AI Resolution Rate: How to Define and Measure Automated Resolution

By Linh Ngo · Updated Oct 11, 2026

AI resolution rate is the share of customer conversations that an AI handled from start to finish, without a human stepping in, and where the customer's actual problem was solved. The second half of that sentence is the one that gets dropped. A conversation that never reached an agent is contained. A customer who never opened a ticket is deflected. Neither tells you the problem went away.

That gap matters because the three numbers can differ by a wide margin on the same traffic, and every vendor draws the line in a slightly different place. This guide separates the terms, shows how three major platforms count a resolution in their own documentation, and walks through a way to measure it that you can defend in a budget meeting. It sits inside the AI for Customer Service collection, so start with the overview of AI in customer service if you want the wider picture.

Key Facts

• Containment, deflection and resolution measure different things: whether a human was avoided, whether a ticket was avoided, and whether the problem was solved. Only the third is an outcome.
• Intercom counts a Fin resolution as confirmed (the customer says the answer helped) or assumed (the customer leaves without asking for more help) (Intercom).
• On its automated-resolution platform in place before resolution tiers launched on May 18, 2026, Zendesk evaluates a messaging conversation 2 hours after the first message by default, extendable to 72 hours, and an email conversation 72 hours after the first email; an LLM verifies each counted resolution (Zendesk).
• Ada requires a resolved conversation to be relevant, accurate, safe and contained, and marks escalated conversations Not Resolved immediately (Ada).
• The CFPB's chatbot report cites a survey in which 78% of consumers who interacted with a chatbot needed to connect with a human after it failed to serve their needs (CFPB).

Containment, Deflection and Resolution: Three Different Questions

Think of a customer who opens your chat widget at 11 p.m. and asks why their invoice shows two charges. Three things can happen, and each one gets counted differently.

Containment asks: did a human get involved? If the bot talked to the customer and no agent joined, the conversation is contained. It says nothing about whether the customer got a useful answer. Ada's documentation makes the point directly: containment shows the share of conversations that didn't reach a human agent, but it can't show whether the customer was satisfied or gave up (Ada).

Deflection asks: did a ticket get avoided? It's usually measured on self-service, such as the help center and article suggestions. If a customer reads an article and doesn't file a ticket, that's a deflection. Anton de Young, writing on Zendesk's own blog, defines the classic self-service score as total help center users divided by total users in tickets, so a 4:1 result means four people try self-service for every one who submits a request. He also admits that Zendesk once hit an almost 40:1 ratio that still didn't prove the help center was preventing tickets (Zendesk blog). A reader who found the answer and a reader who gave up look identical in that number.

Resolution asks: was the problem solved? This is the only one of the three that describes an outcome. It needs a judgment about the conversation, either from the customer (they confirm), from a reviewer, or from a model reading the transcript.

The relationship is a funnel. Every resolved conversation is also contained, but plenty of contained conversations aren't resolved. If your dashboard shows one number labelled "automation rate" or "AI success", find out which of the three it is before you quote it.

How Vendors Count Resolution Differently

The big helpdesk and AI-agent platforms all publish a resolution metric, and they don't agree on the rules. The differences below come from each vendor's own documentation, so they describe how that product counts, not how the industry should.

Intercom (Fin). A resolution is counted after Fin's last answer when the customer either confirms the answer was satisfactory (a confirmed resolution) or exits the conversation without requesting further assistance (an assumed resolution). Intercom's article lists what isn't counted: a reply to a greeting, a clarifying question the customer never answers, and conversations where Fin says it has no answer and a teammate takes over. A resolution can also be reversed if the customer returns to the same conversation for more help (Intercom). The notable design choice is the assumed resolution: silence counts as success.

Zendesk (AI agents). Zendesk reports an "automated resolution" when the AI agent resolves an issue without human help, and its documentation says an LLM verifies each one. A conversation needs an identified use case and a response, and only conversations that weren't escalated are evaluated. On the platform in place before resolution tiers launched on May 18, 2026 (accounts stay on it until they move to tiers), the evaluation timing depends on the channel: email ends 72 hours after the first email, messaging ends 2 hours after the first message by default (configurable up to 72 hours), and voice ends at hangup (Zendesk). The notable design choice here is the verification step, plus a waiting period that varies by channel.

Ada. Ada treats a conversation as resolved only if it's relevant (the AI understood the question), accurate (the information was correct and current), safe (it stayed respectful and avoided harmful topics) and contained (no handoff). Idle web chat and social conversations end automatically after 24 hours, email after 72, and an escalated conversation is marked Not Resolved immediately after the handoff (Ada). The notable design choice is that accuracy is part of the definition, not an afterthought.

Put side by side, three platforms give three answers to the same questions.

None of these is wrong. They answer slightly different questions, which is exactly why you can't compare "62% resolved" from one tool against "48% resolved" from another. And if you run more than one AI layer, say a helpdesk bot plus a separate voice agent, you'll have two definitions living in one report.

