Many customer service teams now use AI and see rising resolution rates from their artificial colleagues in dashboard metrics and AI Analytics. But why does the human team still have just as much work as before? Our guide to measuring AI performance in customer service sheds light on what these numbers really mean.
TL;DR: AI performance in customer service cannot be measured by automated responses alone. Metrics such as Verified Resolutions, repeat contacts and the actual reduction in workload for the support team are what matter. Only the combination of these KPIs shows whether AI is truly performing successfully.
Back to the Resolution Rate: It counts completed conversations. But if a customer simply stops replying out of frustration after receiving an unsatisfactory answer, the conversation appears exactly the same as one in which the issue was genuinely resolved. The more interactions are automated, the greater the impact of this distortion, and customer satisfaction can decline despite all the green numbers.
Gartner has measured this: Only 14% of customer service issues are fully resolved through self-service, including AI. What happens to the rest? Channel switching and silent customer churn.
There is therefore no way around identifying the right metrics if you do not want AI in customer service to push your company in the wrong direction. Customers leave a brand after just one unresolved interaction, according to the Zendesk CX Trends 2026. Counting that as a success risks nothing less than the company’s long-term performance.
The right KPIs for AI Analytics
A meaningful assessment of AI performance in customer service essentially comes down to 4.5 questions:
- How many conversations does the AI handle?
- Which of these does it resolve completely, and which does it escalate to the team?
- How many customers contact you again afterwards about the same issue?
- And how much work does the AI actually save the team in escalated cases?
Example: Zendesk AI
This is broadly what the Zendesk dashboard for AI agents shows: Conversations, Automated Resolutions, Escalations and Dropped Conversations, supplemented by customer ratings and the content used.
Dropped Conversations deserve particular attention: They are recorded neither as automated resolutions nor as escalations and may therefore indicate abandoned interactions, comprehension issues or missing escalation paths.
These numbers only become meaningful when you compare them with reality: Is the overall ticket volume decreasing for the topics handled by AI? Are tickets reopened less frequently? Are repeat contacts declining?
You should also evaluate repeat contacts across channels wherever possible. A customer who calls or sends an email after an AI chat will not necessarily appear as a reopened conversation.
Traditional metrics such as Suggestion Rate or Click-through Rate can help identify fundamental issues such as missing content or gaps in coverage, but they are no longer the core metrics for evaluating modern Zendesk AI. The relevant KPIs also depend on the tasks you have assigned to your AI. You can find an overview on our AI in customer service page.
Contained, Verified, Assisted: Three different outcomes
Zendesk now distinguishes between different conversation outcomes. It is worth examining them separately, because each one reveals something different about what actually happened during the interaction:
Contained Resolution
A Contained Resolution means that the customer did not request any further assistance after receiving the AI response. This is a useful signal for topics with high automation potential, but it is not proof that the issue was resolved. The metric becomes far more meaningful when combined with repeat contacts, reopened tickets and changes in ticket volume.
After 72 hours without another request, Zendesk also uses an LLM to review the conversation. If the review does not confirm a satisfactory resolution, the conversation remains classified as a Contained Resolution.
Verified Resolution
A Verified Resolution goes one step further: After 72 hours without another request, an LLM reviews the conversation and confirms that the issue was resolved fully and satisfactorily. Broken down by contact reason, channel and request complexity, this metric provides a much more reliable indication of where automation is genuinely effective.
Important context: Since May 2026, Zendesk has counted both Contained and Verified Resolutions towards the Automated Resolution Rate. The rate can therefore increase even if the proportion of demonstrably resolved issues remains unchanged.
Assisted Escalation
An Assisted Escalation describes a case in which the AI has already collected information and, ideally, completed useful preparatory work before a human agent takes over. This is often the best possible outcome: For complex issues such as complaints, individual contract questions or technical problems, a well-prepared handover is more valuable than forced and incomplete end-to-end automation.
The intent has been identified and the relevant information is already available, allowing the human agent to start with the necessary context. With solutions such as Aircall, this is already possible over the phone as well.
However, the status alone does not show whether this creates actual workload reduction. You also need to examine whether handling times and follow-up questions decrease and whether the information handed over is complete and useful.
Conclusion: A high automation rate alone provides little value. Good AI Analytics show whether AI reliably resolves suitable issues, prepares more complex cases effectively and genuinely reduces repeat contacts and additional work.
Learn more about Zendesk Advanced AI Agents
What good results look like
There is no single correct target value for customer service, just as there is no universally correct customer service model without considering the specific company. Benchmarks from other organisations are also rarely directly transferable. An online shop with a high volume of standard requests about delivery status, returns or password resets has fundamentally different automation potential from a B2B company whose support team mainly handles individual technical questions and contract discussions.
It is more useful to distinguish between different request types:
- Which topics can be fully automated?
- Which should be prepared by AI and then handed over?
- Which should go directly to the support team from the outset?
