{"id":9050,"date":"2026-10-26T14:34:38","date_gmt":"2026-10-26T09:04:38","guid":{"rendered":"https:\/\/www.anakage.com\/blog\/?p=9050"},"modified":"2026-10-26T15:48:53","modified_gmt":"2026-10-26T10:18:53","slug":"ticket-deflection-rate","status":"publish","type":"post","link":"https:\/\/www.anakage.com\/blog\/ticket-deflection-rate\/","title":{"rendered":"Ticket Deflection Rate: Why the Metric Misleads IT Teams"},"content":{"rendered":"

Ticket deflection counts the tickets that never opened. It does not count the problems that got solved. A user who gives up on your portal and walks to a colleague’s desk counts as deflected. Your dashboard reads that as a win.<\/p>\n

Gartner surveyed 5,728 people and found 73% use self-service somewhere in their journey, while only 14% fully resolve their issue there. That gap is what the deflection number hides.<\/p>\n

Every figure below links to its source. Where no source exists, this post says so.<\/p>\n

\n

Ticket deflection.<\/strong> A support request that is handled without a human agent, usually through a self-service portal, knowledge base article or chatbot.<\/p>\n<\/blockquote>\n

How Ticket Deflection Rate Is Calculated<\/h2>\n

There is no standard formula. Two widely used definitions produce different answers from the same traffic.<\/p>\n

Zendesk calculates it as total help centre users divided by total users in tickets. That returns a ratio, so the answer might be 4.<\/p>\n

Decagon calculates deflected contacts divided by total contact attempts, times 100. That returns a percentage, so the answer might be 80%.<\/p>\n

Same traffic. One vendor reports 4, another reports 80%.<\/p>\n

The denominators differ and the time windows differ. The two also disagree on whether an abandoned session counts as deflected.<\/p>\n

So a rate from one tool cannot be compared with a rate from another. A benchmark quoted without its formula means very little.<\/p>\n

For reference, Decagon publishes these bands. E-commerce 55% to 75%, SaaS 40% to 60%, and financial services and healthcare 25% to 45%.<\/p>\n

Gartner’s Own Numbers Show the Gap<\/h2>\n

The Gartner survey behind the 14% figure ran in December 2023 with 5,728 participants and was published in August 2024.<\/p>\n

Two findings sit side by side:<\/p>\n

    \n
  • 73% of customers use self-service at some point in their service journey<\/li>\n
  • 14% fully resolve their issue there<\/li>\n<\/ul>\n

    Even for issues customers described as “very simple,” the full resolution rate was 36%.<\/p>\n

    Eric Keller, Senior Director of Research in Gartner’s Customer Service and Support practice, said of the finding that “it’s concerning to see that so few fully resolve there.”<\/p>\n

    A deflection metric counts the 73%. The 14% is the number that matters.<\/p>\n

    What Ticket Deflection Costs When Users Come Back<\/h2>\n

    Here is where the metric stops being merely inaccurate and starts costing money.<\/p>\n

    Published cost-per-ticket benchmarks, compiled from MetricNet, HDI and Forrester figures, put the tiers roughly here:<\/p>\n

    \n\n\n\n\n\n\n\n\n
    Tier<\/th>\nAverage cost per ticket<\/th>\nRange<\/th>\n<\/tr>\n<\/thead>\n
    Tier 0, self-service<\/td>\n$1 to $4<\/td>\n$0.50 to $5<\/td>\n<\/tr>\n
    Tier 1<\/td>\n$22<\/td>\n$6 to $35<\/td>\n<\/tr>\n
    Tier 2, desktop support<\/td>\n$70<\/td>\n$35 to $90<\/td>\n<\/tr>\n
    Tier 3<\/td>\n$104+<\/td>\n$90 to $150+<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n

    Now take a desk handling 1,000 tickets a month, with a tool reporting 40% deflection.<\/p>\n

    If all 400 deflected tickets genuinely resolve, the monthly cost falls from $22,000 to $14,200. That is a $7,800 saving, and it is the number in the business case.<\/p>\n

    But deflected users do come back, and by then the problem is often worse. A laptop that would not connect has now missed a deadline. It arrives as a Tier 2 ticket instead of a Tier 1 one.<\/p>\n