Real Examples by Function

The same label hides different realities depending on where the AI sits.

Customer support, Tier 1. A customer asks for a password reset link and gets it. That's a clean resolution: the task was completed, and the customer has no reason to come back. By contrast, a customer asks about a refund policy, the bot quotes last year's version, and the customer closes the window. Under an assumed-resolution rule that's counted as solved. Under an accuracy criterion it isn't.

Billing and finance operations. A bot explains how invoices are issued but can't see the customer's account. The customer then emails, which opens a ticket. Containment on the chat looks perfect. The resolution rate for the underlying issue is zero, and the real cost was two contacts.

IT help desk. An employee asks the AI how to connect to the VPN. The AI returns the right article, and the employee connects. A week later the same employee asks again because the steps changed. Look at repeat contacts per person, not just per conversation.

E-commerce. "Where is my order?" is a good fit for AI because the answer is a lookup. If the AI says "in transit" and the package is three days late, the conversation is resolved on paper and a complaint on the phone the next day.

Onboarding and customer success. A new customer asks how to import contacts. The AI gives generic steps that don't match their plan. They never reply. Silence reads as success, but activation stalled.

The pattern across all five: the metric is easiest to fool where the customer can quietly walk away.

How to Measure AI Resolution Rate Honestly

Here's a method that works whichever platform you use. It doesn't need a data science team, only discipline about definitions.

1. Write your own definition first. One sentence: "A conversation is resolved when the AI gave a correct answer or completed the task, no human replied, and the customer didn't contact us again about the same issue within a set window." Pick the window to match your product. A password reset can use a short one. A billing question might need longer.
2. Separate the three numbers. Report containment, deflection and resolution as three lines. Never let one stand in for another.
3. Decide how silence is treated. Run the metric both ways: confirmed resolutions only, and confirmed plus silent. The gap between them is your uncertainty, and it's worth showing leadership.
4. Add a repeat-contact check. For each counted resolution, look for the same customer reaching you again on the same topic through chat, email or phone. A customer who comes back through another door was never resolved. This is the check that catches the quiet failures the examples above describe.
5. Sample and grade by hand. Pull a random set of conversations your AI counted as resolved, and have a reviewer mark each one right, partly right or wrong. Do it weekly at first. If the reviewer disagrees with the platform on a large share, the platform number is optimistic. Use the same sample to feed better source content back into the knowledge base.
6. Segment by topic and by risk. An overall rate hides that password resets resolve well and billing disputes don't. Break the rate down by contact reason, and keep refund, legal and safety topics out of any "automation target".
7. Pair it with experience measures. Compare CSAT for AI-handled and human-handled conversations separately, and track the escalation rate. Rising resolution with falling CSAT usually means customers are being worn down, not helped.
8. Fix the denominator. Decide whether you divide by all conversations, by conversations the AI attempted, or by conversations in scope. State which one in every report. Changing the denominator quarter to quarter is the easiest way to manufacture progress.
9. Review the definition when the product changes. New channels, new tools for the AI to take actions, or a vendor's rule change all shift what counts.

If you only have time for one extra step, make it the repeat-contact check. It's cheap, it works across tools, and it measures the thing customers care about.

Pitfalls to Avoid

Treating silence as satisfaction. Customers leave for many reasons, including giving up. The CFPB's report on bank chatbots describes customers trapped in "doom loops" of repetitive, unhelpful responses with no route to a human, and cites a survey where 78% of consumers who interacted with a chatbot needed to connect with a human after it failed to serve their needs (CFPB). Those customers may show up as contained.

Comparing rates across vendors. As the tables above show, platforms differ on wait times, who judges, and whether accuracy is checked. Compare trends inside one definition, not headline figures across definitions.

Setting a single target for everything. A blanket goal pushes teams to widen scope into topics where the AI shouldn't act. Decide what to automate and what to assist first, then set targets per topic.

Hiding the exit. Making it harder to reach a person raises containment and lowers trust at the same time. What customers expect from AI service covers why a clear route to a human matters.

Ignoring the knowledge base. Resolution rate is often a content problem. A wrong or outdated article produces a confident wrong answer, and your retrieval setup, such as RAG-based ticket deflection, is only as good as what it retrieves.

Reporting without a human check. An LLM judging an LLM can share the same blind spots. Keep the manual sample, even when the platform verifies automatically.

What a Healthy Report Looks Like

A one-page monthly view is enough. Show the three numbers (containment, deflection, resolution) with their definitions in a footnote. Show resolution split by topic. Show the confirmed-only and confirmed-plus-silent figures side by side. Add the repeat-contact rate, CSAT for AI versus human, and the share of sampled "resolved" conversations that a reviewer rejected. When the roles on your team start to shift as AI handles more volume, the people side needs the same level of reporting as the tooling.