Skip to content

AI resolution rate: how to measure it without fooling yourself

Resolution, deflection, transfer and reopen rates measure different things. See the formulas, an illustrative example and the traps that inflate an AI resolution rate.

Marlos Carmo

Founder of Tolky

Updated on November 16, 2026

10 min read
AI resolution rate: how to measure it without fooling yourself

Share

When a platform says AI resolves half of conversations, the right question is: resolves according to whom? For some, resolved means a conversation that ended without a person. For others, it means the customer did not come back within seven days. The two definitions produce very different numbers for the same operation.

The AI resolution rate is the most quoted and most poorly measured metric in customer service with AI agents. It invites error because the easiest way to raise it is to make the AI hold on to the customer rather than solve their problem.

This guide separates the four numbers that tend to get confused (resolution, deflection, transfer and reopen), shows how to calculate each and points out where the math tends to mislead. The values in the examples are illustrative.

Updated in November 2026.

What is the AI resolution rate?

The AI resolution rate is the percentage of conversations completed by artificial intelligence without needing a person on the team. The most common calculation is:

Resolution rate (%) = conversations closed by AI alone ÷ total conversations in the period × 100

Illustrative example: in one month, the operation had 1,000 conversations. The AI closed 600 without a transfer. The raw rate is 60%.

The word "raw" matters. That number only says the AI finished the conversation. It does not say the customer's problem was solved. For that, you need to look at what happened afterward, which is why the metric should never appear alone. The glossary sums up the rule: resolving a lot with dissatisfied customers is not a gain.

Resolution, deflection, containment and transfer: what is the difference?

The four terms measure different moments of the same conversation, and confusing them is the main source of inflated numbers. The table shows what each one answers.

MetricQuestion it answersShort formula
ResolutionDid the AI solve the problem?Closed by AI with no reopen ÷ total
ContainmentDid the AI keep the conversation, without transferring?Conversations without transfer ÷ total
DeflectionDid the contact stop reaching a person?Contacts resolved with no repeat contact ÷ total contacts
TransferHow often did the AI pass to a person?Transferred conversations ÷ total

Containment is the easiest number to inflate: the AI just has to not offer a transfer. A customer who gives up because they could not reach a person counts as "contained", and that is a failure, not a result. Resolution requires that the problem has ended.

Deflection is a volume view: how many contacts did not become human work. It is useful for sizing cost, but only honest when it discounts those who came back asking for the same thing. See the difference in detail in ticket deflection with AI.

How do you define what counts as "resolved"?

A conversation should count as resolved only when the reason for contact was addressed and the customer did not have to return about the same issue within a defined window. Write the rule before measuring. Four practical decisions:

  1. Closing signal. Did the customer confirm, did the flow reach its end, or did the conversation just go quiet? Silence is not confirmation.
  2. Reopen window. For how many days can a conversation resume on the same issue and still be considered unresolved? Pick a period and keep it.
  3. Type of case. An information question and a system action (a duplicate invoice, order status) do not share the same yardstick. Measure by issue category.
  4. Who decides. The classification should come from auditable rules or sample review, not only from the AI's own judgment.

Without these rules, two people looking at the same dashboard reach different rates. It is the same definition problem found in AHT, SLA and first response time.

Why is the reopen rate the reality check?

The reopen rate measures how many conversations that looked resolved came back, and it is the number that separates resolution from containment. The formula:

Reopen rate (%) = conversations resumed on the same issue within the window ÷ conversations closed by AI × 100

Continuing the illustrative example: of the 600 closed by AI, 90 came back about the same issue within seven days. The reopen rate is 15%, and net resolution is (600 − 90) ÷ 1,000 = 51%.

The raw number was 60%; the net is 51%. The nine-point difference is what the raw calculation hid. When someone compares platforms or periods, ask whether the number is raw or net of reopens.

A high reopen rate usually points to an incomplete answer, an outdated knowledge base or automatic closing for inactivity too early. To understand why a conversational AI agent behaves differently from a fixed-flow chatbot, read conversational AI is not a chatbot.

How do you read the transfer rate?

The transfer rate shows how often the AI recognized it needed a person, and a low rate is not a synonym for quality. An AI that never transfers may be holding cases it should pass on. One that transfers too much may be poorly trained.

What matters is the reason for the transfer. Group them into four causes:

  • Customer request: they wanted to talk to a person.
  • Lack of confidence: the AI had no basis to answer safely.
  • Business rule: the case requires approval or human judgment, such as policy exceptions.
  • Flow failure: the conversation got stuck or looped.

The first two are healthy when handled well. The fourth is a defect to fix. And the quality of the handoff counts: a transfer with no context makes the customer repeat everything and raises handling time. The 7 rules for AI to human handoff detail how to preserve context, and the term is in the glossary as handoff.

