What is deflection rate?

Deflection rate is the share of potential support contacts that are resolved through self-service, such as help articles, account tools or an AI assistant, before the customer reaches a staffed channel. It estimates contacts that never happened. Because those are invisible by definition, deflection is the easiest support metric to overstate, and AI agents make overstating it easier still.

Why does deflection rate matter once agents act?

Deflection is the economic case for self-service: every contact avoided is cost avoided. An AI agent placed at the front of the support journey is often justified almost entirely by projected deflection.

The difficulty is that deflection does not distinguish between a customer who found their answer and one who gave up. A customer who asks the agent how to dispute a charge, gets a link to a policy page and leaves may be deflected. They may also have abandoned the dispute they were entitled to raise. In a regulated firm, the second case is a customer outcome problem, which is why deflection is not resolution.

How is deflection rate measured in practice?

There is no single formula, which is part of the problem. Common approaches include:

• Help article views or agent sessions that are not followed by a contact within a set window.
• Customers who answer "yes" to "did this answer your question?"
• Modelled estimates comparing contact volumes before and after a self-service change.

Each overstates in its own way. A session with no follow-up contact may mean success or surrender. A "yes" from a small minority of users says little about the rest. Strong measurement pairs deflection with signals that the issue did not come back: repeat contacts in other channels, complaints and outcomes such as completed tasks.

An example from a mobile network

Take a mobile operator that adds an AI agent to its help pages for billing questions. Within weeks, billing calls fall and the agent is credited with high deflection. A month later, complaints about unexpected roaming charges rise.

The team traces it: customers asking the agent why their bill was high were given a general explanation of roaming tariffs, and many accepted it without realising a misapplied bundle was the cause. They were deflected from a call that would have found and fixed the error. The operator changes the brief so the agent checks the account for that specific fault, and hands over when it cannot.

How is deflection different from containment?

Deflection happens before a contact: the customer never opens a ticket, never starts a chat with support, never calls. Containment happens during a contact: the customer did start a conversation with an automated agent, and it ended without a person.

The line blurs when an AI agent sits on the help page itself. Many firms count the same conversation as deflected (no ticket raised) and contained (no human involved). Pick one definition per channel and write it down, or the two figures will be added together and claimed twice.

Which terms sit next to deflection rate?

• Escalation rate: the share of conversations handed to a person.
• Self-service success rate: the share of self-service attempts where the customer completed the task.
• Our guide on measuring a support agent beyond resolution rate sets out a balanced set of measures.
• An AI support agent that knows when to say no hands over rather than deflects when the case needs a person.