Agency-Built Automations vs Product Chatbots: A Practitioner Checklist for Ops Engineers

Disclosure: I build at NxFlowAI (nxflowai.com). This is a practitioner checklist, not a vendor ranking.

If you are shortlisting an AI automation agency for small-business workflows, the engineering questions matter more than the homepage adjectives. Here is the list I wish every ops engineer brought to the first call.

1. Idempotency at the edge

WhatsApp webhooks and form posts get retried. If your lead table treats every delivery as new, you will create duplicate owners and double messages. Ask: what is the idempotency key, and where is it stored?

2. Confidence thresholds with a human path

"AI replies to everything" is a product slogan. In production you want: auto-reply only below a risk threshold; otherwise queue for a person with the draft attached.

3. System of record vs system of inbox

WhatsApp is rarely the system of record. CRM (or even a disciplined sheet) is. Ask which fields are written, on which events, and what happens when the CRM write fails.

4. Observability

Who gets paged when template sends start failing? Where are structured logs? Is there a replay?

5. Exit

Can another engineer reconstruct the workflow from docs and credentials alone?

When a product chatbot is enough

If the problem is FAQ deflection on a single channel with no CRM write and no money language, a packaged bot may be enough. We wrote out that fork in custom AI vs off-the-shelf chatbots. If the problem is a multi-system handoff, start with a workflow audit before you buy seats.

NxFlowAI sequence: audit → build → monitor. First architecture/risk audit ~72 hours. We are not NexFlow/Nexusflow.