Mike HeapMike is an experienced Product Manager who focuses on all the “non-development” areas of My AskAI, from finance and customer success to product design, copywriting, testing and more.
Table of Contents
- Deflection rate, in more depth
- How is deflection rate measured?
- What's a "good" deflection rate?
- Common misconceptions about deflection rate
- Misconception 1: deflection rate is the same as resolution rate
- Misconception 2: a higher deflection rate is always better
- Misconception 3: deflection and containment are the same metric
- What deflection rate is NOT, and how it relates to its neighbors
- How does My AskAI handle deflection?
- FAQs
Created time
Jun 2, 2026 09:51 PM
Title length (<60)
Author
Ecomm?
Image
Publish date
Jun 7, 2026
Slug
what-is-deflection-rate
Featured
Type
Article
Ready to Publish
Ready to Publish
Deflection rate is the percentage of customer support contacts handled by self-service or AI before they ever reach a human agent.
Deflection rate is the easiest support metric to celebrate and the easiest to fake: you could hit a "perfect" 100% just by making the human impossible to reach. Here's how I read it straight: the formula, what counts, real benchmarks, and why deflection isn't resolution.
Deflection is a funnel-top metric. It counts whether a contact reached a human. Whether the customer's problem got solved is a separate question.
So a help-center article that answered the question counts as deflected. So does a community thread, or a chatbot that ended without an escalation.
That gap is the whole point of this page (and the reason I wrote it). Resolution tells you the issue was solved; containment tells you the conversation stayed in the bot; deflection only tells you a person was never involved.
If you're here, I'd bet you're trying to work out one thing: is your deflection number a real win, or a stat that's quietly hiding customers who gave up?
Deflection rate, in more depth
TL;DR: Deflection counts contacts that never reached a human, via help center, community, or a bot that didn't escalate. It's a funnel-top efficiency metric, and it says nothing about whether the customer was actually helped.
Deflection started life as a self-service and call-center number. Zendesk calls ticket deflection "the currency of self-service": every question your help center answers is a ticket your team never has to touch.
That idea moved off IVR phone trees and knowledge bases and onto chatbots and AI agents. Which is exactly why the word now gets used for three slightly different things (and why I keep having to untangle it on calls).
It's an efficiency metric. Quality is a separate question. A high deflection rate tells you fewer contacts reached your team, and nothing on its own about whether those people left happy or fuming.
The edges get blurry once a bot is in the mix, and most vendors use the words loosely. The cleanest way I've found to keep them apart is by where they fire: deflection fires pre-ticket, at the funnel top, when a contact never reaches a human at all. Containment fires in-channel, when a conversation stays with the bot start to finish.
Alhena puts the line plainly. Deflection "measures tickets that never reach a human agent. Containment rate measures conversations the AI handles end-to-end without escalation."
How is deflection rate measured?
TL;DR: Deflection rate = contacts handled without a human ÷ total inbound contacts × 100. The soft spot is the denominator, because the tickets that were never opened can't be counted.
The formula everyone agrees on is simple:
Deflection rate = (contacts handled without a human ÷ total inbound contacts) × 100
Decagon writes it as (Total Issues Resolved via Self-Service / Total Issues Submitted) × 100. Alhena writes it as (Self-Service Resolutions / Total Support Requests) x 100. Same formula, just worded differently.
Here's the catch: it's the denominator. The classic self-service version is really (potential tickets minus actual tickets) ÷ potential tickets, and "potential tickets" is a number you can't actually count. You've no way of counting the tickets nobody ever opened.
So teams approximate (this is where it gets squishy): help-center sessions that didn't lead to a ticket, or bot chats that didn't escalate. Where you draw that line decides how flattering the number looks.
There's a second catch, and it's in the numerator. A customer who didn't understand the answer and just left still counts as deflected under most definitions. Decagon even notes that some vendors count "interactions where users engaged with self-service tools but did not escalate." Engaged-then-gave-up looks identical to solved-and-happy in the deflection column.
