Deflection rate is the share of inbound customer support conversations resolved without a human agent. It is the headline metric for any AI support investment, and one of the most often misrepresented.
How to calculate deflection rate
The honest formula is simple: divide the number of conversations closed without a human agent ever responding by the total number of inbound conversations in the same period. Some teams add nuance — only counting conversations where the customer did not later return with the same issue — but the base formula is enough to start.
Deflection rate = conversations resolved without a human / total inbound conversations
Why vendor numbers are misleading
Vendors quote deflection rates of 70%, 80%, sometimes 90%. These numbers are usually computed from a denominator that excludes conversations the AI never tried to answer, or that counts any AI message as a "deflection" regardless of whether the customer was actually satisfied. The honest denominator is total inbound — every conversation that came in, regardless of channel or topic.
A realistic target
For SMB and mid-market support teams, real-world deflection rates land between 20% and 60%. The upper half of that range is achievable when:
- A large share of inbound questions are repetitive (order status, password reset, plan questions).
- Your knowledge base is up-to-date and well-structured.
- AI has access to live data — order status, account state, subscription details — not just static help articles.
- You have clear rules for when AI should escalate to a human.
If your conversation mix is mostly account-specific or transactional, expect to land in the 20–35% range. That is still a real win — it means a third of your conversations no longer need an agent.
How to improve it
Three levers move deflection rate. First, expand the knowledge base — every well-written article unblocks a category of questions. Second, give AI access to live data through integrations so it can answer "where is my order" and "when does my plan renew" with real values. Third, audit AI conversations weekly: look for cases where the AI tried but the customer escalated, and add the missing knowledge.
Bottom line
Track deflection from your own data, not a vendor dashboard. Aim for 20–35% in the first quarter, then push toward 50%+ as your knowledge base and integrations mature. Anyone promising 80% on day one is selling you on a denominator, not a result.