Solid walkthrough. The part I’d add before production is a durable decision envelope between n8n and HubSpot: source lead ID, normalized identity key, model and prompt version, qualification evidence, proposed action, and an idempotency key. Without that, retries can create duplicate deals or send the same follow-up twice, and you can’t reconstruct why a score changed. I’d also keep scoring separate from side effects: let the model propose, then let deterministic rules decide whether a CRM write or email is allowed. Disclosure: I’m Ege, co-founder/CEO of Atlantic in SF—we work on this coordination layer. Happy to compare failure cases if useful.