Plan requirement
| Subscription | Suite Team or higher, Support Team or higher |
| Access | Admin |
Seven numbers that describe an AI agent. Six are on the reporting dashboard under their own names; the seventh is the one that tells you the most and you have to assemble it.
| What you want to know | Where it comes from |
| How much is the agent resolving? | Automated resolutions, shown as a percentage as well as a count. |
| How often does it hand over? | Assisted escalations overall, and Escalated conversations per contact reason. |
| Did it understand at all? | Understood conversations, against Unassisted conversations: the ones that only produced system replies. |
| Did the customer leave? | Drop-offs and % Drop-off rate: conversations that were neither resolved nor escalated. |
| Were customers happy with it? | BSAT score, the ratings of 4 and 5 as a share of responses, with Ratings requested and Ratings given beside it. |
| Which articles is it leaning on? | Usage per knowledge source: the share of conversations that used that article. |
| Do people come back with the same question? | Not on the dashboard. You assemble it from tickets raised shortly after a resolution. |
Read them in pairs
A high resolution figure with high repeat contact means customers are giving up rather than being helped. Either number alone tells the wrong story, and the pair is unambiguous. That is why the row nobody can read off a dashboard is the one worth the effort.
The quicker version of the same check is the drop-off rate, which is on the dashboard. Every conversation ends resolved, escalated or dropped, so a resolution rate that looks strong beside a large drop-off share is describing a queue that emptied because people walked away.
BSAT is the quality measure people overlook, because it sits apart from your ordinary satisfaction score. Watch Ratings given alongside it for the same reason you watch the response rate on CSAT: a fine score from almost nobody is not a fine score.
Two that mislead on their own
Total conversations rises with traffic and says nothing about quality.
Automated resolutions as a raw count is useful for billing and weak as a measure: a busy month looks like an improvement. Read the percentage instead, which the dashboard gives you.
Getting the numbers out without the download button
The dashboard has an Export CSV button, and pressing it every Monday is a job nobody keeps up. There is a data export API that returns the same reporting data, conversation by conversation, as JSON.
Three things decide whether it is worth wiring up. It refreshes once a day, at midnight UTC, so it is a reporting feed and not a live one. The download links it hands back expire a day after you ask for them, so whatever you build has to ask again rather than store a link. And it draws on the same API rate limit as everything else in your account, so a heavy nightly job competes with your integrations.
One more thing decides whether the figures match your own reporting: an export for a date holds the conversations that ended on that date, counted in UTC. A conversation that runs past midnight lands on the following day, and for a business running on Dutch time the day boundary sits an hour or two inside your working day. Compare like with like, or the daily totals will not reconcile with anything else you count.
History reaches back to 1 January 2024. If you need older than that, or want us to build the feed with you, ask us.
Compare against before, and read transcripts
Whatever the agent handles, compare it against the same period before it existed. That is the comparison the business is asking about.
And read transcripts alongside the numbers. Numbers say something changed; transcripts say what. Ten conversations a week keeps the figures interpretable, and it is the habit that keeps a deployment improving.
Comments
0 comments
Article is closed for comments.