The deflection metric that lies to you

Most support orgs report a deflection rate somewhere between 30 and 50%. I’ve yet to see a team where that number, calculated honestly, comes in above 22.

The gap is in the counting. A ticket that the AI assistant “resolves” but the customer reopens 36 hours later is still in the deflection bucket. Same goes for the ticket where the customer abandons the channel and switches to email — the original conversation closed, so the reporting layer marks it as deflected. The reporting layer doesn’t know the customer is still trying to solve the same problem. It just sees a closed ticket.

What an honest count looks like

A useful deflection metric tracks the customer’s full journey on a single intent, not the lifecycle of any one ticket. If a customer pings the chat bot, gets an answer, and doesn’t escalate within seven days on the same intent, that counts as deflection. Anything else — reopen, channel switch, new ticket on the same root cause within the window — does not.

Under that definition, the teams I’ve worked with land somewhere between 14% and 22% real deflection on their first deployment. After a quarter of tuning the knowledge base against actual reopen patterns, that number climbs into the high twenties. None of them get to 50.

Why the inflated number matters

Support leaders presenting 45% deflection up to the CFO are setting up an expectation the program can’t survive. Six months later, AHT on escalated tickets has gotten worse — the AI is now handling the easy stuff, which makes the agent queue denser by composition. The executive sponsor starts asking why the projected budget impact didn’t land. The honest answer is that the 45% was a measurement artifact, but by then the program’s reputation is the harder problem to fix than the metric ever was.

The teams that hold their executive sponsorship longest are the ones that report 18% deflection honestly in month one, walk it up to 24% by month four with the work to back it up, and avoid promising the world before the program has earned it. Less impressive on the first slide. Much more durable.

What to actually measure

Four numbers are worth tracking instead of one inflated headline figure: reopen rate on AI-resolved tickets (broken out by intent class), channel-switch rate within 7 and 14 days, CSAT on AI-resolved interactions compared to agent-resolved ones for the same intent, and repeat-contact rate from the same customer within 30 days. If those four are trending the right way and your deflection number is moving with them, the program is working. If your deflection number is moving and they aren’t, the deflection number is lying to you.

Our auto ticket triage playbook is built around instrumenting these honestly from the first deployment day rather than retrofitting the measurement after the program is in trouble.