That's also why the same word spits out wildly different numbers across vendors (I run into this on nearly every call). Here's what each platform actually counts as "handled without a human":
Vendor | What they call it | What counts as "deflected" (no human) |
My AskAI | AI resolution rate (we don't report deflection) | conversation not escalated to a human, with escalation kept deliberately easy |
Zendesk | Ticket deflection / self-service | a help-center view or bot answer that doesn't open a ticket |
Fin (Intercom) | Resolution (deflection-adjacent) | a configured outcome; a procedure ending in handoff can count |
Decagon | Deflection rate | a self-service or bot interaction not routed to a human |
Gorgias (Automate) | Automation / deflection | the customer didn't reply |
Forethought | Deflection rate (cost proxy) | a self-serve or AI contact not escalated to a live agent |
We report AI resolution rate rather than deflection, for reasons I'll get to. The short version: a "handled without a human" count is only as good as how easy it is to reach a human in the first place.
What's a "good" deflection rate?
TL;DR: Across 195 real deployments the median AI-handling rate is about 70%, and deflection-labeled stats land there too (most between 62% and 82%). Always read it next to CSAT.
I'll start with real data, because most "good deflection rate" answers are hand-waved. Across 195 real AI deployments spanning more than 55 vendors, the median AI-handling rate sits at 70%, with the middle half of the field between 56% and 80%.
The stats explicitly labeled deflection (a smaller set of 17) land in the same place: a 70% median, most between roughly 62% and 82%. So 70% is the scores on the doors number for the whole field.
Three caveats ride along with those numbers, and I'd take all three seriously. They're an aggregate rather than a head-to-head between any two vendors, and they're directional, since every vendor defines its metric differently.
They're also a self-selected ceiling (published figures are marketing wins), so I'd take the true field average as a notch lower. Treat the whole table with a small grain of salt.
The label moves the number more than the actual product does, and that's the part that trips people up. In the same dataset, stats called "resolution" run a 72.5% median against 61% for "automation", with deflection sitting in between. Buyers comparing vendor headline numbers are usually comparing different yardsticks rather than different tools.
There's a second trap worth naming. Classic self-service deflection and AI-agent deflection aren't the same measurement, and their benchmarks disagree by a mile.
Help-center deflection bands run low (Alhena puts average at 20-40%, top performers at 80%+). AI-agent deflection in our field data clusters near 70%, so neither is wrong; they're counting different funnel stages with different denominators.
A "good deflection rate" means nothing until you know which one you're quoting.
Tier | Range (AI-agent deflection) | What it usually means |
World-class | 80%+ | mature deployment, APIs connected, transactional vertical |
Solid | 65-75% | knowledge base optimized, common questions covered |
Average | 40-65% | partial coverage, a meaningful tail still escalates |
Basic | 20-40% | rule-based or thin-knowledge bot |
Industry changes the floor and ceiling a lot. In the field data, education (~81%) and travel (~78%) deflect high; telecom (~58%) and manufacturing (~64%) deflect low. Transactional ecommerce queues (where-is-my-order, returns, sizing) deflect well because the answers are lookups.
High-judgment and regulated queues should deflect lower, because more of those tickets ought to reach a person. Read every deflection number next to its CSAT. I can't stress that one enough.
Common misconceptions about deflection rate
TL;DR: Three myths do the most damage: that a deflected ticket is a solved one, that higher is always better, and that deflection equals containment. All three fall apart the moment you look at CSAT.
Misconception 1: deflection rate is the same as resolution rate
Deflection proves a human wasn't reached. Resolution proves the issue was solved. They're not the same, and the gap between them is exactly where unhappy customers hide.
Even Fin, Intercom's own AI, says it out loud: "What matters is whether the issue was resolved, regardless of who handled it," and "without resolution, deflection can actually harm customer experience by leaving issues unresolved and forcing customers to find alternative ways to get help." DigitalGenius titled a whole piece "Ticket Deflection vs. Resolution Explained."
Here's the cleanest way I've found to see the flaw. You could post a perfect 100% deflection rate tomorrow by deleting every route to a human.
Nobody reaches you, so nobody is "un-deflected", and the number is a lie. Optimizing for deflection on its own quietly rewards exactly that move.
