Resolution rate is the only support metric that matters

Most support AI is measured by deflection: how many conversations never reached a human. It is an easy number to move and a bad number to trust. A customer who gave up after three unhelpful article links counts as "deflected." So does one who churned.

Resolution is a higher bar

We measure a conversation as resolved only when all three are true:

• The customer's request was completed — refund issued, setting changed, question answered with a cited source.
• The customer did not reopen the conversation within 7 days.
• No human touched it (otherwise it counts as a handoff, which has its own metric).

That definition is deliberately strict. It is the reason our public number is 70% and not 95% — and the reason CSAT goes up when agents take over tier one, instead of quietly collapsing behind a prettier dashboard.

What it changes in practice

When you optimize for deflection, the agent learns to end conversations. When you optimize for resolution, it learns to finish jobs. Those are different products:

deflection:  "This article may help → [link] → conversation closed"
resolution:  "Duplicate charge found on inv #1042 → refund issued → customer notified"

The second one is auditable. Every action lands in the log with the evidence the agent acted on, so when your finance team asks why a refund happened, the answer is one click, not an archaeology project.

If you take one thing from this post: ask any support-AI vendor how they define "resolved," and what happens to the number when a customer replies "that didn't help." The answer tells you what their product is actually optimizing.