Reply classification and workflow reactions
Automation 12 min
0/21
Business problem
When a campaign grows, the inbox drowns: 30 replies a day, each needs a decision — who's interested, who's asking, who's unsubscribing. Without classification a salesperson reads everything in order and the most important replies (INTERESTED) wait the longest. Without automation the reply class gets lost between email and a notebook.
Proven method
The classification workflow in 3 steps:
- FETCH: a reply (through the email API or an email tool) enters the system,
- CLASSIFY: AI reads the reply and returns JSON: {
category
:INTERESTED
,confidence
: 0.95,summary
:wants a demo, asks about price
,next_action
:scheduling
} — categories: INTERESTED (they want to continue), NOT_INTERESTED (a refusal), QUESTION (they ask), OBJECTION (an objection), OUT_OF_OFFICE (pause), UNSUBSCRIBE (they unsubscribe), WRONG_PERSON (the wrong person), REFERRAL (a referral), - REACT: a workflow by category: INTERESTED → an alert + notification + optionally a calendar link; QUESTION → an AI response draft → human approval → send; OBJECTION → an AI response proposal from the objection library → human; OUT_OF_OFFICE → pause the sequence for [X] days; UNSUBSCRIBE → remove from the campaign + mark; WRONG_PERSON → AI proposes the right role → a research task; REFERRAL → a thank you + ask for the contact. Rules:
- confidence in JSON (low confidence < 0.7 → always a human),
- treat NOT_INTERESTED with respect (it can be 'not now' — mark for nurture instead of aggressive selling),
- human-in-the-loop: a reply is always approved by a human before sending (even an AI draft),
- classification without an action is useless — every category has a defined reaction.
Market example
- BAD
- Manual, inconsistent execution or chaotic automation lacking data validation.
- GOOD
- A reply: 'Hi, thanks for the email. Yes, we'd like to see a demo. How much does it cost?' → AI: {
category
:INTERESTED
,confidence
: 0.97,summary
:asks for a demo and a price
,next_action
:scheduling
}. Workflow: an alert to the salesperson + a response draft ('great, [link] to the calendar, price from X — happy to discuss during the demo') + a task 'book the demo'. A reply: 'I'm on vacation until the 20th' → {category
:OUT_OF_OFFICE
,confidence
: 0.99, ...} → a 14-day sequence pause. A reply: 'No, thank you' → NOT_INTERESTED → mark + nurture in 90 days.
Ready-to-use prompt
You are a reply classification architect. Context: I have a campaign: [PRODUCT]. I get [X] replies/day. My categories: [LIST — use the 8 from the course]. Email tool: [TOOL]. CRM: [CRM]. Task: Design the classification system: 1. A JSON schema: categories (8) + confidence + summary + next_action (with types) 2. A classification prompt: how to ask AI for a category (rules: what each category means, what to watch out for) 3. Confidence thresholds: what to do at low certainty (< 0.7) 4. Workflow reactions: a table category → action (alert/draft/pause/remove/research) 5. NOT_INTERESTED: how to treat it (nurture vs removal) — a rule 6. Human-in-the-loop: which replies need a human before sending 7. Test: 6 sample replies → categories (practice) Format: 7 sections.
Model rekomendowany: Claude 3.5 Sonnet / GPT-4o Copy and paste into ChatGPT, Claude or Perplexity
ALTERNATYWNE SCENARIUSZE
Variations & alternative scenarios
Basic version: 4 categories (interested/not/question/ooo).
Full version: 8 categories + confidence + reaction automation.
Practical exercise
Write the classification JSON schema (8 categories). Define the workflow reaction for each category (a table). Test on 6 sample replies (simulating AI).
Zapisano w przeglądarce