    \n\n\n\n\n\n\n\n\n\n
    Share of deflected users who return at Tier 2<\/th>\nMonthly cost<\/th>\nActual saving<\/th>\nShare of the promised saving<\/th>\n<\/tr>\n<\/thead>\n
    0%<\/td>\n$14,200<\/td>\n$7,800<\/td>\n100%<\/td>\n<\/tr>\n
    10%<\/td>\n$16,900<\/td>\n$5,100<\/td>\n65%<\/td>\n<\/tr>\n
    25%<\/td>\n$20,950<\/td>\n$1,050<\/td>\n13%<\/td>\n<\/tr>\n
    29%<\/td>\n$22,000<\/td>\n$0<\/td>\n0%<\/td>\n<\/tr>\n
    40%<\/td>\n$25,000<\/td>\n\u2212$3,000<\/td>\nWorse than no tool<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n

    At a 29% return rate the saving disappears completely.<\/strong> Above that, a tool reporting 40% deflection is costing you money.<\/p>\n

    The deflection rate reads 40% in every row of that table.<\/p>\n

    Two caveats on those numbers. The unit costs are third-party benchmark averages, not your costs, and the return rates are illustrative rather than measured. Run the same model with your own cost per ticket and your own re-contact data, which is the point. Nobody publishes an industry figure for how often deflected users escalate.<\/p>\n

    What IT Admins Say About Ticket Deflection<\/h2>\n

    An r\/sysadmin thread from October 2025 asked whether enterprise service desks are being automated into a corner. It drew 302 upvotes and 149 comments. The poster described the loop:<\/p>\n

    “Your chatbot resolved my ticket in 3 seconds by sending me a knowledge base article I’d already tried. Ticket closed, metrics look great, but my laptop still won’t connect to the VPN and now I’ve wasted 20 minutes in a loop.”<\/p>\n

    The top reply, at 293 upvotes, was one question: “Why are you automating the close of the ticket when the issue isn’t resolved?”<\/p>\n

    A reply at 80 upvotes made the comparison that lands hardest: “A real live Service Desk tech would get reprimanded for doing this, why the hell would we allow the AI to do it and call it good?”<\/p>\n

    Another, at 15 upvotes: “A new L1 tech would get talked to if they just sent a KB doc and closed the ticket without knowing if it helped.”<\/p>\n

    A second thread, from March 2026, was titled “‘Ticket deflection’ feels like a vanity metric. what are you measuring instead.” Its author had been shopping for tools:<\/p>\n

    “I literally talked to every big vendor and and I keep getting pitched ‘ticket deflection’ like its the #1 thing that matters… in pratice people still get blocked, still DM the team, still reopen the same thing, or they just give up and try again later. The bot gets a win and we get the pile of hot mess.”<\/p>\n

    Three specific failures came up repeatedly across both threads.<\/p>\n

    Users learn to route around the system<\/h3>\n

    At 27 upvotes: “The best a bot is going to get you is a delayed and angrier user. The worst is they all avoid your ticketing system and just call you or show up at your desk.”<\/p>\n

    One admin described how that spreads: “once you start getting the exception tickets. The ones that a tech needs to actually be engaged on the requesting user becomes aware of who that is and they tend to start skipping the bot entirely for that type of ask. Ultimately undermining any kind of metics gathering that IT is doing on ticket type and getting us right back where we started like there was never a queue to begin with.”<\/p>\n

    So you lose the work and the data about the work. Your category reporting degrades at the same time as your queue appears to improve.<\/p>\n

    Guides do not help users who cannot follow them<\/h3>\n

    At 6 upvotes, and the most useful sentence in either thread: “end users are not tech savvy enough to self resolve problems. We could lay out step by step guides and as soon as users see they need to do something, they go into paralysis, freeze up and open a ticket anyways. If AI could actually do the work to fix stuff it would be of value. Right now, all it can really do is point users to KB articles or help open a ticket which IMO just slows resolution times down.”<\/p>\n

    The distinction there is between telling a user what to do and doing it for them. A knowledge base article is an instruction. Most deflection counts instructions.<\/p>\n

    The metric punishes the people doing the hard work<\/h3>\n

    One engineer at 36 upvotes: “When I was doing T3 escalations I got a negative performance review because my tickets closed and resolution times were far lower than T1. I’m sorry dude I can’t order a new hardware revision from our manufacturer in the same time it takes help desk to reset a password.”<\/p>\n

