Mike HeapMike is an experienced Product Manager who focuses on all the “non-development” areas of My AskAI, from finance and customer success to product design, copywriting, testing and more.
Table of Contents
- Autonomous resolution, in more depth
- How does autonomous resolution work in practice?
- What's a good autonomous resolution rate?
- Common misconceptions about autonomous resolution
- Misconception 1: deflection or containment is the same as autonomous resolution
- Misconception 2: the AI sent a reply, so the ticket is resolved
- Misconception 3: a higher autonomous-resolution rate is always better
- Misconception 4: "autonomous" means no human is ever involved in support
- What autonomous resolution is NOT, and how it relates to adjacent terms
- How does My AskAI handle autonomous resolution?
- FAQs
Created time
May 29, 2026 01:36 PM
Title length (<60)
Author
Ecomm?
Image
Publish date
May 31, 2026
Slug
what-is-autonomous-resolution
Featured
Type
Article
Ready to Publish
Ready to Publish
Autonomous resolution is a support request an AI agent handles from start to finish - no human touches it - and the customer's underlying issue is actually solved.
The word doing the work in that sentence is "autonomous." It means zero human involvement on that ticket: nobody drafted the reply, nobody approved it, nobody picked it up after the AI had a go.
And "resolution" means the issue genuinely went away for the customer. A reply going out and a ticket quietly auto-closing doesn't clear that bar.
The term gets stretched all the time. A lot of what vendors badge as "autonomous resolution" is really deflection (the customer never reached a human) or containment (the ticket closed in-channel, even if the person gave up). If you're on this page, I'd bet you're trying to work out whether a vendor's "85% autonomous resolution" claim is real, or whether you're being sold a deflection number wearing a resolution label.
We run AI customer service inside 200+ ecommerce and SaaS businesses, and I've watched this number get reported honestly and dishonestly. The difference comes down to one design choice, and I'll get to it.
Autonomous resolution, in more depth
TL;DR: Autonomous resolution has three parts that all have to be true: the AI handled the ticket alone, no human took over, and the customer's issue actually got solved.
A ticket only counts as an autonomous resolution if three things are all true. Miss any one of them and you're measuring something else.
First, it has to be AI-only. No human drafted, edited, or approved the reply (not even a quick tweak before it sent). If an agent worked from an AI-suggested draft and sent it, that's AI-assisted, which is a useful thing but a different thing.
Second, there has to be no handoff. The conversation never escalated to a person.
This is where autonomous resolution differs from a handed-off ticket: the autonomous number is the set of conversations that never hit an escalation trigger (we obsess over exactly where that trigger sits). As Cobbai's write-up on autonomous resolution concedes, "smooth escalation protocols are vital," so the honest autonomous count is whatever's left after those protocols fire.
Third, the issue has to be actually resolved. The customer's problem got solved, and ideally they confirmed it (or at least didn't reopen the ticket). A reply that goes out and a ticket that auto-closes 72 hours later still leaves the problem sitting there.
There's one point underneath all three that most definitions skip, and in my experience it's the one that matters most. The honesty of an autonomous-resolution number depends entirely on how easy it is to reach a human.
When escalation is frictionless, the number takes care of itself: anything that genuinely got resolved counts, and anything that didn't goes to a person. It only becomes inaccurate when a product is designed to make the human hard to reach, so customers who weren't actually helped get logged as "resolved" because they had nowhere else to go.
The edges get fuzzy in a few predictable places (this is where the arguing happens). Confirmed resolution versus assumed resolution: do you trust a reopen, a CSAT score, or silence?
Then there are action-requiring tickets, like a refund actually processed, versus answer-only ones. And the gray zone where an AI draft went out unedited but a human was technically "in the loop."
How does autonomous resolution work in practice?
TL;DR: A ticket arrives, the AI understands it, grounds an answer or takes an action, then closes it as resolved. If confidence is low it hands off instead, which doesn't count as autonomous.
Mechanically, an autonomous resolution is a short pipeline that runs without anyone watching. Here's the shape of it for a typical support ticket (this is roughly how ours runs).
