# Reply classification

> How incoming replies get sorted into intents, why natural language resists clean sorting, and what one misclassification costs an outreach system.

**Incoming replies are sorted into intents — interested, not interested, ask later, referral, out-of-office, unsubscribe, and bounce-as-reply — so automation can act on each differently. Classification is hard because language is ambiguous — sarcasm, multi-intent replies, and forwarded threads all mislead. Misreads are costly, so systems trade precision against recall deliberately.**

_Status: Reviewed — human-edited, facts not yet confirmed against the product._

## What it is

Reply classification reads each incoming response to outreach and assigns it an intent.
The working taxonomy is stable across the industry:

- **Interested** — wants to talk; route to a human fast.
- **Not interested** — a clear no; stop outreach and record the outcome.
- **Ask later** — a timing objection; pause and resurface at the named time.
- **Referral** — points to a colleague; a new contact, not a dead end.
- **Out-of-office** — automatic; outreach should continue after the return date.
- **Unsubscribe** — a legal signal, not a preference; suppress everywhere, permanently.
- **Bounce-as-reply** — a delivery failure arriving as a reply; mark the address invalid.

## Why it works this way

Each intent demands a different action, and several are opposites: an out-of-office
should not stop outreach, an unsubscribe must. A single "replied" flag cannot drive
automation safely, so classification sits between the inbox and every downstream rule —
stopping, pausing, routing, suppressing.

## How it behaves in practice

- **Language misleads.** "Great, just what I needed — another cold email" is positive words carrying negative intent.
- **Replies mix intents.** "Not the right person, try Ana — and take me off your list" is a referral and an unsubscribe in one message.
- **Forwarded threads confuse attribution.** The reply may come from someone you never contacted, quoting the person you did.
- **Automatic replies mimic humans.** Out-of-office messages with personal tone and personal replies with template structure cross in both directions.
- **Precision and recall trade off.** Tightening "interested" precision drops warm-but-vague replies; loosening it floods humans with false positives.

## Common misconceptions

| Belief | Reality |
| --- | --- |
| "Any reply means stop the sequence" | An out-of-office should not stop outreach; only genuine responses should |
| "All misreads cost the same" | Continuing past an unsubscribe risks compliance; stopping on an out-of-office loses one prospect |
| "Keyword rules are enough" | Sarcasm, mixed intents, and forwarded threads defeat keyword matching |
| "High accuracy means safe automation" | Aggregate accuracy hides which intents fail — one rare class can carry most of the cost |

## How Reply implements this

Reply's [reply management capability](/capabilities/reply-management) sorts replies
into inbox categories automatically: an AI analyzes each reply and assigns one of
**Interested**, **Not interested**, **Do not contact**, **Not now**, or **Forwarded** —
plus any custom categories you create. Only one category is allowed per thread; the
single exception is **Meeting intent**, a subcategory that can sit alongside a category
(Interested + Meeting intent). Categorization covers email, LinkedIn, and LinkedIn
voice-message replies, and categories can drive triggers — move to another sequence,
pause, and similar actions. Two operational details matter:
[reply detection](/behavior/reply-detection) usually takes up to 3-4 hours to sync a
response, and the categorization AI understands English only, so replies in other
languages need a manual category. Out-of-office replies are handled automatically by
default — the contact is paused with an Out-of-office status and resumed after a set
number of days, because the system cannot read the return date from the email.
Conversations are stored per contact
([conversations specification](/specifications/conversations)), and the human handoff
is covered in [process positive replies](/workflows/process-positive-replies).

## Where this breaks down

No classifier reads minds — ambiguous replies exist that two humans would label
differently, and any automated system inherits that ceiling. The costly failures cluster
in the rare classes: unsubscribes and legal requests are a sliver of volume but carry
most of the risk, which is exactly where sampled human review belongs.

## Related

- [The AI SDR model](/learn/ai-sdr-model)
- [Sequence design](/learn/sequence-design)
- [Reply management in Reply](/capabilities/reply-management)
- [Reply detection rules](/behavior/reply-detection)
- [Why a reply was not detected](/troubleshooting/reply-was-not-detected)

## Build with Reply

- REST API: [docs.reply.io](https://docs.reply.io/api-reference/introduction) — inbox and conversation operations
- MCP: [agents.reply.io/mcp](https://agents.reply.io/mcp) — inbox categories and reply handling