Triage a shared inbox: what it is, how urgent, and whether a draft may go out

The step

Inbox automation has two halves. Writing the reply is the half that gets all the attention. Deciding whether this message should be answered automatically at all, by whom, and how soon, is the half that decides whether the project survives contact with real customers.

• Intent from a prompt. “Classify this email as order, complaint, question or spam.” The model answers in prose, sometimes with a fifth category, and a message that is half complaint, half order lands in one bucket with no trace of the other.
• Urgency from keywords. “urgent”, “asap”, “lawyer”. Everyone writes “urgent”. The one that matters doesn’t.
• Nothing fits. The message is in the wrong language, a vendor pitch, an out-of-office. A prompt forces it into the closest category; your workflow then does something with it.

Each of those is a pick between options you know in advance. A decision model scores the options and tells you how sure it is, so the workflow can be careful exactly where it should be.

What the request looks like

{
  "model": "saina-helm-2-0.8b",
  "state": "Subject: Order 48213\n\nHi, the blender arrived yesterday with a cracked jug. I'd like a replacement sent before the weekend, I'm hosting on Saturday. If that's not possible please refund me. Receipt attached.",
  "mode": "decision",
  "threshold": 0.7,
  "min_margin": 0.15,
  "questions": {
    "intent": {
      "type": "single_choice",
      "question": "What is this message about?",
      "options": {
        "damaged":  "Damaged or defective item received",
        "status":   "Where is my order",
        "return":   "Return or refund of an undamaged item",
        "question": "Pre-sales or product question",
        "other":    "None of the above"
      }
    },
    "urgency": {
      "type": "rating",
      "question": "How soon does this need a reply?",
      "levels": ["This week", "Today", "Now"]
    },
    "auto_ok": {
      "type": "yes_no",
      "question": "Can a standard reply be sent without a person reading this first?"
    },
    "tags": {
      "type": "multi_choice",
      "question": "Which apply?",
      "options": { "refund_request": null, "deadline": "Mentions a date or deadline", "attachment": "Refers to an attachment", "upset": "Customer is upset or threatening" }
    }
  }
}

And what comes back:

{
  "answers": {
    "intent": {
      "type": "single_choice",
      "selection": "damaged",
      "reason": "accepted",
      "probabilities": {
        "damaged": 0.88,
        "return": 0.08,
        "status": 0.02,
        "question": 0.01,
        "other": 0.01
      },
      "confidence": 0.85
    },
    "urgency": {
      "type": "rating",
      "expected_level": 1.21,
      "reason": "accepted",
      "probabilities": {
        "0": 0.12,
        "1": 0.55,
        "2": 0.33
      },
      "levels": [
        "This week",
        "Today",
        "Now"
      ],
      "level": 1,
      "confidence": 0.55
    },
    "auto_ok": {
      "type": "yes_no",
      "yes": 0.41,
      "no": 0.59,
      "confidence": 0.18,
      "selected": null,
      "reason": "below_threshold"
    },
    "tags": {
      "type": "multi_choice",
      "selections": [
        "refund_request",
        "deadline",
        "attachment"
      ],
      "reason": "accepted",
      "memberships": {
        "refund_request": 0.86,
        "deadline": 0.93,
        "attachment": 0.95,
        "upset": 0.18
      }
    }
  }
}

Illustrative answer: the shape is exact, the numbers show a typical case. Note the gate question falling back: the model is not sure a canned reply is safe here, so the workflow sends it to a person with the other three answers already filled in.

Four questions, one read of the email. The intent is clear and accepted. The urgency comes back as a probability-weighted level on your scale. The tags are independent, so three can be true at once. And the question that matters most, auto_ok, came back below_threshold: that is the gate doing its job. The agent still drafts; a person presses Send.

Putting it in place

Make “none of the above” a real option

Give the intent question an explicit other option with a description, and treat a low top score as a signal too. Between the two, off-category mail stops being forced into a bucket. (Spotting the message that matches none of a hundred intents is on our roadmap; with a handful of intents and an explicit other, it works today.)

Gate on two numbers, not one

Let a reply go out automatically only when auto_ok clears your threshold and no risk tag (upset, a legal word, a refund promise) is above its own. Everything else is drafted and queued for a person. Start strict, loosen with data.

Wire it in

In n8n: Gmail or IMAP trigger → the Saina Helm node → Switch on intent, with the node’s Fallback output going to a review label. Start from the node’s support-routing template and swap in the questions above. From code, it is the same request to api.saina.run/v1/ask or to a server you run with pip install 'saina[local,server]', so the mail never leaves your infrastructure.

Use distribution mode while you are calibrating, decision mode with thresholds once you have picked them.

What it won’t do

• It does not write the reply and does not extract fields (order numbers, dates). Keep your LLM or your parser for that; Helm decides whether and where.
• It reads what you send. For a long thread, send the latest message and a short summary of the rest rather than the whole thread.
• Text only with Helm 2. If the decision depends on an attached image, use Helm 2 Omni and send the image in the same request.

Frequently asked questions

Read next