For ~5k leads/month, I would not build the core follow-up generation layer entirely inside Zapier.

The strongest architecture is usually:

Zapier = orchestration + triggers + integrations
Your backend + OpenAI API = intelligence layer

Why? Because once AI-generated follow-ups become part of your actual revenue workflow, three things start to matter very quickly:

Cost efficiency
Reliability
Control over prompts, retries, memory, and personalization

At ~5k leads/month, you’re already entering the zone where “AI inside Zapier” can become operationally expensive and harder to maintain. Multiple automation operators on Reddit and automation communities report that Zapier AI steps often introduce significant markup versus direct API usage, especially once workflows scale.