Most “AI for business” pitches are a demo that works once. The hard part is not the model. It is everything around it: messy inputs, systems with no API, and a failure mode where the thing is confidently wrong and nobody notices for a month.
How I build automation that survives production
- Deterministic guardrails. The model proposes; validated code decides. Anything consequential is checked before it commits.
- Loud failure. A broken sync alerts a human. Silent data loss is the failure that actually costs money.
- Human-owned architecture. You get a system your next developer can read, not a black box only I understand.
- n8n orchestration so workflows are visible and editable rather than buried in someone’s script.
Where it genuinely pays off
- Inbound enquiry triage and routing
- Quote and proposal drafting from structured event data
- Document and invoice extraction into your accounting stack
- Bridging legacy systems that will never get a modern API
And where it does not: if a problem is better solved by a scheduled export and twenty lines of code, that is what I will build. Cheaper to run, far less to go wrong.