The 30-second answer

Same dataset, two very different numbers: 100 customers start a chat, 60 of them never reach a human (deflection rate = 60%), but of those 60, only 35 actually got their problem solved (resolution rate = 35%). Deflection counts conversations that ended. Resolution counts problems that were fixed.

How to measure it honestly

Two signals together. Either alone misses something.

  • The customer explicitly said it was solved. A CSAT survey response, a thumbs-up at the end of the chat, a "thanks that worked." Explicit confirmation.
  • The customer did not come back about the same thing within 72 hours through any channel. Implicit confirmation — the absence of recontact.

Why both: CSAT misses the customers who don't answer the survey, and the customers who don't answer the survey are disproportionately the ones who didn't get help. Recontact misses the customers who switched providers or complained externally instead of contacting you again. Together, the two cover most of the field.

Why CFOs care about this number, not containment

Containment tells you how much labour the AI absorbed. Resolution tells you how much customer value the AI created. The first reduces a cost line. The second protects revenue.

At any company where churn is the real risk, resolution matters more than containment. A 70% containment rate sounds great until you discover that the 30% recontact buried inside it is correlated with a 5-point drop in renewal rate. That number — the dollar impact of the resolution gap — is the question a finance review actually wants answered.

Where vendors report it

A few are starting to. Intercom Fin has begun surfacing a "real resolution" metric in some dashboards. Most other vendors still report only containment by default. If you have ticket IDs across channels (chat, email, voice, social), you can compute resolution yourself in a spreadsheet without vendor cooperation — and you should.

What "good" looks like

  • Above 70% — the AI is genuinely solving problems for most customers it handles.
  • 50-70% — the AI is helpful but the recontact tail is large. Tune your knowledge base; tighten escalation criteria.
  • Below 50% — the dashboard is telling you a much better story than your customers would. The AI is functioning as a deflector, not a resolver. Either pull back the autonomy or invest seriously in fixing the underlying gaps.

The honest framing for vendor pitches

When a vendor quotes containment, ask: "What does the resolution number look like on the same dataset?" If they can produce it, you've found a vendor that measures honestly. If they can't, you've found out something useful about how they think about the work.