposed trigger, fields, rules, and actions.
- Remove unnecessary or sensitive context.
- Edit incorrect assumptions.
- Approve or reject the proposal.
- Validate and publish separately using the normal release process.
#Use the approval inbox for guarded actions
A guarded tool call pauses before execution. The reviewer checks the target, payload, risk, and business reason, then approves or rejects. Approval and execution are distinct events so the audit trail can show who decided and what eventually happened.
#Common mistakes
- Sharing secrets or unnecessary customer data with the provider.
- Treating a high confidence score as approval.
- Combining proposal approval and production publishing into one unchecked action.
#Verify the result
- The proposal targets an inactive draft.
- Shared context is minimal and redacted.
- A human owns the final decision.
- Execution evidence is reviewed after approval.