The AI sales agent metrics that matter are the ones that follow a chat through to money: conversations, handovers, baskets, paid orders and response time. Chat count on its own tells you the agent is busy, not that it is selling. This article defines five numbers to track, works through a real example from a UAE bakery and shows how to review them weekly. It is part of our guide to AI WhatsApp agents.
Why chat count is not enough
When an AI agent goes live, the first number everyone looks at is how many chats it handled. It is easy to measure and usually large, which feels good. But a high chat count can hide problems:
- customers asking the same question twice because the first answer was unclear;
- many chats ending without a basket or a request;
- handovers that nobody picked up;
- baskets created but never paid.
To judge whether the agent is helping the business, follow the path from first message to paid order, and look at where customers drop out.
Five metrics to track
| Metric | What it counts | What it tells you |
|---|---|---|
| 1. Conversations | Distinct customer chats the agent handled in the period | Demand reaching WhatsApp |
| 2. Handovers | Chats passed to staff, by reason | Where the agent’s knowledge or rules stop |
| 3. Agent orders or baskets | Orders or baskets created in the chat by the agent | Whether chats turn into buying intent |
| 4. Paid orders | Agent orders that were actually paid | Real sales, not intentions |
| 5. First response time | Time from the customer’s first message to the first reply | Whether customers are kept waiting |
For service businesses such as clinics or salons, replace baskets and paid orders with appointment requests and requests your team confirmed. The logic is the same: count intention, then count what actually happened.
Two ratios are worth calculating from these numbers:
- Chat-to-order rate: agent orders divided by conversations.
- Order-to-paid rate: paid orders divided by agent orders.
The first shows how well the agent turns questions into baskets. The second shows how well baskets turn into payment, which often depends on follow-ups and the payment step itself.
A worked example from a UAE bakery
Our b2r bakery case study shows these metrics in practice. b2r (Bake to Roots) is a UAE bakery selling breads, pastries and Mouneh. After its AI sales agent went live on WhatsApp, the case-study page reports, for the first 30 days to 1 October 2026 (source: 800website CRM):
- 45 chats
- 13 agent orders
- 4 paid orders
Reading the funnel:
- 13 of 45 chats led to an agent order, so roughly 3 in 10 conversations became a basket.
- 4 of 13 agent orders were paid, so most baskets did not turn into payment in that period.
The second ratio is where to look first. Possible reasons include customers comparing options before paying, payment links reaching people at a busy moment, or a preference for paying on collection. The b2r setup includes 2h, 24h and 48h follow-ups, which are defaults that can be changed, so the next step would be to check whether unpaid baskets received follow-ups and how customers responded.
These are early figures from one business in its first month. They show how to read the funnel, not what any other business should expect.
Reading handovers and errors
Handovers are not failures. They show the boundary of what the agent should handle. Group them by reason:
| Handover reason | If it is rising |
|---|---|
| Customer asked for a person | Check whether replies are too long, unclear or robotic |
| Complaint or refund | Look at the operational cause, such as late deliveries |
| Agent not confident | Add the missing knowledge |
| Large or custom order | A good sign: route it to sales quickly |
Errors are replies that were wrong: a wrong price, an out-of-date policy or a product that no longer exists. Count them during your weekly review and fix the source information each time. A falling error count is one of the clearest signs the agent is improving.
Reviewing weekly
A short, fixed routine keeps the numbers useful.
- Pull the same five metrics for the week, in the same format.
- Compare with last week and with your own baseline from before the agent, if you have one.
- Read ten chats, including two that ended without an order.
- Pick one improvement: a knowledge fix, a follow-up change or a payment step.
- Write it down and check its effect next week.
Keep the review to 30 minutes. One change a week, measured properly, beats a list of ten ideas that never get checked.
How Dub Ai helps
Dub Ai sends reports on sales, leads, conversations and AI usage to you on WhatsApp whenever you ask, with Excel and PDF versions. You can request a weekly sales report that compares the week with the one before. The Help Centre article on getting reports on WhatsApp lists what you can ask for. Clients use the WhatsApp assistant and the Dub Ai team rather than a separate system, and our team helps you read the numbers and choose the next improvement.
Frequently asked questions
What is a good chat-to-order rate?
There is no single figure that applies to every business. It depends on what you sell, your prices, how customers find you and how many chats are support rather than sales. The useful comparison is with your own baseline: measure the rate before the agent, then track it week by week and look for steady improvement after each change.
How often should I review the agent?
Weekly in the first two or three months, when most knowledge gaps and wording problems appear. Once errors and unexpected handovers settle, a fortnightly or monthly review is usually enough. Review straight away after any change to products, prices, policies or opening hours, because those are the moments when wrong answers are most likely.
Book your demo on WhatsApp: +971 56 552 6662
Sources
- b2r case study: online shop and WhatsApp AI agent (Dub Ai, figures from 800website CRM, 30 days to 1 October 2026)