The three numbers everyone reports are deflection, escalation and containment, and all three can be improved by making the agent worse. Deflection rises when customers give up. Escalation falls when the handoff is hard to find. Containment counts an ended conversation as a success. None of them describes whether the customer got what they came for, which means the dashboard needs at least one number that only comes from reading transcripts.
Deflection rate
What it counts: conversations that ended without a human involved, as a share of total conversations.
What it hides: the difference between resolved and abandoned. A customer who got their answer in two turns and a customer who typed three increasingly annoyed messages and closed the window both count as deflected. So does a customer who then phoned instead.
Make it useful by: subtracting conversations that produced a case, a call or a repeat session within seven days. That adjusted number moves differently and is much harder to game.
Escalation rate
What it counts: conversations transferred to a person.
What it hides: whether escalating was the right call. A low escalation rate can mean the agent is competent or that the escape hatch is buried. A rising escalation rate often means the agent has got better at recognising what it should not attempt, which is a quality improvement that looks like a regression on the chart.
Make it useful by: splitting it by trigger. User requested, out of scope, repeated failure, action error. Four very different stories currently averaged into one number. Repeated failure is the one to drive down. User requested may be fine.
Containment
What it counts: broadly, conversations the agent handled end to end. Definitions vary between products and reports, so establish exactly what your figure includes before quoting it to anyone.
What it hides: everything about correctness. An agent that confidently gives a wrong answer contains perfectly. This is the metric most likely to be presented at a steering committee and the least connected to whether the thing works.
The numbers worth adding
- Repeat contact within seven days. The single best proxy for whether the conversation actually resolved anything. It needs no transcript reading.
- Turns to resolution. Track the distribution, not the mean. The long tail is where the loops live.
- Action failure rate, per action. Flows throw errors and the agent often apologises without telling anyone. This is ordinary monitoring and it is frequently absent.
- Sampled transcript accuracy. Fifty conversations a week, read by a human, scored correct or not against what the customer asked. Slow, boring, and the only direct measure of quality you will get.
- Refusal rate on in-scope questions. Cases where the agent should have helped and did not. Invisible everywhere else.
Reading them together
Single metrics mislead. Combinations tell you things:
High deflection with high repeat contact means the agent is ending conversations without resolving them. This is the most common bad pattern and it looks excellent on a deflection-only dashboard.
Low escalation with falling satisfaction means customers cannot get out. Check how the escape hatch is worded and how many phrasings trigger it.
High containment with a high turn count means people are working hard to get answers. Containment is being earned through customer effort, not agent competence.
Stable numbers with rising complaints means something changed that your instrumentation does not cover. Model updates happen on a schedule you do not control, so keep a fixed regression set and rerun it rather than trusting production aggregates alone.
Practical instrumentation
Conversation transcripts and session records are data in the org, which means they are queryable and reportable like anything else. What is captured, where it lives and how long it is retained depends on the channel and the release, so confirm the current object model in documentation before building reports on assumed field names.
Two things to set up early, because retrofitting them is painful. Store the escalation reason as a value you can group by, not as free text. And tag conversations with the topic that handled them, so you can see which parts of the agent are carrying the volume and which were built for nobody.
Frequently asked
What is a good deflection rate? There is no useful benchmark, because it depends entirely on what you routed to the agent. A narrow agent on five well-documented intents will beat a broad one and be more valuable.
Should we survey customers after an agent conversation? Yes, and expect low response rates skewed towards the annoyed. Useful as a signal of change over time, not as an absolute score.
How much transcript reading is enough? Enough that someone on the team has read a real conversation this week. Teams that stop reading transcripts stop noticing failures within about a month.
