A dashboard should show whether your customer service works. With an AI agent this gets harder: many conversations run without people, and the numbers often come from the agent’s vendor itself. This page shows which metrics matter, what each one measures exactly and how to tell whether a number is telling the truth.
The metrics at a glance
Most teams need only a few numbers from three groups: speed, resolution and the customer’s verdict. Each number comes with a cross-check, a second number that shows whether the first one misleads.
| Metric | what it measures | Cross-check |
|---|---|---|
| First reply time | Time from arrival to the first reply to the customer. At Zendesk, for example, up to the first public reply by a staff member. | The customer’s total waiting time, because a fast first reply can hide a long wait afterwards. |
| Resolution time | Time from arrival to resolution. There are two readings: up to the first resolution, or up to the last one if a case was reopened. | Reopened cases, because a short first resolution time is no use if the customer comes back. |
| First contact resolution | Share of requests resolved at the first contact. | A short customer survey, because internal counting often runs too high. |
| AI agent resolution rate | Share of conversations the agent resolved on its own. Every vendor counts this differently, see below. | Samples of conversations read by a person, and repeat contacts. |
| Handover rate | Share of conversations that go to people. More on the page When an AI agent hands over to people. | Handovers without reason and missing handovers. |
| Satisfaction (CSAT) | Share of customers who rate a contact as good. At Zendesk, for example, the share of tickets rated good out of all rated tickets. | The survey’s response rate, because few answers give a skewed picture. |
| Customer effort (CES) | How much effort customers had to put into their request. | Repeat contacts and channel switching. |
| Recommendation (NPS) | How likely customers are to recommend the company. | Better suited to the whole company than to a single contact. |
| Cost per contact | Service costs divided by the number of contacts in the same period. | Compare only with your own earlier number. More on the page What does an AI agent in customer service cost?. |
What “resolved” means for an AI agent
The agent’s resolution rate is the number that can mislead the most, because every vendor counts it differently:
- Intercom Fin counts a conversation as resolved when the customer confirms the answer was satisfactory, or when they leave the conversation without asking for further help. It counts as left after 24 hours without a reply. If the customer comes back in the same conversation, the resolution is deducted again.
- Zendesk only counts verified resolutions: cases without a takeover by a person, where in addition a language model checks the conversation text. The details are on the page What does an AI agent in customer service cost?.
- Salesforce Agentforce separates “deflection” (the conversation ended without a handover) from “task resolution” (the conversation was rated as fully resolved) and also tracks an escalation and an abandonment rate.
Our reading: numbers from different systems cannot be compared directly. “Without handover” does not mean “resolved”. Before the start, put in writing what counts as resolved for you, and measure it yourself.
Where dashboard numbers mislead
A dashboard can glow green while customers give up. Three findings show why:
- Self-service resolves less than it seems. In a Gartner survey of 5,728 customers in December 2023, only 14% of issues were fully resolved in self-service, and even for “very simple” issues only 36%. Self-service here means all self-help options, not only AI agents. The most common reason for failure: in 43% of cases customers could not find relevant content.
- People who have a bad experience switch channels. In a Gartner survey of 3,566 customers in February and March 2026, only 27% said they would try a chatbot again after a negative experience. They are more likely to choose another channel. Gartner therefore warns explicitly against building chatbots as “containment traps” that keep customers away from people. 87% of respondents said access to a human agent is essential when companies use generative AI in service.
- Working towards a number changes it. This is the so-called Goodhart principle, in the wording of the anthropologist Marilyn Strathern (1997): “When a measure becomes a target, it ceases to be a good measure.” It is a general principle, not a study of service metrics. For a dashboard it means: a single number as a target invites improving the number instead of the service.
Our conclusion: count repeat contacts across all channels as a cross-check on the resolution rate. Anyone who calls or writes on the same day after a chat was probably not resolved, even if the chat counts as resolved.
Benchmarks: what exists and what does not
Many benchmarks circulate online, for example for a “good” first contact resolution rate, satisfaction or reply time. For most of them we found no independent source with a disclosed method.
One exception is first contact resolution at the vendor SQM. SQM measures among its own clients, over 500 inbound customer service call centres in North America, with a survey after the call. For 2022 SQM gives an average of 71% and 70 to 79% as “good”. According to SQM, internal counting of first contact resolution runs 10 to 20% too high compared with the survey. This applies to phone in North America, not to chat, email, Germany or AI agents.
The standard for customer contact centres, DIN EN ISO 18295-1, also sets metrics only “where needed and as far as necessary”, according to its summary. We could not confirm any public target values from it.
The most reliable benchmark is therefore your own: measure your numbers before an AI agent starts and during the trial, and compare against that baseline afterwards.
Metrics per employee: involve the works council
In Germany, under section 87(1) no. 6 of the Works Constitution Act (BetrVG), the works council has a say in “the introduction and use of technical devices designed to monitor the behaviour or performance of employees”. According to the Federal Labour Court, “designed” means objectively suitable; the employer’s intention does not matter (order of 13.12.2016, 1 ABR 7/15). According to an order of 23.10.2018 (1 ABN 36/18), even an Excel sheet with times can be enough, and it does not matter whether the data is actually evaluated.
