Which chat support KPIs matter? Learn how to measure resolution, handoff, speed, satisfaction and cost, with formulas and a simple example.
Most teams that add chat to their website, WhatsApp or Messenger start by watching one number: how many chats came in. That number tells you about demand, not about quality. If you want to know whether chat is helping customers and saving your team time, you need a small, balanced set of measures.
This guide lists the KPIs worth tracking, how to calculate each one, and how to read them together so you do not optimise one at the expense of another.
Start with the question each KPI answers
A KPI is only useful if it changes a decision. Before building a dashboard, group your measures by the question they answer:
- Is anyone using it? Volume and channel mix.
- Does it solve problems? Resolution and containment.
- Is it fast and smooth? Response time and handoff time.
- Are customers happy? Ratings and sentiment.
- Is it worth the money? Cost per resolved chat and credit usage.
- Does it make money? Leads captured and conversions assisted.
Five to eight KPIs are enough for most small and mid-sized teams. More than that, and nobody looks at the report.
Volume and demand KPIs
Conversations by channel
Count conversations per channel each week: website widget, WhatsApp, Messenger, email and so on. Channel mix shows where customers prefer to reach you and where to invest in training content.
Conversations by topic
Volume alone hides the story. Group conversations by topic (shipping, returns, pricing, login problems) and look at the top five. If one topic produces a large share of chats, that is a candidate for a better help article, a product fix or a dedicated flow.
Peak hours
Plot chats by hour and day of week. This drives staffing and shift planning for the conversations your AI agent cannot finish alone.
Resolution KPIs
Containment rate
Containment is the share of conversations the AI agent handled without a human taking over.
Containment rate = conversations with no human handoff / total conversations
It is popular because it is easy to compute, but it can mislead. A customer who gives up and leaves also looks "contained". Never read containment on its own.
Resolution rate
Resolution rate asks whether the customer's problem was actually solved. The most reliable signals are:
- an explicit thumbs-up or positive rating
- a conversation that ends without the customer repeating the question
- a human reviewer marking the chat as resolved
Resolution rate = chats marked resolved / total chats
For a reviewed sample, this is much more honest than containment.
Repeat contact rate
If the same person writes again about the same issue within a few days, the first answer probably failed. Track repeat contacts per customer or per topic. A rising repeat rate is an early warning that your knowledge base is out of date.
Unanswered or fallback rate
Count the conversations where the agent said it did not know. These are your best content roadmap: each one is a missing answer.
Speed KPIs
First response time
For an AI agent this is close to instant, so it is more useful as a measure of your human channels (email, escalations). Track it by channel and by hour.
Time to human
When a chat is handed to a person, how long until someone picks it up? This is the number customers feel most. Measure it from the moment of handoff to the first human message.
- Compare the median, not just the average, because a few long waits distort averages.
- Break it down by shift, so you can see whether evenings or weekends are under-covered.
Time to resolution
Measure from first message to resolved. Report AI-only chats and human-assisted chats separately, because mixing them hides both stories.
Satisfaction KPIs
Customer rating
A simple thumbs up or down on a reply or conversation gives you a satisfaction percentage:
Satisfaction = positive ratings / total ratings
Remember that only a minority of customers rate. Treat the figure as a signal, and read the low-rated conversations themselves.
Sentiment
Sentiment analysis classifies conversations as positive, neutral or negative. It covers customers who never click a rating, so it complements ratings. Watch for negative sentiment clustering around a topic or a time of day.
Escalation reasons
When a chat is escalated, record why: the customer asked for a person, the agent lacked an answer, the issue was sensitive, or the customer was frustrated. Reasons tell you whether to fix content, tune guardrails or add staff.
Cost and efficiency KPIs
Cost per resolved conversation
Cost per resolved chat = (platform cost + human time on escalations) / resolved chats
Include the human time spent on handoffs, not just the subscription. A hypothetical example: if a shop receives 300 chats a month and the agent resolves 200 of them without help, the cost of those 200 is just their share of the subscription. The remaining 100 carry staff time too. Seeing the two groups apart makes the trade-offs clear.
Credit or usage consumption
Many AI platforms bill by replies or credits. Track consumption per agent and per month so you can forecast and spot a runaway agent early.
Agent time saved
Estimate hours saved as contained chats multiplied by your typical handling time. Label it as an estimate in your reports, and validate the handling time by sampling real chats.
