Workflow Automation · Read + Write

Email triage + draft reply

Augment Rep · High Reason · Medium

The problem

Inbox is the new task queue. 80 emails a day; 50 are FYI, 20 need a 30-second reply, 10 actually need thinking. Mental cost of triaging is the worst part.

The AI approach

Classify each email by intent + priority, draft a reply for the routine ones, surface the ones that need real attention.

The outcome

~70% drafted

70% of emails arrive with a draft already attached. The user edits or hits send. Zero-inbox by 10am.

Try itInput → Process → Output

Input — inbox · last hour
Process — AI pipeline
1Classify intent + priorityReadPredictive
2Apply routing rulesRulesSymbolic
3Draft reply for routine emailsWriteGenerative
4Surface human-needed firstRulesSymbolic
Output — triaged inbox
Click Run demo to classify each email and draft routine replies.
Triaged · 5 emails 3 drafted · 1 escalated
Sample draft (Sam Lee): "Hi Sam — Wednesday 2pm works on my side. I'll send a calendar invite. Best, ..."

Three AI types in this use case

SymbolicRouting rules: priority sender list, escalation paths, urgent-keyword overrides, auto-archive list.
PredictiveIntent classifier (FYI / question / scheduling / action) + priority predictor (sender, recency, thread depth).
GenerativeLLM drafts reply in your voice using past-thread context; flags the ones that need your judgement.

The stack

  • Email API · Gmail / Outlook
  • Intent · fine-tuned BERT classifier
  • LLM · Claude with voice prompt
  • Sandbox · drafts only, no auto-send

When this works

  • You have ≥ 6 months of past replies for the LLM to learn voice from
  • Most emails fall into a stable set of intents
  • You retain final-send authority — drafts stay in Drafts

When it fails

  • Sensitive negotiations — model misjudges tone
  • Ambiguous email — should this reply happen at all?
  • Voice mimicry too good — recipient thinks it is you, but it is wrong
  • Auto-send turned on without sampling