What should a chatbot do when it cannot answer?

A well-designed fallback follows a hierarchy: first, attempt one clarifying question to recover (this resolves 25-40% of would-be fallbacks). Second, offer the closest KB article or alternative resource. Third, hand off to a human with full context. Fourth, log the failed query so it becomes the next KB article. Never just say "I don't understand" — that is a dead end.

What confidence threshold should trigger fallback?

For most deployments, 0.55-0.65 is the sweet spot. Below that, the LLM is likely guessing — risk of hallucination outweighs the value of attempting an answer. Higher-stakes industries (medical, finance, legal) should set thresholds at 0.70+. Pair the threshold with a clarification step before declaring failure to recover gracefully.

How do I prevent fallback from feeling like a dead-end?

Always offer at least two paths out: a clarifying question ("Can you tell me more about X?"), a relevant KB article, or a human handoff. Never end on "Sorry, I can't help." Acknowledge the limitation, offer concrete next steps. Across our deployment data, this pattern alone improves post-fallback CSAT by 1.4 points.

How do I measure fallback rate?

Fallback rate = (Fallback events) / (Total bot turns) × 100. Healthy SaaS chatbots run 8-15% fallback rate. Above 20% indicates poor KB coverage or weak retrieval; below 5% suggests the bot is over-confident and may be hallucinating. Pair with answer-accuracy sampling — low fallback + low accuracy is the worst combination.

AI Chatbot Fallback Strategies

Name: EzyConn
Brand: EzyConn

Every chatbot fails sometimes. The difference between a great deployment and a frustrating one is what happens next. Here's the full fallback playbook — confidence thresholds, clarification prompts, escalation, and the gap-loop that turns failures into improvements.

12 min readUpdated May 17, 2026Operations

Try EzyConn Free

Never end on "sorry"

The single biggest fallback mistake is declaring failure without offering a path forward. Even a graceful failure must include at least one of: a clarifying question, an adjacent article, a human handoff, or an alternative channel. Dead-end fallbacks tank CSAT by 1.4-2.0 points vs ones with a next step.

6 Fallback Strategies, Ranked by Recovery Rate

1. Clarifying question first

25-40% of would-be fallbacks

Instead of declaring failure, ask one specific clarifying question. "Could you tell me which plan you're on?" recovers far more conversations than "I didn't understand."

2. Offer adjacent KB content

15-25% deflection on fallback

Even if no exact match, surface the 2-3 nearest articles. Users often find their answer in an adjacent article.

3. Smooth handoff to human

Always option after 2 failed attempts

Pre-empt frustration. After the second fallback in a session, automatically offer to connect with a human. Pass full transcript.

4. Alternative channel offer

8-12% choose alternate

Some queries are better solved over phone or email. Offer the right channel when the chat clearly is not working.

5. Schedule callback

10-18% choose callback

When humans are not immediately available, callback scheduling beats indefinite waiting. CSAT for scheduled callbacks averages 4.5+.

6. Gap-loop logging

Long-term: -2-5% fallback/month

Every fallback becomes a candidate KB article. Weekly review of top fallback queries drives compounding deflection gains.

Fallback Message Templates

Clarifying

"I want to make sure I help correctly — could you tell me [specific dimension]? For example: are you on Free, Pro, or Team?"

Adjacent content

"I don't have a direct answer for that, but these articles might help: [article 1], [article 2]. Want me to walk through either?"

Soft handoff

"That's outside what I can answer accurately. Let me connect you with a team member — they'll see our conversation and pick up right here."

Channel offer

"This sounds like something better discussed by phone. Want me to schedule a 15-minute call with our team?"

After-hours fallback

"I can't fully resolve this without a teammate, and our team is offline until 9am. Want me to email a summary so they can respond first thing?"

The Gap-Loop (Long-Term Strategy)

Fallback queries are gold. Every "I don't know" from the bot is a signal of a missing KB article. The gap-loop:

1. Log every fallback. Capture the user's exact question, timestamp, and conversation context.
2. Weekly review. Cluster fallback queries by topic. Identify the top 5-10 patterns.
3. Write KB articles. Each pattern becomes a new article. Or update an existing one with the missing terminology.
4. Re-test. Run the failed queries against the updated bot. Confirm resolution.
5. Compound. Over 90 days, this single discipline drops fallback rate 8-15 percentage points.

Confidence Threshold Tuning

Set your initial fallback threshold at 0.60. Sample 100 conversations weekly: look for hallucinations (lower threshold), unnecessary fallbacks (raise threshold), and edge cases the system handled correctly. After 30 days, most deployments settle at 0.55-0.70. Highly regulated industries should run higher (0.70-0.80) to favor handoff over false-confident answers.

Frequently Asked Questions

Healthy fallback rate?

8-15% for most SaaS chatbots. Higher signals KB gaps; lower may signal hallucination.

Best first response?

A clarifying question, not a failure. Recovers 25-40% of would-be fallbacks.

Long-term improvement?

Run the gap-loop weekly. Compounds to 8-15 point fallback reduction in 90 days.

Turn fallbacks into KB articles

EzyConn surfaces the top 10 fallback queries every week, ranked by impact. Write the article, watch the rate drop.

Start Free

Last updated May 17, 2026. View more guides.