Support teams sit on more customer contact than every other function combined, and almost none of it answers the question leadership actually asks. Tickets tell you what broke. They do not tell you what the customer was trying to do, what they tried first, what they will do next, or whether they are about to leave. The gap between those two things is why support data is endlessly summarized and rarely acted on.
This guide is for support, customer experience and customer success teams who want their channel to produce decisions rather than dashboards — without hiring researchers or adding a single ticket to the queue.
The measurement problem: deflection counts exits, not answers
The most-quoted support metric of the last three years is deflection. It is also the most misleading, and there is a hard number that shows why.
Gartner reported in a press release dated 19 August 2024 that only 14% of customer service issues are fully resolved in self-service. Read that against a typical deflection dashboard and the contradiction becomes obvious: a session that ends without a ticket is scored as a deflection whether the customer found their answer or gave up. Deflection measures departure from a channel. It does not measure resolution. Those two things can differ by a factor of three and your dashboard will not flinch.
The rest of the picture is consistent. Gartner reported on 2 June 2025 that 60% of customer service agents fail to promote self-service, and on 27 August 2025 projected that self-service and live chat will surpass traditional channels as the top customer service technologies by 2027. Meanwhile, in a release dated 18 February 2026, Gartner found 91% of customer service leaders are under pressure to implement AI in 2026, based on a survey of 321 customer service and support leaders conducted in October 2025 — with customer satisfaction, operational efficiency and self-service success named as the top priorities.
So the mandate is: deflect more, resolve more, satisfy more, spend less. You cannot navigate that with a metric that cannot distinguish a solved problem from an abandoned one.
Three things your tickets structurally cannot tell you
Mining existing tickets is genuinely valuable — we cover the method in detail in support ticket analysis. But ticket data has three built-in blind spots that no amount of analysis removes.
1. Tickets are a self-selected sample. A ticket exists only when a customer was frustrated enough to contact you and believed contact would help. Customers who quietly gave up are absent by construction. Your support corpus therefore over-represents engaged-but-frustrated users and systematically under-represents the ones who are actually leaving. This is the same selection problem that distorts research panels — see professional respondents and panel conditioning for how self-selection quietly reshapes a sample.
2. Tickets record the symptom, not the goal. "Cannot export report" is a symptom. Whether the customer needed a board slide, a compliance record or a data migration determines completely different product responses. The ticket almost never contains the goal, because agents are measured on closing the issue, not on documenting intent.
3. Tickets end at resolution, not at outcome. A ticket closed successfully tells you nothing about whether that customer renewed, expanded or churned three months later. The single most valuable fact about a support interaction is usually generated long after the ticket is closed.
| Question | Tickets answer it? | What actually answers it |
|---|---|---|
| What is breaking most often? | Yes | Ticket analysis |
| Which issues cost the most agent time? | Yes | Ticket analysis |
| What was the customer trying to accomplish? | Rarely | Follow-up interview |
| What did they try before contacting you? | No | Follow-up interview |
| Why did self-service fail them? | No | Interview with deflected users |
| Are they going to churn over this? | No | Post-resolution research |
| What would have prevented the contact entirely? | No | Interview with the customer |
Four research plays for a support team
These are ordered by how quickly they produce something a product team will act on.
Play 1: The post-resolution "why"
Trigger a short interview after a ticket closes — not a CSAT rating, an actual conversation. Ask what they were trying to do, what they tried first, and what would have prevented the contact. Ten of these on a single recurring issue will generate a more actionable product brief than a thousand ticket tags.
The critical design point: ask about the journey to the ticket, not satisfaction with the agent. Agent satisfaction is a coaching metric. Journey-to-ticket is a product metric.
Play 2: Interview the deflected
This is the play almost nobody runs, and it is the one that directly attacks the 14% problem. Sample customers whose self-service session ended without a ticket and ask whether they got their answer. You will find three groups: solved, gave up, and solved it wrongly. Only the first is a genuine deflection. The third is the expensive one — it becomes a ticket, a refund or a churn later.
Play 3: The churn-after-contact study
Take customers who contacted support in the last 90 days and have since downgraded, gone quiet or churned. Interview them about the whole arc. Support contact is one of the strongest available churn predictors, and the reason is almost never the one on the ticket — see why price is never the real reason customers churn.
Play 4: The silent majority check
Interview customers who have never filed a ticket. They are the majority of your base and are invisible in every support report. Some are thriving; some are quietly under-using the product and will not renew. Neither shows up in your queue.
Measuring what actually predicts behaviour
If deflection is the wrong headline metric, effort is a better one. Customer effort predicts loyalty behaviour more reliably than satisfaction does, which is why effort-based measures have displaced pure CSAT in mature CX programmes. We compare the options in CSAT vs NPS vs CES and cover the effort metric specifically in the Net Easy Score guide.