Meaningful benchmarks are therefore usually created internally: Compare the same contact reasons before and after automation and separate the results at least by channel, language and complexity.
A low overall rate may indicate missing content, but it may equally mean that the selected topics are simply too complex for full AI automation. Conversely, a high rate that hides a large number of repeat contacts does not represent workload reduction either. It is merely an additional, unproductive interaction that costs your company money and frustrates customers.
Good to know: Since May 2026, advanced AI agent features have been included in all Zendesk Suite and Support plans. Each plan includes a base allowance of Automated Resolutions. However, only Verified Resolutions count towards this allowance; Contained Resolutions and Assisted Escalations do not. Reporting and billing metrics are therefore not fully aligned. Source: Zendesk
Common misinterpretations of AI dashboards
- high Resolution Rate ≠ resolved issues
- few escalations ≠ successful AI
- many Automated Resolutions ≠ reduced workload
- green KPIs ≠ satisfied customers
Whether AI genuinely reduces the workload in customer service cannot be determined from the AI dashboard alone. What happens in the rest of the support operation matters just as much: Is ticket volume decreasing for automated contact reasons? Is handling time falling for escalated cases? Does the team require fewer agent hours overall? Only when AI KPIs and traditional service metrics move in the same direction does genuine workload reduction occur rather than merely better-looking reports.
Where to optimise effectively
A common response to weak AI KPIs is to create more content: more help centre articles, more FAQs and more guidelines. Unfortunately, this often has little effect because the problem lies in the quality of the existing content rather than the quantity.
Customer service AI works with the knowledge you provide, and a help centre written for humans is not automatically ready for AI self-service. Your articles should therefore be unambiguous, up to date and action-oriented. Internal terminology that is obvious to the support team but meaningless to customers is one of the most common problems, because AI uses these terms without questioning them.
Instead of launching a broad content initiative, it is usually more effective to focus on the most frequent contact reasons: Are they covered properly? Are the next steps clear? It is also worth reviewing individual conversations: Where do customers abandon the interaction? When would an earlier escalation have been more appropriate? Which topics regularly produce Verified Resolutions and could be expanded further?
Automated and Verified Resolutions can be monitored continuously. A deeper review by contact reason and content gap is better conducted at longer intervals, otherwise there is a risk of mistaking isolated outliers for recurring patterns.
Putting your own numbers into context
When customer service AI does not deliver the expected results, the problem usually lies in the interaction between different elements: Content, processes, routing and reporting are not working together cleanly and logically. What the dashboard shows may be true, but it is not the whole truth.
As one of the most experienced customer service consultancies and a Zendesk Premier Partner licensed for AI, we encounter situations like these regularly. Whatever tool you use, the best first step towards improvement is usually a customer service clarity workshop . Two experts, one day, and a clear understanding of where you stand and what to do next.
In the meantime, if you would like to explore the topic further, we particularly recommend our customer service strategy page and our guide to genuine customer service efficiency . Or speak to us directly about your questions and goals:
Strategy + Implementation: What sets Leafworks apart
Leafworks helps businesses of all sizes to make effective use of AI in customer service – from strategy through to implementation and ongoing optimisation. Having worked on nearly 1,000 client projects, we know which metrics really make a difference and which adjustments you need to make to ensure that AI doesn’t just produce good reports, but actually takes the pressure off your team.
The first step? Clarity! If you’d like to find out more about how we take a holistic approach to customer service and improve it in the long term, we also recommend our customer service strategy page.
Frequently asked questions about AI KPIs in customer service
Which KPIs are most important for AI in customer service?
The most important KPIs include Automated Resolutions, Verified Resolutions, Assisted Escalations, repeat contacts and ticket volume by contact reason. Only by looking at these metrics together can you determine whether AI is actually resolving customer issues and reducing the workload for your support team.
What is a Verified Resolution?
In Zendesk, a Verified Resolution is a confirmed automated resolution. After the customer has not requested any further assistance within a defined period, a Large Language Model (LLM) reviews the conversation. Only if the issue is assessed as genuinely resolved is the conversation counted as a Verified Resolution.
What is the difference between a Verified Resolution and a Contained Resolution?
A Contained Resolution means the customer did not request any further assistance after the AI response. A Verified Resolution goes one step further: an LLM additionally confirms that the issue was most likely resolved successfully. This makes it the more meaningful metric.
How does Zendesk AI measure success?
Zendesk provides metrics such as Conversations, Automated Resolutions, Verified Resolutions, Assisted Escalations and Dropped Conversations. However, these figures only become meaningful when combined with traditional customer service KPIs such as ticket volume, handling time, repeat contacts and customer satisfaction ratings.
Why isn't the Resolution Rate enough?
The Resolution Rate only shows how many conversations were closed. It does not distinguish between issues that were genuinely resolved and conversations that customers abandoned out of frustration. Metrics such as Verified Resolutions, repeat contacts and changes in ticket volume are needed to determine whether AI is actually improving customer service.