Misconception 2: a higher deflection rate is always better
Past a point, a rising deflection rate is a warning sign. Alhena's own guidance is blunt: "A deflection rate above 90% without matching resolution and CSAT scores can signal that customers are being blocked from reaching agents rather than genuinely helped."
In high-judgment queues (fraud, billing disputes, anything emotional) a fast handoff to a person is the right outcome. A high deflection rate there means the AI is clinging to tickets it should let go. The real danger is a high number, manufactured by making the human hard to reach.
This isn't just a vendor talking point; the people running support say it themselves. One operator post puts it as "deflection rate is the worst metric to celebrate in customer support."
Another, from PolyAI's CEO: "'We deflected 70% of all contact.' That number is meaningless without another question: what happened next?"
Misconception 3: deflection and containment are the same metric
They fire at different points (this is the distinction I labor over most). Deflection is pre-ticket: the contact never reached a human. Containment is in-channel: the conversation stayed with the bot end to end, follow-up questions included.
A help-center article that stops a ticket being opened is deflection. A chatbot that handles a five-message back-and-forth without escalating is containment. Treating them as one number is how vendor stats stop being comparable (I see it all the time).
What deflection rate is NOT, and how it relates to its neighbors
TL;DR: Deflection sits next to resolution, containment, automation, and self-service rate. They're four different events on the customer journey, so tracking them as one number hides more than it shows.
Deflection sits in a family of metrics that support teams throw around as if they're synonyms. They really aren't, and I'll gently push back any time someone treats them as one number.
Term | Definition | Difference from deflection |
AI resolution rate | the customer's issue was actually solved, end to end, by the AI | deflection ignores whether anything got solved |
Containment rate | the conversation stayed in the automated channel | containment is in-channel; deflection is pre-ticket |
Automation rate | the share of support work the AI touched | touching a ticket isn't deflecting or resolving it |
Self-service rate | the customer found the answer without contacting you at all | self-service is the main mechanism that produces deflection |
Autonomous resolution | the AI solved it start to finish, confirmed | the real version of "deflected and actually solved" |
the escalation event itself | deflection measures the absence of a handoff |
The mental model I use is a spectrum from soft to strict. Deflection and containment are the soft end: they describe where a ticket went. Resolution and autonomous resolution are the strict end: they describe whether the customer got what they came for.
If you've only got room to track one number, track the strict one, and read it next to CSAT.
How does My AskAI handle deflection?
TL;DR: We report AI resolution rate rather than deflection, and keep escalation easy on purpose. A deflection number is only as trustworthy as how easy it is to reach a person.
We don't optimize for deflection, and we don't report it as our headline metric. We report AI resolution rate.
A conversation counts as resolved when it wasn't escalated to a human. And the honesty of that number comes entirely from one thing: escalation is deliberately easy.
A customer can ask for a person in plenty of ways, and the AI also hands off when it can't answer, when it spots frustration, or when a ticket hits a topic the team has set for escalation. We don't claim to know an issue was truly solved without the customer confirming it, and we don't invent ways to pretend we do.
That stance comes straight from how the metric gets gamed. A deflection number is only as good as how easy it is to reach a person, and the cleanest way to fake one is to bury the escalation path. I'd rather report a defensible number on an easy-to-escape system than a flattering one on a trap.
The proof is in customer numbers read next to satisfaction rather than deflection on its own.
Swytch deflects 81% of support contacts, more than 4,050 tickets a month, across its Zendesk setup.
TravelJoy went from 24% on Zendesk's own AI agent to 80% with us, at 86% AI CSAT. Their head of customer service, Alan Pugh, put it as "you're beating Zendesk's AI agent 76% to 24% on AI deflection. Huge."
Zinc handles 68% of its queries with response times under 60 seconds and a 97% CSAT. The deflection figure only means something because the CSAT is sitting right next to it.
Two limits worth naming. We've no voice product, so call deflection and IVR containment aren't us; the deflection we measure is chat, email, and helpdesk tickets.