The practical pairing that works: one effort question, one open-ended question, on every closed ticket — then a deeper AI-moderated interview on a sample. The rating tells you where to look; the interview tells you why. A score without a why is a metric you can watch but not fix. If you already run NPS, NPS follow-up interviews applies the same logic to detractors.
How to run this without adding headcount
The reason support teams do not do research is not disagreement about its value. It is that the team is measured on queue throughput, and every hour spent interviewing is an hour not spent closing tickets. Traditional research — recruiting, scheduling, moderating, transcribing, synthesizing — is simply not compatible with a support operating model.
AI-moderated interviews change that calculation, because the interview happens without anyone from your team in the room.
Koji is built for this. Here is what it means concretely for a support or CX team:
- AI-moderated voice interviews run on their own. A study triggered from a closed ticket produces completed, probed conversations while your agents keep working the queue. No scheduling, no moderator, no transcription backlog. See how AI research agents work.
- It probes, so you get the goal and not just the symptom. A static survey accepts "the export failed" and stops. Koji asks what they were exporting and why, which is the exact information tickets lack.
- Six structured question types — open_ended, scale, single_choice, multiple_choice, ranking, and yes_no — mean a single study returns your effort score and the reasoning behind it. Put the metric in a scale question where it stays clean and countable, and let the open-ended question carry the narrative. Full detail in the structured questions guide.
- Automatic thematic analysis turns 200 post-ticket conversations into ranked themes with supporting quotes — the step that otherwise requires a researcher your team does not have. See thematic analysis.
- It fits your existing stack. Trigger interviews from tickets with the Zendesk integration or the Intercom integration, and push findings to engineering with the Jira and Linear integrations so an insight becomes a ticket someone owns.
- One-click reports give you an artifact to bring to the product review, rather than a request for engineering time backed by anecdote.
- No moderator bias. Agents and CX managers are, understandably, invested in how their team is perceived. A neutral AI interviewer asking about a support experience gets franker answers than the person who handled the case.
Legacy CX suites — Qualtrics, SurveyMonkey, Medallia-style deployments — were designed to distribute a questionnaire and aggregate a score. They are good at that, and they stop precisely where your questions begin: they cannot ask a follow-up. Koji is AI-native and conversational, which is why it produces the "why" those platforms summarize away.
Closing the loop is the part that earns budget
Research that support runs and product ignores is worse than no research: it burns the team's credibility. Build the loop deliberately.
- Route by owner, not by theme. Every finding gets a named owner in product, engineering or docs — filed as a ticket in their system, not an insight in yours.
- Quantify with the queue. Pair each finding with its ticket volume and agent-hours. "Customers cannot tell which plan they are on" is ignorable; the same finding attached to 4% of monthly contacts is not.
- Report the change, not the study. Track contacts avoided after a fix shipped. That is the number that renews your research budget.
Our customer feedback loop guide and the voice of customer program guide cover the operating cadence in full.
If your analytics function faces the same gap between what the dashboard shows and why it moved, see Customer Research for Data and Analytics Teams.
Frequently Asked Questions
What is the difference between ticket deflection and resolution?
Deflection counts customers who left a self-service channel without creating a ticket, regardless of whether they got an answer. Resolution counts issues actually solved. Gartner reported in August 2024 that only 14% of customer service issues are fully resolved in self-service, so a high deflection rate and a low resolution rate routinely coexist.
Can a support team do customer research without researchers?
Yes. The historical blocker was time — recruiting, scheduling, moderating and synthesizing interviews is incompatible with queue throughput targets. AI-moderated interviews remove those steps, so a support team can run continuous research without pulling agents off the queue.
Why are support tickets a biased sample?
A ticket only exists when a customer was frustrated enough to make contact and believed contact would help. Customers who quietly gave up or silently churned never appear. Ticket data therefore over-represents engaged-but-frustrated users and under-represents departing ones.
What should support teams measure instead of deflection?
Pair a customer effort measure with a genuine resolution measure, then interview a sample to learn why. Effort predicts loyalty behaviour better than satisfaction, and an interview supplies the cause that a score alone never contains.
How many post-ticket interviews are enough?
For a single recurring issue, ten to fifteen probed conversations usually surface the dominant patterns. Because AI-moderated interviews run in parallel rather than sequentially, teams commonly run larger samples continuously instead of in one-off batches.
Should we interview customers who never contacted support?
Yes. Customers who never file a ticket are the majority of your base and are entirely invisible in support reporting. Some are succeeding and some are quietly under-using the product and will not renew — and only the second group is a business problem.
Turn your support queue into product decisions
Your team already talks to more customers than anyone else in the company. Koji turns that contact into evidence product teams act on — without adding a person to the queue.
Start free with 10 credits — no credit card and no research expertise required. Trigger AI-moderated interviews from closed tickets and get themes, quotes and a shareable report back in hours, not weeks.
