Free worked example · Synthetic records · Runs only in your browser · No real messages or transactions
Workflow Demo 3: AI-Assisted Support Drafting with Human Approval
Intended User and Problem
User: Operations owner of a support inbox for a subscription service.
Problem: Agents need speed for routine replies, but must avoid risky auto-responses and never auto-approve refunds.
Visual Workflow
Inputs / Outputs / Assumptions
Inputs: ticket email, sentiment, request type, urgency, amount at risk, customer tier.
Outputs: draft reply, escalation flag, approval state, and final action outcome.
Assumptions: all examples are fictional. No AI model is called: replies are fixed placeholder templates. Risk fields are supplied inputs, not verified classifications. Simulated approval is not authenticated and cannot authorize any real action.
Synthetic Data (Fictional)
Press "Run local simulation" to run drafting and approval checks.
Notes and CTA
No payment/refund decisions are made automatically in this workflow; those cases are escalated for human action.
CTA: Review this in tandem with the Small-Business AI Runbook framework or request an async blueprint enquiry.
Read this before adapting the workflow
No AI model is called: responses are fixed placeholders. A human must insert and verify instructions from an approved knowledge base before real use. Risk labels are supplied inputs, not verified classifications. The example approval name is not authenticated. Production needs role permissions, approval records bound to the exact draft version and durable delivery keys. No refund, account change or real message occurs here.
Implementation references (not endorsements or connected integrations): OpenAI guidance on human oversight and adversarial testing