    Goodhart’s Law Is the Underlying Problem<\/h2>\n

    A commenter at 49 upvotes named it: “as soon as you start using them to measure, you run into Goodharts Law.”<\/p>\n

    \n

    Goodhart’s Law.<\/strong> When a measure becomes a target, it stops being a good measure, because people optimise for the number rather than the thing it was meant to represent.<\/p>\n<\/blockquote>\n

    Another comment showed the second-order version. A service desk manager aware of Goodhart’s Law tried stack-ranking his agents against each other instead:<\/p>\n

    “This only made things worse, because it was just a different form of Goodhart’s law – don’t be at the bottom of the stack-rank. And it also unconsciously told us that we’re in competition with each other, so knowledge-sharing went down the drain. And also because we could pick incoming tickets from the queue, of course the most difficult, time-consuming tickets sat for days in the queue until users complained to him.”<\/p>\n

    A shorter version, at 3 upvotes: “The iron law is you get what you measure, management generally only interested in ticket closures not whether problems actually solved.”<\/p>\n

    And at 36 upvotes, the structural diagnosis: “It’s because the measurements aren’t about end 2 end outcomes, they are about step outcomes.”<\/p>\n

    Deflection is a step outcome. It measures one handoff, not whether the user got back to work.<\/p>\n

    The Klarna Reversal<\/h2>\n

    Klarna ran the largest public experiment in optimising for deflection, and then undid it.<\/p>\n

    In February 2024 the company said its AI assistant could do the work of 700 customer service agents. That was Klarna’s own claim. Within a month the bot was handling around 75% of customer chats, about 2.3 million conversations, across more than 35 languages. Headcount fell 22% to roughly 3,500, mostly through attrition during a hiring freeze.<\/p>\n

    In May 2025 CEO Sebastian Siemiatkowski told Bloomberg that Klarna was hiring humans again. He said the chatbots were cheaper but produced “lower quality” output, and that the AI-first strategy “wasn’t the right path.”<\/p>\n

    His conclusion: “Really, investing in the quality of human support is the way of the future for us.”<\/p>\n

    The deflection numbers were never in dispute. The bot did handle 75% of chats. The problem was what that 75% actually delivered.<\/p>\n

    What to Measure Instead of Ticket Deflection<\/h2>\n

    Five measures, drawn from what practitioners in those threads said they trust.<\/p>\n

      \n
    1. Re-contact rate.<\/strong> How many users come back about the same issue within 7 and 30 days. This is the single number deflection hides, and your ticketing system already holds the data.<\/li>\n
    2. Cost per resolution, not cost per ticket.<\/strong> Divide total support cost by issues actually closed and not reopened. A deflected ticket that returns at Tier 2 shows up properly here.<\/li>\n
    3. Time to first human touch.<\/strong> One admin with 20 years in the field wanted exactly three things: “time to first touch, volume by category and priority, and TTR.” None of them is deflection.<\/li>\n
    4. Change in satisfaction, measured per ticket.<\/strong> One vendor employee in the thread, who disclosed their affiliation, argued for tracking the change rather than the score. It shows which interactions make users angrier. Another manager grew satisfaction from 60% to over 90% and gave it veto power over every other metric.<\/li>\n
    5. Issues prevented at source.<\/strong> Count the problems that never became tickets because the condition was fixed first. This is the hardest to measure and the only one that genuinely reduces work.<\/li>\n<\/ol>\n

      Goodhart’s Law applies to all five. Each becomes gameable the moment it becomes a target. Use several together, and treat any single number as a reason to investigate rather than an answer.<\/p>\n

      Shift Left Means Prevention, Not a Chatbot<\/h2>\n

      The most constructive comment in the October thread, at 14 upvotes, corrected the term everyone was using:<\/p>\n

      “Shifting left, in the context of user issues, has always been to prevent the users from having issues in the first place — not sending in a chatbot.”<\/p>\n

      The same admin listed what that looks like in practice. Stop requiring password rotation, which removes the resets rather than deflecting them. Phase out client VPNs, which removed their VPN tickets entirely. Plan for users to need software. And then this:<\/p>\n

      “We collect telemetry about the client hardware, how it’s performing, and its connections. These handle a lot of the routine anguish that frustrates service desks: overheating machines, insufficient uplinks, complaints about performance.”<\/p>\n

      Prevention and deflection look similar on a ticket-volume chart and are opposites in practice. One removes the problem. The other moves the conversation.<\/p>\n