In Germany, if a dashboard records values per person or allows conclusions about individuals, it can be a technical device under section 87(1) no. 6 BetrVG. If there is a works council, it must then be involved, unless a statutory or collective agreement provision applies. This also applies in small teams if values can be attributed to individuals. Have this checked for your case.
In 2022 the German Data Protection Conference called for legal limits on performance and behaviour monitoring and for a general ban on permanent surveillance. That is a demand, not current law. This page is not legal advice.
How to build your dashboard
- Write down definitions: What counts as a contact, a resolution, a handover, and when does the clock start? Without definitions you compare numbers that mean different things.
- Measure a baseline: Record your numbers before the agent starts. Only then can you see later what has changed.
- A cross-check for every number: resolution rate with repeat contacts, first reply with waiting time, satisfaction with response rate.
- Values per person only after clarification: Clarify beforehand whether you show values per employee, with the works council if there is one.
- Read conversations: Numbers show where something stands out. Whether a case was really resolved only shows in samples of conversations read by a person. How to do this is under Measuring handovers.
Your next step
Want to know which metrics suit your customer service? We talk about it in a first call of 30 minutes. Get in touch
What could AI take off your plate?
In a first call (30 minutes) we look at your work and tell you openly what is worth it and what is not.
Frequently asked questions.
Which metrics belong on a customer service dashboard?
A few from three groups: speed (first reply time, resolution time), resolution (first contact resolution, the AI agent's resolution rate, handover rate) and the customer's verdict (satisfaction, effort, recommendation). Each number comes with a cross-check.
What is a good first contact resolution rate?
There is no generally valid value. The vendor SQM gives an average of 71% for inbound phone call centres in North America in 2022, measured with a survey after the call. For chat, email and AI agents we found no confirmed value. Compare with your own baseline.
Is a high automation rate good?
Not by itself. A conversation without a handover is not automatically resolved; the customer may also have given up or switched channels. Check the rate against repeat contacts and samples read by a person.
CSAT, CES or NPS: what should I measure?
Satisfaction (CSAT) and customer effort (CES) fit a single contact, recommendation (NPS) fits the whole company better. CES measures perceived effort. Repeat contacts and channel switches are better measured separately.
Does a service dashboard need the works council?
In Germany, if a dashboard records values per person or allows conclusions about individuals, it can be subject to co-determination under section 87(1) no. 6 BetrVG. If there is a works council, it must then be involved, unless a statutory or collective agreement provision applies. Have this checked for your case; this is not legal advice.
How is your website doing?
The check is free and without obligation. You get the report by email.
Sources.
- Zendesk Help Center: About native Support time duration metricsVendor / primary · As of
- Zendesk Help Center: First reply time and requester wait timeVendor / primary · As of
- Zendesk Help Center: Analyzing your Support ticket activity and agent performanceVendor / primary · As of
All 16 sources
- Zendesk Help Center: About automated resolution tiersVendor / primary · As of
- Intercom: Fin AI Agent resolutionsVendor / primary · As of
- Salesforce Trailhead: Monitor Agent Metrics and ScoresVendor / primary · As of
- SQM Group: What Is a Good FCR Rate? (2022)Vendor / primary · As of
- Bain & Company: Measuring your Net Promoter ScoreVendor / primary · As of
- Harvard Business Review: Stop Trying to Delight Your Customers (Dixon, Freeman, Toman)Vendor / primary · As of
- Gartner: Only 14% of customer service issues are fully resolved in self-serviceVendor / primary · As of
- Gartner: Only 27% of customers would try a chatbot again after a negative experienceVendor / primary · As of
- German Works Constitution Act (BetrVG), section 87Vendor / primary · As of
- Federal Labour Court, order of 13.12.2016, 1 ABR 7/15Vendor / primary · As of
- Federal Labour Court, order of 23.10.2018, 1 ABN 36/18Vendor / primary · As of
- German Data Protection Conference: resolution on employee data protection (29.04.2022)Vendor / primary · As of
- DIN Media: DIN EN ISO 18295-1:2017-10Vendor / primary · As of
- Zendesk Help Center: time metrics, first reply and wait time, satisfaction, automated resolution tiers
- Intercom: Fin AI Agent resolutions, as of 17.09.2026
- Salesforce Trailhead: Monitor Agent Metrics and Scores
- SQM Group: What Is a Good FCR Rate? (2022), the vendor’s survey among its own clients
- Bain & Company: Net Promoter Score
- Harvard Business Review, July/August 2010: Stop Trying to Delight Your Customers, origin of the Customer Effort Score
- Gartner, 19.08.2024: self-service and Gartner, 02.09.2026: chatbots after a negative experience
- Marilyn Strathern: “Improving ratings”: audit in the British University system, European Review 5(3), 1997, pp. 305–321
- Section 87 BetrVG, BAG 1 ABR 7/15, BAG 1 ABN 36/18
- German Data Protection Conference: resolution on employee data protection, 29.04.2022
- DIN Media: DIN EN ISO 18295-1:2017-10



