Originally published at ictdesk.net
Here is the trap a lot of support teams are walking into with AI this year. Deflection rate looks great on a dashboard, ticket volume drops, the bot handled it, everyone claps. But deflection only counts the conversations a human did not touch. It says nothing about whether the customer's problem got solved. Optimise for the wrong number and you can cut volume while your satisfaction score quietly falls.
The two numbers are not the same
Deflection counts a ticket a human skipped. Resolution counts a problem that got solved. A bot can deflect a conversation by answering confidently and wrongly, and the customer gives up rather than pushing through, which shows up as a win and a lost customer at the same time.
This is not a small risk. Forrester expects about three in ten firms to actually damage their customer experience in 2026 through poorly built AI self-service, and surveys keep finding that most customers wish companies would stop bolting AI onto support. The technology is not the problem. Measuring the wrong thing is.
What makes AI actually resolve a ticket
The teams reporting the highest real resolution rates all describe the same handful of things, and none of them is a cleverer way to avoid a human.
First, the agent is grounded in your knowledge, so answers come from your docs and past tickets rather than a guess. Second, it can take action, look up an order, reset a setting, change a status, instead of only describing the steps. Third, it does a clean handoff when it is stuck, passing the full context to a person early rather than looping the customer. And fourth, you measure resolution and satisfaction, not deflection, so a bad answer cannot hide as a ticket the bot technically handled.
What good looks like
Set expectations honestly. Early AI deployments land around thirty to fifty percent resolution, mature workflows reach fifty to seventy, and deeply integrated agents that can take action hit seventy to eighty-five on well-scoped issues. Those are resolution numbers, not deflection numbers, and they only get there when the AI is wired into your systems and your help desk is treated as a knowledge and process problem, not just a software purchase.
FAQ
Is deflection a useless metric? Not useless, just incomplete. It tells you volume moved off humans. Pair it with resolution and CSAT so you know the volume moved because problems got solved.
Why does chasing deflection hurt CSAT? Because it rewards answering rather than solving. A confident wrong answer counts as a deflection and leaves a frustrated customer.
What is a realistic AI resolution rate? Thirty to fifty percent early on, up to seventy to eighty-five for integrated, action-taking agents on well-defined problems.
Does an AI agent replace my team? No. It resolves the well-scoped, repetitive tickets and hands the rest to people with full context, which is where the good numbers come from.
How do I start measuring resolution? Tag whether the customer's issue was actually solved, tie it to a follow-up satisfaction signal, and stop treating an untouched ticket as a solved one.
Deflection is easy to celebrate and easy to fake. Resolution is the one your customers actually feel, so it is the one worth building around.
Top comments (0)