Practitioner Notes: Designing a Custom AI Workflow Before You Pick a Model

I build at NxFlowAI. Steal the checklist.

If you are hiring a custom AI workflow agency, ask them to show the state machine before the prompt library. This is written for people who build for owner-run shops and clinics, where the owner is also the approver.

Minimal states that matter

1. Ingest: message/event arrives; idempotency key set.
2. Classify: intent label your team already uses (catalogue, booking, price check, support).
3. Route: named owner + next step written to CRM/sheet.
4. Draft: model suggests a reply (optional).
5. Approve: human gate for price/promise; auto-send only for low-risk FAQs.
6. Monitor: alerts on webhook/CRM failures; documented pause/rollback.

Skip steps 4-5 if you only need routing. Many teams do. The expensive mistake is buying step 4 while steps 1-3 are undefined.

Anti-patterns

• Training a bot on a FAQ PDF while ownership is undefined.
• Letting the vendor own the WABA "for convenience."
• No dedupe when the same webhook is delivered again after a non-200 response.
• "We'll monitor later" with no named alert path.

What a good discovery artefact looks like

A one-page map, a risk list (including wrong-price send), and an implementation plan you could hand to another builder. We time-box a first audit around 72 hours.

DIY map first: write the path from first message to owned next step for one real enquiry type.

NxFlowAI (nxflowai.com) is Mumbai-based and remote-first; not NexFlow/Nexusflow. [e-mail]