How to Measure Chatbot Deflection Rate (Simple Guide)
Prices and limits below come from each vendor's published pricing pages and documentation. This is not a paid hands-on test.
Your chatbot is answering questions all day. But is it actually solving problems, or just handing them off to your team with a friendly "let me get a human"? There is one number that tells you the truth: your deflection rate.
Deflection rate is the share of conversations your bot handles alone, with no human stepping in. It is the simplest way to see whether the bot is earning its keep. Get it wrong, though, and you will either celebrate numbers that mean nothing or fix problems that do not exist.
In this guide, we will walk through how to measure chatbot deflection rate properly — the formula, the data you need, the traps to avoid, and three tools that fit small businesses.
What chatbot deflection rate actually means
Deflection rate is a percentage. It compares the conversations your bot fully resolved to the total conversations it handled.
Say 100 people chat with your bot this week and 60 of them never need a human. On the basic definition, that is a 60% deflection rate, and the other 40 conversations went to your team.
That word "deflected" confuses people, because the two readings point in opposite directions. In this guide, a deflected conversation means one the bot passed to a person, so a high deflection rate is good and a low one means your bot is acting like a receptionist that forwards every call. Some support dashboards use the word the other way round, so check the label in your own tool before you compare anything.
One warning before we go further: "resolved" has to mean something specific. A visitor who closes the chat window out of frustration is not a resolved conversation. We will come back to this, because it is the biggest way people fudge this metric.
The formula (and the one adjustment most people miss)
The basic formula looks like this:
- Deflection rate = (conversations resolved by the bot ÷ total bot conversations) × 100
That is it. But raw totals hide a lot. A bot that only handles "what are your hours?" will look excellent on this number while your team drowns in real questions.
So use a second version alongside it, which we call the honest rate:
- Honest deflection rate = (bot-resolved conversations ÷ conversations the bot was *expected* to handle) × 100
To get that, tag each conversation with a topic when it closes — orders, returns, pricing, hours, complaints. Then check which topics the bot should have solved. If your bot is supposed to handle order tracking and most order-tracking chats still reach a human, you have a fixable problem rather than a marketing win.
Here is the pattern worth watching for. A store runs the simple formula and gets a strong headline number. Split the same conversations by topic and returns come out far weaker: the bot was confidently giving the wrong return instructions, so people asked again and again. Identical overall figure, completely different story — and only the split tells you where to fix something.
Step one: define what counts as "resolved"
Before you touch a dashboard, write down your definition. Pick one:
- The visitor got an answer and did not ask for a human.
- The visitor got an answer and did not come back with the same question within 24 hours.
- The visitor completed a goal — placed an order, booked an appointment, found the tracking link.
The third option is the strongest, but it needs your bot connected to your store or booking system. The first is the easiest and the weakest. Start with the second and move to the third as you get comfortable.
Write the definition down and do not change it mid-month. Otherwise your trend line is meaningless.
Step two: collect the right data
You need three things from your chatbot tool: total conversations, conversations with a human handover, and a way to tag topics.
Most tools give you the first two out of the box. The third is where the good tools separate themselves. If you cannot tag or filter by topic, you are stuck with the vanity number.
Also pull these alongside deflection:
- **Containment by channel.** Web chat, WhatsApp and Instagram often behave very differently.
- **Repeat-contact rate.** Same visitor, same question, within a few days. This is your early warning system for bad answers.
- **Handover reasons.** Why did the bot give up? Missing knowledge, angry customer, or a question it was never trained on?
In practice, handover reasons are usually the most useful single report. They tell you what to fix next week.
Step three: pick a tool that makes this easy
You do not need enterprise software to track this. Here are three that cover the range, from budget-friendly to built for growing support teams.
