Deflection Rate
The percentage of incoming support conversations the AI chatbot resolves without human agent involvement. The primary KPI for AI chatbot ROI. Industry benchmark: 40–65% for well-configured deployments.
Deflection rate is the headline metric for AI chatbot ROI. A 60% deflection rate means 60% of your incoming support tickets were handled entirely by the AI — no human agent involved. The remaining 40% escalated to a human.
Why it matters
Deflection rate drives the math that justifies AI chatbot investment. If a human CS agent handles 10 tickets per hour at $35/hr, each ticket costs $3.50 in labour. A chatbot deflecting 1,000 tickets/month at 60% saves 600 tickets Ã- $3.50 = $2,100/month in gross labour cost.
The gotcha: deflection rate benchmarks vary widely by source. Vendors claim 70–90%; independent benchmarks show 40–65% for teams in the first 6 months of deployment. The gap is explained by:
- Vendor benchmarks are from mature deployments — 12–18 months in, with heavily tuned knowledge bases
- Independent benchmarks include the tuning period — month 1–3 deflection is often 20–35%, rising as the KB improves
- Definition differences — “deflection” means “no human involvement” in some platforms; “customer marked satisfied without escalation” in others. These produce different numbers from the same conversation logs.
In our test methodology
We measure deflection rate against a standardised set of 50 real support transcripts drawn from 3 industries (DTC ecommerce, SaaS, service business). We replay these against each platform’s AI with an identical knowledge base. Our benchmarks:
| Platform | Deflection rate (our 50-transcript test) |
|---|---|
| Intercom Fin | 76% |
| Tidio Lyro | 72% |
| Freshchat Freddy AI | 58% |
These are controlled-test numbers. Real-world performance depends heavily on knowledge base quality, query type distribution, and configuration. Expect your month-1 deflection rate to be 15–20 percentage points below these benchmarks while your KB matures.
Deflection rate vs containment rate
These terms are often used interchangeably but have subtly different meanings in different platforms:
- Deflection rate — the AI handled the conversation without human involvement (could mean the customer got an answer, or the customer abandoned the chat)
- Containment rate — the AI kept the conversation “contained” — no human was involved AND the customer interaction was completed (not abandoned)
Containment rate is a stricter metric and generally runs 5–15 percentage points lower than deflection rate for the same platform. Intercom calls it “containment rate” in some of its documentation; Zendesk uses “deflection rate.” Always clarify the definition before comparing vendor benchmarks.
Realistic month-1 expectations
Honest expected deflection rate for a first AI chatbot deployment, by KB quality:
| KB quality | Month-1 deflection | Month-6 deflection |
|---|---|---|
| No KB (AI only from training data) | 10–20% | 20–30% |
| Basic FAQ document (5–10 articles) | 25–35% | 40–50% |
| Comprehensive KB (50+ articles, synced) | 40–55% | 60–75% |
| Mature KB + product data integration | 50–65% | 70–85% |
Don’t expect “90% deflection in month 1” — that’s the marketing-page lie. Real operators building serious AI chatbot deflection see 40–65% at the 6-month mark with dedicated KB investment.