Traps that inflate the resolution rate

Most impressive resolution numbers come from a measurement choice, not a real improvement. The most common traps:

Counting the customer who gave up as resolved

Someone who abandons the conversation without an answer was not served. If the dashboard counts "no transfer" as success, abandonment becomes resolution.

Excluding from the denominator what the AI cannot do

If the calculation covers only the topics the AI handles, the rate rises, but the operation still has the rest in the human queue. Report both views: over the total and over the scope.

Mixing messages with conversations

Processed messages and resolved conversations are different units. A long conversation produces many messages and a single conversation. Do not apply the rate of one unit to the other.

Measuring without quality sampling

Automatic numbers need human review by sample. Read conversations marked as resolved and check whether the customer was really served. A weekly sample already reveals error patterns.

Ignoring satisfaction

High resolution with low satisfaction means the customer was contained, not served. Track a satisfaction measure, such as NPS, alongside resolution.

How do you build a reliable resolution dashboard?

A reliable dashboard reads five numbers side by side, by topic and by channel. The table below is a starting point.

IndicatorWhat it showsReading in context
Raw resolutionConversations closed by AI aloneStarting point, never alone
ReopensHow much of the "resolved" came backDefines net resolution
Transfer by reasonWhy the AI passed to a personShows where to adjust the knowledge base and rules
Time to transferHow long the customer waited before reaching a personDisguised containment shows up here
Satisfaction (NPS)How the customer rated the serviceValidates the resolution

With this dashboard, the decision stops being "did the rate go up or down" and becomes "on which topic does the AI resolve well, where does it transfer for good reason and where does it need new content". To connect these numbers to financial return, see how to measure the return of AI agents.

What does Tolky show about resolution?

On the platform, Tolky records more than 67% autonomous AI resolution, without human intervention, as an aggregate platform figure. It is a platform metric, not a promise of results for your operation and not the result of a specific customer. The definition, period and measurement method are described in the Tolky platform conversation report 2026.

Each operation's result depends on the topic, the quality of the knowledge base, the channel and the transfer rules. Tolky is a conversational AI CRM in which the AI resolves what it can and passes to the team, with the conversation's context, when the case calls for a person. The platform does not promise full automation, and it is worth talking to the team to understand how to measure resolution in your case before setting a target.

Frequently asked questions

What is the difference between resolution and deflection?

Resolution measures whether the AI completed the conversation. Deflection measures how many contacts stopped reaching a person. Deflection is only reliable when it discounts customers who came back asking for the same thing, because in that case the human work was merely postponed.

What is a good AI resolution rate?

There is no single number. The rate varies by topic: simple questions tend to have a high rate and complex complaints a lower one. Measure your baseline by category, track the trend and always look at reopens and satisfaction alongside.

Are containment and resolution rates the same thing?

No. Containment says the conversation was not transferred; resolution says the problem was solved. A customer who gives up counts as contained, but not as resolved.

How do I know whether the AI is resolving or just holding the customer?

Compare raw resolution with resolution net of reopens, read a sample of conversations and watch time to transfer and satisfaction. If reopens are high and satisfaction is low, the AI is holding.

Does transferring to a human worsen the resolution rate?

It lowers the raw number, but it is part of a well-designed service. Cases that require judgment or approval should go to a person. What you avoid is transfer with no context, which increases time and frustration.

Do Tolky's 67% resolution figures apply to my company?

Not as a promise. It is an aggregate platform figure. Each operation's result depends on topic, knowledge base, channel and rules. The path is to measure your own baseline.


Conclusion

A reliable AI resolution rate needs a written definition, reopens discounted, transfers read by reason and satisfaction alongside. Without that, the number rises without the customer noticing any improvement.

See the plans, read the platform report or talk to the team →. Let's talk?

Share

Tags

  • AI resolution rate
  • how to measure AI resolution
  • ticket deflection
  • handoff rate to human
  • reopen rate
  • AI agent metrics
  • containment vs resolution
  • #conversational CRM

About the author

Marlos Carmo

Founder of Tolky

Marlos Carmo is an AI entrepreneur and founder of Tolky, the conversational-era infrastructure and AI CRM that unifies intelligent service, multi-channel support (such as WhatsApp and voice), live CRM, and operational intelligence in a single ecosystem. He is a finalist for the SXSW Innovation Awards and a member of Francesco's Economy, a global network of young entrepreneurs focused on innovation and social impact. He works connecting Artificial Intelligence and digital transformation in projects for large organizations.

Next step

Want to bring this to your operation?

Tati, Tolky's AI assistant, understands your scenario in a few questions, suggests the best path and connects you with someone from the team, with full context.

Read also