Chatbot resolution rate benchmarks: what good looks like in 2026

Ask five vendors what a good chatbot resolution rate is and you will get five numbers, because each of them is counting something different. One independent survey of 5,728 customers found that only 14 percent of support issues were fully resolved in self-service, while one vendor says its customers average a 76 percent resolution rate. Both can be true, and neither tells you what to expect in your own inbox.

This post collects the published figures in one place, explains why they differ, and gives a realistic way to set your own target. We have not run a benchmark of our own, and nothing here is Klevere data. Every number below comes from the source linked beside it, with the year and the scope stated so you can judge how far it travels.

Quick answer

There is no single good resolution rate. Independent research shows 14 percent of issues fully resolved in self-service, while vendors report averages between 41 and 76 percent of conversations on their own platforms. Measure resolution as a confirmed outcome, not containment, set a baseline from your own transcripts, and judge progress against that.

Resolution, containment and deflection: three different measures

Most disagreement starts with definitions. Containment is the share of conversations a bot handles without passing the customer to a human. Fin's glossary gives the formula as conversations handled without a human transfer divided by total conversations handled by automation, and its worked example is 10,000 conversations with 2,500 transferred, which gives 75 percent. The same page points out that containment counts abandoned, trapped and rerouted conversations as successes, because it only records that no human was involved.

Resolution rate asks a harder question: was the problem actually solved? On the same glossary page, Fin counts a conversation as resolved when the customer confirms the answer helped or does not ask for more help after the answer. That is a looser test than a confirmed fix, but it is still stricter than containment. Zendesk uses a similar idea for billing. Its support documentation says that paying per automated resolution means you pay only for customer requests that were successfully resolved by the AI agent without escalation to a human, and that a large language model verifies resolved conversations.

Deflection is the third term and the loosest. It usually means a customer who would have raised a ticket did not, perhaps because they read an article or talked to a bot. It says nothing about whether they got what they needed. A customer who gives up and phones instead may still count as deflected from the ticket queue.

Here is how the three compare in practice.

When you read any benchmark, check which of these it reports before comparing it with anything else. A 70 percent containment rate and a 70 percent resolution rate describe very different customer experiences.

What the published figures say

The list below sets out the sources we could verify, with the figure, year and scope for each. Read the scope first. Some figures measure issues, some measure conversations, some measure opinions, and one is a projection.

Notice how little overlap there is. The independent figures measure customer experience across channels, the vendor figures measure conversations on the vendor's own platform, and the Salesforce figure measures what staff believe. None of them is a controlled measure of a chatbot resolving the same set of questions.

Why the numbers disagree

Four differences explain most of the gap between 14 percent and 76 percent.

The first is the unit. Gartner counts issues across a customer's whole journey, including complex ones. Vendors count conversations their bot received, which already excludes the customers who went straight to a phone line or email because they knew the problem was complicated. A bot that never sees the hard cases will look better than one that sees everything.

The second is selection. A vendor average is taken across customers who bought the product, kept it, and, in many cases, tuned it with the vendor's help. A survey of consumers includes every poorly configured bot they have met. Neither is wrong, but they answer different questions.

The third is the definition of resolved. As the Fin glossary notes, a conversation can count as resolved if the customer does not ask for more help. Silence is not always satisfaction. Some customers leave because they gave up. A stricter definition, such as an explicit confirmation or the absence of a repeat contact within a set period, produces lower figures.

The fourth is complexity. The CFPB's review says chatbots may be useful for resolving basic inquiries, but their effectiveness wanes as problems become more complex. It also describes what it calls doom loops, where automated responses fail to resolve an issue and lead customers round repetitive answers without an offramp to a human. Gartner's own split points the same way, since even very simple issues were fully resolved in self-service only 36 percent of the time. Your mix of simple and complex questions will move your rate more than the choice of tool.

What a realistic range looks like

Our reading of the evidence, not a benchmark we have measured, is that a first deployment should aim to resolve a meaningful share of one narrow, well-documented category of questions rather than a headline percentage of everything. The independent data says broad self-service resolves a small share of issues. The vendor data says well-tuned deployments on selected customers can reach the 40 to 76 percent range of conversations. The gap between those two is where most real projects sit, and where you land depends on how repetitive your enquiries are and how good your source material is.

A practical way to set a target is to start from your own history. Sample a few hundred past conversations, label each one as answerable from existing documentation, answerable only with account data, or needing human judgement, and count them. The first category is your realistic ceiling for an assistant that reads documents. The second becomes reachable once the assistant is connected to your systems. The third should stay with people, and a good deployment makes that handover easy rather than hiding it.