      Tools in this category are worth separating on that basis. Does the product detect a condition and correct it on the endpoint? Or does it show the user an article and close the ticket?<\/p>\n

      Anakage is one option worth considering here. Its published case study covers a global IT service provider averaging 2,000 incidents and service requests a month, and reports a 30% reduction in incidents and service requests “at source” using automated detection and resolution, alongside a 50% cut in asset verification time. Treat it as a vendor-reported figure from a single named deployment, which is what it is, and ask for the measurement method.<\/p>\n

      What We Could Not Verify<\/h2>\n

      Four gaps, stated rather than filled.<\/p>\n

      No published figure exists for how often deflected users re-contact support. The cost model above uses illustrative rates because the industry data does not exist.<\/p>\n

      The cost-per-ticket tiers are third-party compilations citing MetricNet, HDI and Forrester. The compiler does not map individual figures to individual sources, and labels them estimates to validate against your own environment.<\/p>\n

      The 700-agent Klarna figure was Klarna’s own claim about capability, not a headcount reduction that was independently audited. Reporting at the time did not verify it.<\/p>\n

      Gartner’s 14% finding comes from a customer service survey, not an internal IT service desk survey. The pattern is likely similar in enterprise IT, but we have not found an equivalent study for internal support.<\/p>\n

      FAQ<\/h2>\n

      Q: What is ticket deflection?<\/strong><\/p>\n

      A: A support request resolved without a human agent, usually through a self-service portal, knowledge base article or chatbot. The term counts requests that never reached an agent, which is not the same as problems that were solved.<\/p>\n

      Q: How is ticket deflection rate calculated?<\/strong><\/p>\n

      A: There is no standard formula. Zendesk divides help centre users by users in tickets, which returns a ratio. Decagon divides deflected contacts by total contact attempts and multiplies by 100, which returns a percentage. The same traffic can read as 4 or as 80%.<\/p>\n

      Q: What is a good ticket deflection rate?<\/strong><\/p>\n

      A: Any benchmark is close to meaningless without the formula behind it. Decagon publishes 55% to 75% for e-commerce, 40% to 60% for SaaS, and 25% to 45% for financial services and healthcare. Compare only against your own baseline using one consistent formula.<\/p>\n

      Q: Why is ticket deflection considered a vanity metric?<\/strong><\/p>\n

      A: It measures a handoff rather than an outcome. A user who abandons a chat after an unhelpful answer counts as deflected while their problem is unsolved. Gartner found that 73% of customers use self-service while only 14% fully resolve there.<\/p>\n

      Q: Can ticket deflection actually increase support costs?<\/strong><\/p>\n

      A: Yes, if deflected users return later with a bigger problem. Published tier costs put Tier 1 at $22 and Tier 2 at $70. On those figures, a desk reporting 40% deflection loses its whole saving once about 29% of deflected users come back at Tier 2. The deflection rate does not change.<\/p>\n

      Q: What should I measure instead of ticket deflection?<\/strong><\/p>\n

      A: Re-contact rate within 7 and 30 days, cost per resolution rather than cost per ticket, time to first human touch, change in satisfaction per ticket, and issues prevented at source. Use several together, because any single target gets gamed.<\/p>\n

      Q: What is Goodhart’s Law and why does it apply here?<\/strong><\/p>\n

      A: Goodhart’s Law says that when a measure becomes a target, it stops being a good measure. Once deflection is the target, closing tickets without resolving them raises the score. One admin summarised it as “you get what you measure.”<\/p>\n

      Q: Does shift-left mean deploying a chatbot?<\/strong><\/p>\n

      A: Not originally. As one admin put it, shifting left means preventing users from having issues in the first place. Removing password rotation or retiring a client VPN eliminates those tickets, where a chatbot only changes who the user talks to first.<\/p>\n

      The Short Version<\/h2>\n

      Ticket deflection measures a handoff and gets reported as an outcome. That is why Gartner finds 73% of people using self-service and only 14% finishing there. Track re-contact rate and cost per resolution alongside it, and the picture changes quickly.<\/p>\n

      If you want fewer tickets rather than fewer conversations, ask any vendor one question. Does the tool fix the condition on the endpoint, or hand the user an article? Anakage offers a free 30-minute demo for teams working that out, worth a look if it matches your setup. Book a demo<\/a>.<\/strong><\/p>\n

      Sources<\/h2>\n