Docs
How InboxSquire turns email into agent-ready signal.
A look under the hood: the seven-stage triage loop, the structured output every email produces, and the safety rules that gate every action before it touches your inbox or your fleet.
The triage loop
Seven stages, one direction.
Every connected message runs the same path — from raw email to a structured decision you can act on or hand to an agent.
- Stage 1
Classify
Each message is sorted into a category — hot lead, escalation, finance, vendor pitch, noise, and more.
- Stage 2
Score
A signal score and confidence value rank how much the message actually matters to you.
- Stage 3
Summarize
The thread is condensed to a short, scannable summary so you read the point, not the prose.
- Stage 4
Route
Messages that warrant action are matched to the right fleet agent or your review queue.
- Stage 5
Draft
When a reply is appropriate, a draft is prepared for you to review — never sent automatically.
- Stage 6
Escalate
High-urgency or risk-flagged items are surfaced for a human decision before anything happens.
- Stage 7
Log
Every classification, route, draft, and approval is recorded in your audit log.
Classification output
Each email yields one JSON object.
The classifier emits a single structured record per message. That record is what drives scoring, routing, drafting, and your audit log — and it is the contract the rest of the fleet reads.
{
"category": "hot_lead | customer_escalation | finance | vendor_pitch | sponsor_opportunity | newsletter | internal_alert | spam_noise | unknown",
"signalScore": 0,
"urgency": "low | medium | high | critical",
"confidence": 0,
"recommendedAgent": "Kane | Midas | MiniCFO | HAMR | TextAlerts | AltBot | none",
"recommendedAction": "draft_reply | create_task | create_crm_note | send_sms_alert | archive | label | snooze | ignore | route_to_agent",
"reason": "short explanation",
"approvalRequired": true,
"riskFlags": ["sensitive_data", "legal", "financial", "angry_customer", "unsubscribe", "unknown_sender"]
}signalScore and confidence are normalized values used to rank and gate. approvalRequired and riskFlags are what the safety rules below read before anything acts.
Routing & safety rules
The classifier proposes. You decide.
Routing is a recommendation, not an instruction. These rules sit between the output schema and any real-world action.
Never auto-send. Drafts are prepared for your review and only go out when you approve them.
Never auto-delete. Nothing is removed from your inbox on InboxSquire's initiative.
Low confidence routes to review. When the classifier is unsure, the message goes to a human, not an agent.
Payment, legal, angry-customer, or sensitive-data items always require approval before any action is taken.
Unknown senders asking for credentials, payment, or secrets get risk-flagged and held for review.
Full API and integration docs ship as the beta hardens. Until then, the schema above is the stable contract, and connecting Gmail is opt-in and revocable.