Templates
Inbound email routing
Send each email that reaches a shared address to the right queue, and flag the ones that need a human reply today.
Send each email that reaches a shared address to the right queue, and flag the ones that need a human reply today. Pattern: intent_routing. Model: jev-1.13.0. Outage rule: review.
When to use it
- A shared address such as hello@ or support@ gets mail for several teams, and someone forwards it by hand.
- You parse inbound mail with a webhook and want a queue name back, not a paragraph.
- Pair it with the email triage template when a person's own inbox also needs urgency.
When not to use it
- You need the reply written. That is text generation; route first, then let an LLM or a person reply.
- Routing follows the address exactly, such as billing@ always going to billing. Do that in code.
Questions
| Id | Type | Asks | Answers | Gating |
|---|---|---|---|---|
queue | choice | Which team should handle email, judged by what the sender wants? It arrived at the shared address email.to. | support, billing, sales, partnerships, security_report, spam_or_automated, none_of_these | yes |
needs_reply_today | noul | If email is from a person who wants something, does that person need a human reply today to avoid real harm or a lost customer? | true or false | yes |
Reading the result
- Replace the
queueoptions with your own queues, and keepnone_of_these. - Move the email only when
overallActionisauto.reviewgoes to a person who sorts by hand. Therouteis the queue name. security_reporthas a stricter bar throughperOption, since a missed vulnerability report is costly.needs_reply_todayonly counts when the queue is notspam_or_automated.
Spec
Save it as bandwise/sets/inbound-email-routing.json and edit it for your data.
{
"schemaVersion": 1,
"model": "jev-1.13.0",
"input": {
"schema": {
"type": "object",
"required": [
"email"
],
"properties": {
"email": {
"type": "object",
"required": [
"from",
"to",
"subject",
"text"
],
"properties": {
"from": {
"type": "string"
},
"to": {
"type": "string"
},
"subject": {
"type": "string"
},
"text": {
"type": "string",
"maxLength": 20000
}
}
}
}
}
},
"stages": [
{
"id": "route",
"questions": {
"queue": {
"type": "choice",
"instructions": "Which team should handle `email`, judged by what the sender wants? It arrived at the shared address `email.to`.",
"criteria": {
"support": "A customer needs help using the product or reports something broken.",
"billing": "Invoices, refunds, payment failures, plan changes or cancellations.",
"sales": "A prospect asks about buying, pricing, a demo or a quote.",
"partnerships": "Integrations, resellers, co-marketing or vendor pitches.",
"security_report": "Someone reports a vulnerability or a security concern in the product.",
"spam_or_automated": "Spam, bulk mail, auto-replies, bounces or notifications.",
"none_of_these": null
},
"meta": {
"label": "Queue"
}
},
"needs_reply_today": {
"type": "noul",
"instructions": "If `email` is from a person who wants something, does that person need a human reply today to avoid real harm or a lost customer?",
"criteria": {
"true": "Delay would cost something today: an outage, a blocked customer, a payment problem, a deadline or an angry escalation.",
"false": "A reply within a few working days is fine."
},
"meta": {
"label": "Needs a human reply today"
}
}
}
}
],
"policies": {
"queue": {
"type": "choice",
"gating": true,
"thresholds": {
"high": 0.6,
"medium": 0.35
},
"perOption": {
"security_report": {
"high": 0.8,
"medium": 0.5
}
},
"actions": {
"high": {
"kind": "auto"
},
"medium": {
"kind": "review"
},
"low": {
"kind": "review"
}
}
},
"needs_reply_today": {
"type": "noul",
"gating": true,
"relevantWhen": {
"q": "queue",
"neq": "spam_or_automated"
},
"noul": {
"trueAt": 0.75,
"falseAt": 0.25,
"reviewMargin": 0.1
},
"actions": {
"high": {
"kind": "auto"
},
"medium": {
"kind": "review"
},
"low": {
"kind": "review"
}
}
}
},
"routes": [
{
"when": {
"q": "queue",
"eq": "support"
},
"output": "support"
},
{
"when": {
"q": "queue",
"eq": "billing"
},
"output": "billing"
},
{
"when": {
"q": "queue",
"eq": "sales"
},
"output": "sales"
},
{
"when": {
"q": "queue",
"eq": "partnerships"
},
"output": "partnerships"
},
{
"when": {
"q": "queue",
"eq": "security_report"
},
"output": "security"
},
{
"when": {
"q": "queue",
"eq": "spam_or_automated"
},
"output": "archive"
}
],
"defaultRoute": "triage_by_hand",
"savings": {
"comparatorModel": "claude-haiku-4-5",
"estOutputTokensPerQuestion": 60,
"kind": "decision"
},
"onUnavailable": "review"
}Example states
The expected outcome is what a person would decide. It is not a recorded model answer.
Failed payment
Expected: billing, needs a reply today.
{
"email": {
"from": "ops@customer.example.com",
"to": "hello@yourco.example",
"subject": "Card declined, account locked",
"text": "Our card was declined this morning and now the whole team is locked out. We updated the card but it still says suspended. Please help, we have a client demo at 3pm."
}
}Demo request
Expected: sales, no same-day reply needed.
{
"email": {
"from": "maria@prospect.example.org",
"to": "hello@yourco.example",
"subject": "Demo for a 40 person team?",
"text": "We are comparing tools for next quarter. Could we book a demo sometime in the next few weeks?"
}
}Out of office reply
Expected: spam_or_automated, archived.
{
"email": {
"from": "noreply@partner.example.net",
"to": "hello@yourco.example",
"subject": "Automatic reply: Following up",
"text": "I am out of the office until Monday with limited access to email."
}
}Borderline cases
One case near the line for each question. Use them to test your wording before you trust the thresholds.
queue
A customer asks for a refund because a feature is broken: billing or support?
{
"email": {
"from": "kim@customer.example.com",
"to": "support@yourco.example",
"subject": "Refund please",
"text": "The export has not worked for two weeks and we cannot use the product. I would like a refund for this month."
}
}needs_reply_today
Polite follow-up on an old question, with a soft deadline later in the week.
{
"email": {
"from": "raj@customer.example.com",
"to": "hello@yourco.example",
"subject": "Following up on SSO",
"text": "Just checking in on my SSO question from last week. We present the rollout plan on Thursday."
}
}Try it
Save an example state as state.json, then run the spec locally. Local mode makes no network call and needs no key; answers are synthetic unless a recorded fixture matches.
pnpm bandwise run --local bandwise/sets/inbound-email-routing.json state.jsonBandwise is an independent product built on TypeSafe's System One models. It is not TypeSafe's documentation. For the System One models themselves, see docs.typesafe.ai.