A founder once told me his new chatbot was deflecting 80% of support tickets. Impressive, until we looked closer. Half of those "deflected" chats ended with the customer typing "this is useless" and closing the window. The bot didn't resolve anything. It just stood between people and the help desk, and the dashboard called that a win.
That's the trap with deflection rate. It's one of the most useful support metrics you can track, and one of the easiest to fake, sometimes without meaning to.
What deflection rate is actually supposed to measure
Deflection rate is the share of customer questions resolved without a human agent stepping in. The point is to answer the honest question: how much repetitive work is your self-service handling, so your team can spend its time on the hard cases?
Notice the word resolved. That's the whole game. A conversation only counts as deflected if the customer got a real answer and didn't need a person afterward. A session where someone rage-quit, or emailed you an hour later about the same thing, was not deflected. It was abandoned, and abandonment dressed up as success is worse than no metric at all, because it tells you to keep doing the thing that's failing.
The formula, and the part everyone skips
The basic calculation looks simple:
Deflection rate = fully resolved self-service conversations / total conversations that sought help
The trouble is the denominator and what you allow into the numerator. Here's a cleaner version of the buckets you need to sort chats into first:
- Resolved by self-service: the bot answered, the customer left satisfied, no follow-up.
- Escalated to a human: the bot handed off, on purpose or because it was stuck.
- Abandoned: the customer left without an answer and didn't come back through another channel.
- Not a support question at all: spam, testing, someone browsing.
Only the first bucket belongs in your numerator. Escalations and abandonments do not. And you should strip out the last bucket entirely, because counting "what are your hours" pageview traffic as deflected support inflates the number without meaning anything.
Watch for the follow-up leak
Here's the mistake that quietly wrecks the metric. A customer chats with the bot, seems to get an answer, closes the window, and then emails support twenty minutes later about the exact same issue. Your chat tool marks that first session as resolved. Your help desk logs a fresh ticket. Nothing was deflected, but your two systems each recorded a win.
To catch this, you have to connect the chat log to the ticket system, or at least sample. Take fifty "resolved" chats and check whether the same customer opened a ticket within a day. If a big share did, your real deflection rate is lower than the dashboard claims, and you've found exactly which answers aren't landing.
There's a channel-switch version of the same leak. A customer chats, gets nowhere, and calls instead. Your phone log and your chat log never talk to each other, so the failed chat still reads as resolved. If you take calls, spot-check whether callers tried the bot first. The ones who did are your clearest signal that a self-service answer needs rework.
Pair it with satisfaction, always
Deflection on its own can be gamed by a bot that's confidently unhelpful. The guardrail is to track resolution and satisfaction side by side. If deflection climbs while satisfaction on those chats drops, you're not saving work, you're pushing frustration downstream where it's harder to see.
A short thumbs-up or thumbs-down at the end of a bot conversation is enough to start. What you want is the pattern: high resolution and steady satisfaction together. One without the other is noise.
Watch the trend, not just the level. A deflection rate that climbs while satisfaction holds is a healthy sign your answers are getting better. A rate that climbs while satisfaction slips usually means you've made it harder to reach a human, not easier to get help. Same number moving up, opposite meaning.
Consider a small SaaS company, "Marlowe Books," that runs scheduling software for salons. Before measuring carefully, they claimed 75% deflection. After they excluded abandoned sessions and checked for follow-up emails, the real figure was 48%. That sounds like a downgrade. It wasn't. It was the first honest number they'd had, and it pointed straight at the three question types the bot kept fumbling: refunds, timezone changes, and how to move a booking. They fixed those answers, and real deflection climbed to the low 60s over a month. The honest baseline is what let them improve.
A quick audit you can run this week
You don't need a data team to get an honest read. Work through this checklist:
- Pull one week of bot conversations, not a cherry-picked day.
- Tag each as resolved, escalated, abandoned, or not-support.
- Cross-check a sample against tickets opened within 24 hours.
- Calculate deflection using only fully resolved chats over real support chats.
- Break the number down by question type to see where it's strong and weak.
- Look at satisfaction on the resolved bucket, not the whole pile.
Do that once and you'll trust the number. Do it monthly and you'll watch it move for real reasons.
Reading the number without fooling yourself
Once you have an honest figure, resist the urge to chase a single target. There's no universal "good" deflection rate. For repetitive, well-documented questions, resolving 50 to 70 percent without a human is a realistic range. For nuanced, account-specific, or emotional issues, you want those reaching a person quickly, and a low deflection rate there is correct, not a failure.
The most useful view isn't the headline percentage at all. It's the breakdown. Which question types does self-service handle cleanly, and which ones should always route to a human fast? That's the map you actually act on. In SpideyChat you can read the transcripts alongside the resolved and escalated tags, which makes those patterns easy to spot without exporting anything.
Segment it further where it helps. New visitors and returning customers often behave differently. Deflection during business hours tells a different story than deflection at 2am, when a person answering isn't even an option and the bot is the whole show.
Time-based segmentation often reframes the whole conversation. A modest daytime deflection rate can look disappointing until you notice that at night, when the alternative is nobody, the bot resolves most of what comes in. That after-hours resolution is pure upside. It's work that simply wouldn't have happened otherwise, and it deserves more credit than a raw blended average tends to give it.
Treat deflection as a conversation starter with your own data, not a trophy. Measure it honestly, pair it with satisfaction, and let the weak spots tell you which answers to write next. That's the version of the metric that actually makes your support better instead of just making the slide look good.