And we're priced per ticket and never per resolution, so there's no incentive on our side to inflate a "handled" count. Across the customer base, our agents resolve more than 72% of tickets on a rolling 30-day basis, and 200+ ecommerce and SaaS teams run support this way.
FAQs
What is deflection rate?
Deflection rate is the percentage of customer support contacts handled by self-service or AI before they reach a human agent. I'd call it a funnel-top efficiency metric: it measures whether a person was involved. Whether the customer's problem got solved is a separate question.
What is deflection rate in customer service?
In customer service it usually means the share of incoming contacts your help center, chatbot, or AI agent handles so they never become a ticket a human has to answer. It's tied tightly to self-service: the more questions your knowledge sources answer well, the higher your deflection climbs.
How do you calculate deflection rate?
The standard formula is (contacts handled without a human ÷ total inbound contacts) × 100. The hard part is the denominator (it always is). The classic self-service version uses "potential tickets", which you can't count directly, so teams approximate with help-center sessions or bot chats that didn't escalate.
What is a good deflection rate?
It depends which deflection you mean. Classic help-center deflection runs around 20-40% on average and 80%+ at the top, while AI-agent deflection in our field data clusters near a 70% median (most deployments land between 62% and 82%). Always read the number next to CSAT, because a high rate with a low satisfaction score is a problem in itself.
What is the difference between deflection rate and resolution rate?
Deflection only tells you a human wasn't reached. Resolution tells you the issue was actually solved. A customer who gave up still counts as deflected but not as resolved, which is why I'd reach for resolution rate as the truer measure of whether your support is working.
What is the difference between deflection rate and containment rate?
Deflection fires pre-ticket: the contact never reached a human at all. Containment fires in-channel: a conversation stayed with the bot from start to finish (that pre-ticket vs in-channel split is the bit I lean on). A help-center answer that prevents a ticket is deflection; a bot that handles a full back-and-forth without escalating is containment.
What is ticket deflection rate?
Ticket deflection rate is deflection applied specifically to support tickets: the share of would-be tickets resolved by self-service or AI so they never land in an agent's queue. It's the helpdesk-flavored version of the same metric, and the one we get asked about most.
What is call deflection rate?
Call deflection rate is the voice version: the share of phone contacts handled by an IVR, callback, or self-service option so they never reach a live agent. It's a contact-center metric, and we don't measure it, because My AskAI has no voice product.
What is chatbot or AI deflection rate?
It's the share of conversations a chatbot or AI agent handles without passing them to a human. Because a bot is involved, this version blurs into containment (the two get used interchangeably all the time), so I'd check what a vendor actually counts before you compare numbers.
Is a higher deflection rate always better?
No. Past a point, a rising deflection rate can mean customers are being blocked from reaching agents rather than helped, especially if CSAT is sliding. In high-judgment queues a fast handoff to a person is the right outcome, so a very high deflection rate there is a red flag in my book.
Why is deflection rate controversial?
Because it's easy to celebrate and easy to game. It measures money saved and says nothing about whether problems got solved, and you can push it up just by making it harder to reach a human, which hurts customers while the number looks great. That's why a lot of operators (me included) argue you should report resolution rate paired with CSAT instead.
How can I improve my deflection rate?
Improve the knowledge your self-service and AI draw on, cover the questions that actually drive contacts, and put the AI on the front line for high-volume, low-judgment queries (that's where we see the cleanest wins). But do it the right way: keep escalation easy and watch CSAT, so you're deflecting because customers got answers rather than because they gave up.
Does deflection rate vary by industry?
Yes, a lot. In the field data, education and travel deflect high (roughly 78-81%), while telecom and manufacturing sit lower (around 58-64%). Transactional ecommerce queues deflect well because they're lookups; regulated, high-judgment queues should deflect lower by design.
What resolution rate should I expect from AI support?
Across 195 real deployments the median AI-handling rate is about 70%, with the middle half between 56% and 80%. Treat that as a directional benchmark, never a promise: your own number depends far more on your setup and ticket mix than on your industry, so the one thing I'd push every buyer to do is test it on your own tickets.
