Fallback rate

Fallback rate: definition, calculation and benchmarks

Definition:  
The fallback rate is the metric that shows how often a chatbot or AI agent can't resolve or understand a request on its own and therefore ends the conversation, hands it to a human or gives a stock reply such as "Sorry, I didn't understand that". It is given as a percentage of all conversations and is considered a key indicator of a system's quality.

In DACH customer service, a value below 15 percent is considered solid and below 10 percent very good. The following section shows how the metric is calculated, what causes it and how to lower it.

How is the fallback rate calculated?

The formula is: number of fallback cases divided by the number of all conversations, multiplied by 100. With 1,000 conversations a month and 80 fallback cases, the fallback rate is 8 percent. The lower the value, the more reliably the system works in practice.

It's important to distinguish it from the automation rate: the automation rate measures how many requests were successfully resolved. The fallback rate measures the opposite – failure. The two values complement each other but aren't simply inverses, as some conversations are deliberately passed to agent assist or human employees without any technical comprehension problem. The article efficient customer service automation with AI shows how to achieve a high level of automation in practice.

Closely related but not identical is the containment rate: the share of conversations the bot completes without human escalation. This reveals an important difference in practice: a high containment rate doesn't automatically mean a low fallback rate. According to Zendesk's 2026 CX Trends study, customer satisfaction with AI-powered support fell by 6 points between 2024 and 2026, while the resolution rate rose by 18 points over the same period. A rising resolution rate with falling satisfaction suggests that conversations are being counted as "resolved" without the concern actually having been settled.

What causes fallbacks?

Typical causes of a fallback are unclearly phrased user requests, missing data in the knowledge base, topics outside the trained scope and technical errors when connecting to backend systems such as Shopware or Zendesk. Dialect, slang or ambiguous phrasing also increase the fallback rate, especially in systems with a weak NLP (natural language processing) foundation. An AI agent differs from classic, rule-based chatbots here: it can understand context across several messages and thus reduce fallbacks caused solely by rigid decision trees. The article AI agents vs. chatbots explains the difference in more detail.

What fallback rate counts as good?

In DACH customer service, values below 15 percent are considered solid and below 10 percent very good. Values above 25 percent usually indicate an incomplete knowledge base or too narrow a topic coverage. This order of magnitude matches international benchmarks: according to an overview of chatbot fallback values, classic rule-based chatbots often have a fallback rate of 30 to 50 percent, while modern AI systems based on large language models typically lower this to 5 to 15 percent because they understand context and can handle rephrasing. When fallback values are high, providers like melibo rely on retrieval-augmented generation – access to their own, current knowledge data instead of pure model knowledge – to lower the rate.

How do you lower the fallback rate?

You lower the fallback rate by regularly evaluating failed conversations, expanding the knowledge base with questions that were actually asked and defining clear escalation paths to human employees. A good AI agent also uses fallbacks as a learning signal: every failed request reveals a gap in the knowledge base that can be closed in a targeted way. The article AI agents in service: how they change processes offers an introduction to how such systems work in principle.

Conclusion

The fallback rate shows how robust a chatbot or AI agent is in real operation, not in a test run with selected sample questions. A value below 10 percent is considered very good in DACH customer service, but what matters is less the individual percentage than the trend over time. A high fallback rate isn't just a warning sign but also an indication of where the knowledge base still has gaps. Companies should therefore not look at the metric in isolation but evaluate it together with the automation rate and containment rate to get a complete picture of system quality.