How to classify email replies: six readings, not one label
Classify a reply on several separate questions rather than with one label: who the sender is to you, what the message is about, what they want, how they feel, how urgent it is and how heated it is. Each question drives a different decision, so one combined label always throws something away. Email Digit reads every inbound reply on those six axes, using a fixed vocabulary you can see.
Why one label is not enough
Most reply sorting starts with a single field: positive or negative, interested or not. Take a real reply: “My card was charged twice and nobody has answered my last two emails.” It comes from a paying customer, it is about billing, it is a complaint, it is negative, it is urgent, and the writer is fairly angry. Collapse that into “negative” and you lose the parts that decide what happens next: who should answer, how fast, and in what tone.
The opposite case matters as much. “Love what you are building, can we get a call in this week?” is positive and enthusiastic. If it is from a lead you emailed, it is the reply you were waiting for. If it is from a stranger who wants to sell you something, it is a pitch. A positive/negative label scores both the same.
The six questions worth asking of every reply
Whether you sort replies by hand or with software, these six questions cover what routing needs. They are independent of each other, which is the point: a calm message can be urgent, and an angry one can be about nothing important.
| Question | What it decides |
|---|---|
| Who is the sender to you? | Which team owns it, and whether it is a lead at all |
| What is it about? | Which process it belongs to: sales, support, billing, legal |
| What do they want? | The next action: book a call, answer a question, remove them |
| Sentiment | The tone your answer needs |
| Urgency | How soon it has to be answered |
| Intensity | How carefully it has to be answered |
If you do this by hand, write the allowed answers down. A shared list of categories is what makes two people sort the same message the same way, and it is what lets you count anything later.
The vocabulary Email Digit uses
Email Digit reads each inbound reply and fills in all six at once. The allowed answers are fixed, so the same message cannot come back with a label nobody has seen before:
- Sender type (9): lead, customer, vendor, partner, applicant, press, internal, spam, unknown.
- Category (14): sales, support, billing and payment, complaint, refund and return, scheduling, contract and legal, feedback and praise, partnership, recruiting, unsubscribe, out of office, wrong contact, other.
- Intent (8): interested, not now, question, complaint, unsubscribe, out of office, wrong contact, other.
- Sentiment: positive, neutral or negative.
- Urgency: low, normal, high or critical.
- Intensity: a number from 0 (calm) to 100 (furious).
Two details are worth noticing. “Applicant” and “press” are sender types of their own, so a job application or a journalist’s question is never filed with your sales leads. And intensity is separate from urgency: an out-of-office note is low on both, a polite letter about a contract dispute can be high urgency and low heat, and a furious message about a typo can be the other way round.
The AI is told to judge substance rather than politeness, so a courteous message that is really a dispute is read as a dispute. Its answer has to fit the vocabulary above; anything outside it is replaced with a neutral default such as “other” or “unknown”, never stored as a new label.
What the six readings feed
The readings are not only labels on a screen. Together with your history with the sender, they produce a priority score from 0 to 100 and a set of risk flags, which is how the inbox decides what you see first. They also choose which reply playbook drafts a suggested answer, and an automation can be set to fire on sender type, category, intent, sentiment or urgency. The priority score is fixed arithmetic over these readings, not a second guess by the AI.
Limits worth knowing
- AI classifications are capped per plan. Each plan includes a set number of AI classifications per 30-day period. During the 14-day trial the cap is 500, or your plan’s own cap if that is lower.
- Past the cap, replies are still read, by keyword rules. Nothing is dropped. Keyword rules fill in intent, category, sentiment, urgency and intensity, but they cannot tell who the sender is, so the sender type reads “unknown” until the allowance resets. Each of those replies is marked KEYWORD-SCORED, and the inbox shows a banner with the reset date.
- Unsure readings are marked for review. When the classification is less than confident, the reply is flagged “Needs review” with the reason, rather than quietly treated as certain.
- No accuracy figure. We have not published one, because we have not measured one we would stand behind. When a reading is wrong, you can correct it in the inbox, and the correction can become a rule for that sender or domain.
For the wider triage routine, read Every reply, sorted before you open the inbox, or see how replies are read.