Two warnings apply to any target. First, a rising resolution rate that comes with falling customer satisfaction usually means the bot is closing conversations rather than solving problems. Second, the Qualtrics finding that nearly one in five consumers saw no benefit shows that customers notice when automation is a barrier. Half the consumers in that survey also said they were concerned AI would stop them connecting with a human, so the route to a person needs to stay visible.

Productivity is a different benchmark

Some of the strongest evidence for AI in support does not measure resolution by a bot at all. In the NBER working paper Generative AI at Work, Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied 5,179 customer support agents who were given a generative AI assistant in stages. Productivity, measured as issues resolved per hour, rose by 14 percent on average, with a 34 percent improvement for novice and low-skilled workers and minimal impact on experienced workers.

That result is about agents assisted by AI, not customers served by a bot, so it should not be mixed into a chatbot resolution benchmark. It does matter for planning, because it suggests that the quickest gains for some teams come from drafting and retrieval support for staff, with autonomous resolution layered on later for the categories that suit it. If you are weighing the two approaches, our comparison of AI and human customer support covers when each makes sense.

How Klevere approaches resolution targets

Klevere builds custom AI agents for small and mid-sized businesses, with more than 500 agents deployed across 50 plus projects and 12 industries. We do not start with a target percentage. We start by reading your real conversations and sorting them into the three categories described above, because that tells us what an agent can responsibly take on in the first phase.

We then define resolved in writing before anything goes live. For most clients that means an explicit confirmation from the customer, or no repeat contact on the same issue within an agreed window, plus a sample review by a person each week. Containment is tracked as a secondary figure, never as the headline, for the reasons the Fin glossary gives.

Escalation is designed as part of the product. Every agent has clear conditions for handing over to a person, passes the full conversation across, and logs why it escalated, so the reasons feed the next round of improvements. Our AI support agent page describes the product, and our article on what can go wrong with AI agents in production explains the failure patterns we plan for, including the loops the CFPB describes.

For phone and booking channels, resolution looks different, because the outcome is often an appointment or a completed task rather than an answer. Our guides to AI voice agents for business phone calls and AI appointment booking cover how we measure those. If you want to connect resolution to money, our guide to measuring AI ROI shows how to translate resolved conversations into time and cost.

How we compiled this

We searched for published figures on chatbot and AI support resolution, containment, deflection and related outcomes, then opened each source to confirm that the figure appears on that page. We included regulators, research firms with published samples, an academic working paper, and vendor figures where the vendor's own page states them, labelling vendor figures as such. Sources span the United States, the United Kingdom, Europe and a further set of countries in the Qualtrics and Salesforce samples.

We excluded figures we could not trace to a page, anything from forums or aggregator blogs, and any number that mixed definitions without saying so. Two sources, Gartner's press release and a Zendesk per-resolution price, could not be opened directly in our checks, so we relied on trade press for the Gartner figures and left the price out. Several figures are one to three years old, and the dates are stated above. Treat the whole set as a guide to the shape of the evidence rather than a precise benchmark.

Frequently asked questions

What is a good chatbot resolution rate?

There is no universal answer. Independent research found 14 percent of issues fully resolved in self-service, while vendors report averages of 41 to 76 percent of conversations on their own platforms. A good rate for you depends on how repetitive your enquiries are and how resolution is defined. Set a baseline from your own transcripts first.

What is the difference between resolution rate and containment rate?

Containment is the share of conversations handled without a human, whether or not the customer got an answer. Resolution rate checks that the problem was actually solved, usually by customer confirmation or by no further request after the answer. Containment counts abandoned conversations as successes, so it flatters weak bots.

How do you calculate chatbot deflection rate?

Deflection is usually the number of customers who used the bot or self-service and did not raise a ticket, divided by total contacts. It estimates avoided workload, not satisfaction. To keep it honest, track repeat contacts on the same issue within a set period, since a customer who gives up and phones is not truly deflected.

Why do vendor resolution rates look higher than independent surveys?

Vendors count conversations their bot received, which excludes customers who chose another channel for complex problems. Their averages also come from customers who kept the product and tuned it. Independent surveys count issues across the whole journey. The units and populations differ, so the numbers should not be compared directly.

Does a higher resolution rate always mean better customer service?

No. A bot can raise its resolution rate by closing conversations early or making human help hard to reach. Qualtrics found nearly one in five consumers who used AI for customer service saw no benefit. Pair resolution with satisfaction scores, repeat contact rates and regular human review of sampled conversations.

How long does it take to improve a chatbot's resolution rate?

It depends on your source material and scope. Improvements usually come in steps: fixing gaps in documentation, connecting the bot to account data, and refining escalation rules. Review a weekly sample of unresolved conversations and fix the most common cause each time. Expect gradual gains rather than a single jump.

If you are unsure what your own baseline looks like, a free AI audit is a practical place to start. We review your support conversations, sort them by what an agent can responsibly handle, and tell you where resolution is realistic and where people should stay in the loop.