The four numbers that tell you whether your AI agent is working: containment, handover rate, booking rate and not-sure rate, with healthy ranges
An AI agent produces a lot of numbers and most of them do not matter. Four do, and they only make sense together: an agent with high containment and a low booking rate is answering well and selling badly; one with a high handover rate and a high booking rate is doing the hard part right and needs teaching. Here are the four, what healthy looks like, and what each tells you to do.
Quick answer
Containment is the share of conversations the agent handled without a person; healthy is 60 to 85 percent for most small businesses. Handover rate is the share passed to a person, by reason; 15 to 40 percent is normal, and the reasons matter more than the number. Booking or conversion rate is the share of enquiries that became a booking or order; compare it with your pre-agent rate. Not-sure rate is the share of replies where the agent said it did not know; it should fall from 15 to 20 percent in week one to under 5 percent within a month as you teach. Read them together weekly.
The four metrics
| Metric | Definition | Healthy | If it is off |
|---|---|---|---|
| Containment | Conversations handled with no person involved | 60 to 85% | Low: teach from the not-sure list; check handover triggers are not too broad. Very high: check the agent is not avoiding handover it should make. |
| Handover rate | Conversations passed to a person, by reason | 15 to 40% | Read the reasons. "Asked for a person" on simple questions means answers are weak. Complaints and quotes are correct handovers. |
| Booking or conversion rate | Enquiries that became a booking, order or qualified lead | Above your pre-agent rate | Low with high containment: the agent answers but does not ask for the booking. Add the offer to the instructions. |
| Not-sure rate | Replies where the agent said it did not know | Under 5% after a month | Teach weekly from the unanswered list; write facts the way customers ask. |
Reading them together
- High containment, low booking. Good answers, no ask. Instruct the agent to offer a slot after answering a price or availability question.
- High handover, high booking. The agent qualifies well and hands over at the right moment; teach it the routine answers to lift containment.
- Falling not-sure, flat booking. Knowledge is improving; the offer or the slots may be the problem.
- Rising handover for "asked for a person". Customers do not trust the answers; audit the conversations per the monthly audit.
Supporting numbers
- First response time. Should be seconds for the agent; watch the human response time on handovers, which is where delay now lives. See measuring response time.
- Reply usage. Replies used against the monthly allowance, and the quality rating Meta reports for each number. See daily limits.
- Won and Lost with reasons. The CRM view that turns bookings into revenue and tells you why leads are lost.
Vanity metrics
Total messages sent, conversation count on its own, and "satisfaction" from a thumbs-up prompt tell you little. A busy agent is not a useful one, and customers rarely rate. Judge the agent on what it contained, what it booked, and what it did not know.
The weekly routine
Monday: read the digest, check the four numbers, answer the not-sure list, and read any handover reason that rose. Ten minutes. The teach loop is in reducing repeat questions, and the ROI arithmetic in the ROI worksheet.
Frequently asked questions
What is containment rate for an AI agent?
The share of conversations handled without a person. 60 to 85 percent is healthy for most small businesses.
What is a normal handover rate?
15 to 40 percent. The reasons matter more than the number; complaints and quotes are correct handovers.
How do I measure whether the agent increases bookings?
Compare the share of enquiries that became bookings before and after, from your CRM's Won stage.
What should the not-sure rate be?
It starts around 15 to 20 percent and should fall under 5 percent within a month of weekly teaching.
Which metrics should I ignore?
Total messages, raw conversation counts and thumbs-up satisfaction prompts.
Four numbers, every Monday
Containment, handover, bookings and not-sure, in the digest. 7-day free trial.
Sources and further reading
- Harvard Business Review, The Short Life of Online Sales Leads (Oldroyd, McElheran, Elkington, 2011) · firms that responded within an hour were about seven times more likely to qualify the lead
- Lead Response Management Study (2007) · the five-minute window for contacting a new lead
ReplyKit is an independent product and is not affiliated with or endorsed by Meta or WhatsApp. This article is general guidance for running a small business, not legal, financial or regulatory advice.