Revenue and lead KPIs
For sales-oriented chat, add:
- Leads captured per 100 conversations
- Lead quality, as judged by the sales team
- Assisted conversions, meaning purchases or bookings by people who chatted first
Be careful with attribution. Chat often influences a purchase that completes later on another device, so use these as directional rather than exact.
Putting the KPIs together
No single metric tells the truth. Read them in pairs:
| If you see | Check against | Likely meaning |
|---|---|---|
| High containment | Low ratings or high repeat contacts | Customers give up; answers are weak |
| Low containment | High ratings on human chats | Content gaps, but humans are doing well |
| Fast first response | Slow time to human | Bot is quick, but handoff coverage is thin |
| Falling cost per chat | Falling satisfaction | You may be over-automating |
A hypothetical monthly snapshot helps show how this looks in practice:
In this made-up example, the AI-only resolution share is about 63 percent. That looks healthy, but the 30 abandoned chats deserve a closer read before you celebrate.
A weekly review routine
- Open the list of low-rated and escalated chats.
- Read ten of them and note the cause of each.
- Fix the top cause: add a Q&A, update a source, or adjust the persona.
- Check next week whether that topic improved.
- 01Measure
- 02Read the weak chats
- 03Fix content or rules
- 04Measure again
Common mistakes
- Chasing containment alone. It rewards bots that block people from reaching a human.
- Averaging everything. Use medians and segment by channel, topic and shift.
- Ignoring small samples. If you have 20 ratings, one angry customer swings the number.
- Changing definitions. Decide what "resolved" means and keep it stable, or trends become meaningless.
- Never reading transcripts. Numbers show where to look; transcripts show what to change.
Doing this in AI SNS CHAT
AI SNS CHAT gives you the raw material for most of the KPIs above. It does not claim to be a full business-intelligence tool, so some measures are calculated from exports.
Volume, topics and sentiment. Under Analytics you will find Chats, Topics, Question history and Sentiment views. Topics and Question history show what customers ask most and which questions the agent struggled with. Advanced analytics require the Pro plan, so check pricing if you need them.
Resolution and satisfaction. Open Activity → Chat logs. Customers can give thumbs up or down ratings, and you can filter conversations, select many at once, then flag, escalate, resolve or delete them. You can export the logs to compute resolution rate or repeat contacts in a spreadsheet.
Time to human. Handoffs are routed using Helpdesk → Scheduling, where shifts are set per person and timezone. Handoff alerts can go to Slack, Teams, LINE WORKS, WhatsApp, email or a webhook. When a chat is handed over, the customer sees a "what happens next" line based on the rota, and the system never invents a wait time. Helpdesk needs Standard or higher. Time-to-human itself is something you compute from timestamps in your exports.
Closing the loop. When you find a weak answer, you can take over a chat, and edit a reply into the ideal answer so the agent learns from it. Add missing knowledge from Data sources, such as Q&A entries, files, websites or text snippets. Test changes in the Playground before they reach customers.
Cost. The Usage page shows credits per agent. One reply uses one credit, and credits reset each billing cycle. Use this alongside your handoff counts to estimate cost per resolved chat. The platform has no customer-facing view of underlying AI provider costs, so base your calculation on your plan and credit use.
Leads. With the Collect leads action or a custom form, captured leads appear under Activity → Leads, and each lead also creates a contact. That gives you leads per conversation with no extra tooling.
You can also ask Synthia, the built-in fleet manager, to walk you to the right screen. See the full features list for details.
Frequently asked questions
How many chat support KPIs should a small team track?
Start with five to eight: conversations by topic, resolution rate, time to human, customer rating, sentiment, cost per resolved chat and leads captured if you sell through chat. Add more only when a specific decision needs them.
Is a high containment rate always good?
No. Containment only says that no human joined, not that the customer was helped. Always read it next to ratings, repeat contact rate and a sample of real transcripts.
How do I measure satisfaction when few customers leave ratings?
Combine ratings with sentiment and with a manual review of escalated and low-rated chats. Treat percentages from small samples with caution, and look at trends over several weeks.
How often should I review chat KPIs?
A short weekly review of weak conversations drives improvement, while a monthly review of trends suits managers. Quarterly, revisit your definitions and targets to make sure they still match your goals.