**Tidio** is the one to hand to most small e-commerce stores. Tidio describes live chat as working in under ten minutes, which is the vendor's wording rather than a timed test, and Lyro AI answers from your FAQ automatically. The catch: advanced analytics sit on higher tiers, and Lyro AI credits run out quickly on cheaper plans. Pricing starts at $24.17/mo billed yearly, and there is a free plan with 50 conversations a month — enough to test your measurement setup before you pay anything.
**Freshchat (Freshworks)** suits growing support teams that want everything in one place. It handles web, WhatsApp and Apple Business Messages in a single inbox, and the free plan covers up to 10 agents, which is unusually generous. Freddy AI helps your agents, and the Freddy AI Agent is billed per session as an add-on — the first 500 sessions are free, then $49 per 100 sessions. Paid plans start at $19/agent/mo billed yearly. For this metric you want reporting that separates bot-handled from human-handled chats, so check that the tier you are considering actually exposes it.
**Crisp** is the budget pick. Flat per-workspace pricing from €45/workspace/mo keeps it cheap for big teams, and the chatbot, shared inbox and campaigns are all included. The trade-off is real: AI features are lighter than Tidio or Botpress, and the free plan (2 seats) comes with zero AI credits. If your bot is mostly scripted flows rather than AI answers, Crisp's reporting is plenty.
A quick note on all three: check the vendor's site for current pricing before you commit, since plans change.
Step four: build a simple monthly review
You do not need a data team. Set aside half an hour once a month and answer four questions:
1. What was our deflection rate overall, and by channel? 2. Which three topics had the lowest deflection? 3. What were the top handover reasons? 4. What one change will we make this month?
Then make that one change. Add a missing FAQ article. Rewrite the return policy answer in plain language. Add a button that says "track my order" right at the start.
Teams that chase ten fixes at once learn almost nothing. One change per month, measured against the same definition, tells you what actually works.
The mistakes that make this number lie
These are the traps worth knowing before you trust a single figure.
- **Counting closed windows as resolved.** If the visitor left without a reply, that is not a win. Exclude it or flag it separately.
- **Ignoring the handover button.** Some bots make it so easy to reach a human that deflection looks terrible — while the customer is perfectly happy. Check satisfaction alongside deflection.
- **Comparing across tools.** Every platform defines "conversation" differently. Only compare month to month within the same tool.
- **Chasing a perfect score.** A rate that climbs toward everything-resolved usually means you are blocking people who genuinely need a person. There is no universal target to aim at, so the useful comparison is your own history.
The goal is not a big number. It is fewer frustrated customers and a team that spends its time on real problems.
Frequently asked questions
What is a good chatbot deflection rate?
There is no published universal benchmark worth quoting, because every platform defines a conversation differently and every business has a different question mix. Use your own history as the baseline: fix a definition, measure for a month, then judge each change against it. A rate that falls back while repeat contacts climb usually means the bot is giving wrong answers; a rate that holds while handovers drop is real progress.
How often should I measure deflection rate?
Once a month is enough for most small teams. Measure weekly only when you are actively testing a change, like a new FAQ article or a rewritten answer. Always use the same definition so your trend line stays honest.
Can I measure deflection rate without paid analytics?
Yes. Most chatbot tools show total conversations and human handovers on their basic or free plans, which is all the formula needs. Tidio's free plan, for example, includes 50 conversations a month — enough to build the habit before upgrading.
Does a high deflection rate mean my chatbot is successful?
Not on its own. Pair it with customer satisfaction and repeat-contact rate. A bot that hands nothing to a human but sends people back with the same question three days later is failing quietly, and only the second number will tell you.
Ready to pick one?
Marcus Reed — Marcus Reed is the senior editor at ChatbotRadar. He tracks live chat and AI chatbot tools for small teams, verifying pricing on vendor sites and trialling tools on their free plans. How he tests · Report an error
Disclosure: this article contains affiliate links. If you buy through them we may earn a commission at no extra cost to you. Pricing and features are re-checked regularly but can change — always confirm on the vendor's site.