Blog · Pattern
Stop automating inboxes. Start designing workflows.
Most "AI email automation" is one prompt away from disaster. The fix isn't a better prompt. It's a smaller job for the model, with everything else handled by infrastructure that doesn't hallucinate.
Things the model should not be in charge of.
Hand any of these to the LLM and you're asking probability to do bookkeeping. It will, eventually, get it wrong in a way that's expensive.
- Whether to retry. (That's an idempotency key, not a vibe.)
- Whether the sender is real. (That's a directory lookup.)
- What thread this belongs to. (That's a message ID.)
- Whether a human needs to approve this. (That's policy.)
A pattern that survives production.
- Inbound webhook fires. The platform has already parsed the fields, deduped the thread, and verified the sender against the directory.
- A deterministic router decides: agent, human, or drop. No model yet.
- If an agent: the model gets a narrow job — classify, draft, extract — with the surrounding state passed in as context.
- The output goes through a policy check. High-risk actions queue for human approval.
- The reply is sent into the original thread. The whole exchange is the audit log, automatically.
The best tool for agent communication is the one already running everywhere.
Queues are great inside a system. Chat APIs are great inside a product. Neither reaches the vendor your agent needs to email about an invoice.
Email does. And once the workflow is structured this way — model in the middle, infrastructure on the edges — you can swap models without rewriting the workflow, and you can audit the workflow without trusting the model.
Where assmbl fits.
We handle the unglamorous parts: parsing, threading, sender identity, encryption, retry, billing. You handle the part that's actually interesting — what the agent should do.