Definitions
Deflection rate = conversations closed without a human ÷ total conversations. It answers "how much work did the bot keep away from my team?" — but it can't tell a satisfied customer from a frustrated one who left.
Resolution rate = conversations where the problem was solved ÷ total conversations. It answers "how often did the customer get what they came for?" It's harder to measure, because you need a signal that the problem was actually solved: an explicit confirmation, no repeat contact on the same issue within a set window, or a positive rating.
Containment rate is a third term you'll see, mostly from contact-centre software. It usually means the same as deflection: the customer stayed inside the automated channel. It says nothing about whether they left happy.
A worked example
Say your chatbot handles 1,000 conversations in a month. 200 are handed to your team. Of the 800 no human touched, 550 end with the customer confirming the answer helped, and 250 simply end — the visitor stops replying or closes the window.
- Deflection rate: 800 ÷ 1,000 = 80%
- Resolution rate: 550 ÷ 1,000 = 55%
- The 25-point gap: 250 conversations that look like a success on a deflection dashboard, but may be customers who gave up — and some of them will come back angrier, by email or phone.
Why deflection makes chatbots look better than they are
Deflection rises when a bot is good, but it also rises when a bot is hard to escape. Hiding the "talk to a person" option, looping customers through the same suggestions, or ending conversations on inactivity all push deflection up while making the experience worse. A vendor quoting 90% in the first week is almost always quoting deflection.
Resolution behaves differently: it only goes up when customers actually get answers. That's why it's the number to put in front of your team and your leadership — and why deflection is best read alongside it, not instead of it.
What a realistic resolution rate looks like
Published benchmarks put early AI support deployments around 30–50% resolution, rising to 50–70% once workflows and the knowledge base mature. Where you land depends far more on the quality of your help content than on the AI model: most low resolution rates turn out to be missing or ambiguous articles.
How to measure resolution honestly
A few habits keep the resolution number truthful:
- Count a conversation as resolved only on a positive signal — a confirmation or rating — not on silence.
- Check for repeat contacts: the same customer coming back about the same issue within a few days means it wasn't resolved.
- Weight resolution against CSAT on those same conversations; rising resolution with falling satisfaction means answers are technically right but unhelpful.
- Read a sample of "resolved" transcripts every week, not just the escalated ones.
- Track the questions the bot couldn't answer — each recurring one is a missing article, and fixing it is the fastest way to move resolution.
How nDesk reports it
nDesk's dashboard reports resolution and deflection as separate numbers, alongside the questions that most often ended in an escalation, so you can see both how much work the assistant took off your team and how often it actually solved the problem — and what to write next to close the gap.