## TL;DR
A canned reply on the wrong intent is a confidence problem: the trigger fired on a guess, not a certainty. Gate every canned reply behind an intent-confidence threshold, and below it either ask a clarifying question or stay silent and let the human choose. A reply sent to the wrong intent is worse than no reply; the gate makes "no reply" the safe default.

## The query

```text
support agent's canned reply triggering on wrong intent: confidence gate
```

## Use this when

- Canned replies fire on wrong intents
- Customers get irrelevant automated replies
- You are tuning reply automation
- Auditing automated reply accuracy

## Not for

- Writing the reply copy
- Training intent models
- Personalizing replies
- Reply timing

## Steps

### 1. Require a confidence threshold per canned reply

Each automated reply declares the minimum intent confidence that may trigger it. High-stakes replies (refunds, cancellations) need higher thresholds than informational ones. One global threshold fits nothing well.

Expected output: per-reply thresholds set by stakes.

### 2. Disambiguate instead of guessing below threshold

When confidence is below the threshold but above the noise floor, ask: "Are you asking about X or Y?" A clarifying question converts a likely misfire into a correct routing. Silence is better than a wrong reply; a question is better than silence.

Expected output: a disambiguation path for the middle band.

### 3. Stay silent below the noise floor

Very low confidence means the intent is unknown. Do nothing automated; route to a human or the general queue. An automated reply to an unknown intent is a random reply.

Expected output: no automated reply under the noise floor.

### 4. Log every trigger with its confidence

Record which reply fired, on what intent, at what confidence. The log is how you tune thresholds: misfires cluster just above the threshold, which tells you where to raise it.

Expected output: a trigger log with confidence per fire.

### 5. Review misfires weekly and tune

Sample fired replies, label correct versus wrong intent, and adjust thresholds. New intents and vocabulary shift the confidence landscape; thresholds are maintenance, not set-and-forget.

Expected output: weekly review driving threshold updates.

## Template: the checklist

```text
Canned reply confidence gate:
[ ] Per-reply confidence thresholds set by stakes
[ ] Disambiguation questions for the middle band
[ ] Silence below the noise floor
[ ] Trigger log with confidence per fire
[ ] Weekly misfire review tuning thresholds
```

## Variant phrasings

### auto reply wrong intent

Steps 1 and 2. Threshold, then disambiguate.

### canned response misfiring

Steps 4 and 5. Log the fires, tune weekly.

### intent confidence threshold for replies

Step 1 as the direct answer.

## Why it happens

Reply automation is usually built intent-first: "when we detect X, send Y." Confidence is an afterthought, so the system acts on its best guess even when the guess is weak. Real traffic is full of ambiguous phrasing that matches multiple intents weakly. The gate adds the missing question: "how sure are we?"

## Edge cases

- Multi-intent messages: the message contains two intents. Fire on the dominant one only if confidence is high; otherwise disambiguate.
- Threshold gaming: if agents learn the threshold, some will rephrase to trigger replies. Keep exact thresholds internal.
- New intents with no history: start conservative (high threshold, human review) until the intent's confidence profile is known.
- The clarifying question itself misfiring: keep disambiguation to two options max, with a human fallback.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_R2fyuZYans4LtXQ_fM7xug