- Trigger. A ticket or chat arrives, in the helpdesk inbox or a chat widget.
- Intent understanding. The AI works out what the customer actually means, beyond the literal words they used.
- Grounding. It pulls the answer from your connected knowledge (help center, docs, past answers) rather than inventing one.
- Action. For anything beyond a straight answer, it calls a tool or runs a multi-step task: looking up an order, processing a refund, updating an account over an API.
- Answer and confirmation. It replies, and ideally checks the issue is solved (or sees no reopen and a positive rating).
- Close, or hand off. If it's confident and the issue is solved, the ticket closes as resolved. If confidence is low or the customer asks for a person, it hands off, and that ticket does not count as an autonomous resolution.
Step 6 is the interesting one, and it's where I see the most confusion: where vendors draw that line is where the numbers split apart. Two platforms can both report "70% autonomous resolution" while counting completely different events.
Each major platform counts something slightly different.
Platform | What it counts as a resolution |
My AskAI | The conversation wasn't escalated to a human. A deliberately basic signal, but a defensible one: escalation is made easy (an explicit ask, the AI being unable to answer, signs of frustration, or a set escalation topic), and we don't pretend to know more than that. |
"Fin resolves the issue end to end, or successfully executes a Procedure you've configured to end in a handoff to a human or a workflow." A configured handoff can count. Fin renamed "resolutions" to "outcomes" in late 2025. | |
"A resolved support conversation, a saved cancellation, an upsell, a cross-sell" — outcomes the customer agrees to. | |
Decagon | An internal "AI resolved" tag; the criteria aren't public. |
A 72-hour quiet period; it counts only if the ticket stays closed for 72 hours. | |
It counts when the customer didn't reply. |
Notice the spread (we've made peace with the fact that no two vendors count this the same way). Gorgias counting a non-reply as a resolution is the weakest signal on the list, because a customer who gave up looks identical to a customer who was helped.
Zendesk's 72-hour silence is a slightly stronger version of the same idea. Microsoft goes further still: their Dynamics 365 docs describe autonomous case resolution as the system that "drafts and sends an email," so sending a reply gets treated as resolving the case. None of these definitions is dishonest; they just measure different events and call them the same word.
What's a good autonomous resolution rate?
TL;DR: Roughly 40-65% is average, 65-80% is solid, and 80%+ is world-class. Always read it next to CSAT, because a rate on its own tells you very little.
The honest answer is a range, and it depends heavily on how built-out your setup is. From the rollouts we've run, the bands look like this.
Tier | Rate | What it usually means |
World-class | 80%+ | Mature deployment, APIs connected, a weekly review loop running. |
Solid | 65–80% | Custom answers in place, guidance tuned, tools connected. |
Average | 40–65% | Knowledge base built out, common questions covered. |
Early | 25–40% | Early rollout, knowledge base still thin. |
Those bands aren't theoretical. RecruitCRM started around 35% and climbed to 68%; TravelJoy went from 24% on Zendesk's own AI to 80%; Edel Optics moved from roughly 25% to 79% after connecting live customer data.
The climb almost always comes from the customer's own work: better knowledge, connected APIs, tuned guidance. The model itself stays much the same throughout.
The bands shift by industry, too. A transactional ecommerce queue (where-is-my-order, returns, sizing) can run high, while a regulated or high-judgment one (fintech, healthcare, complex technical support) caps lower, because more of those tickets should go to a person.
A resolution rate on its own, though, is close to meaningless, and I'd never report one without CSAT sitting next to it. A bot can "resolve" a ticket by frustrating someone into giving up, so the number alone tells you very little.
TravelJoy's 80% came with 86% AI CSAT; Zinc runs 68% resolution at 97% CSAT. The reverse is just as telling, and it's my favorite example: Sofar Sounds deliberately runs a triage-first setup at 26% AI resolution but 85% CSAT, routing most tickets to humans on purpose. For their team, a low rate alongside high satisfaction is exactly the right call.
Common misconceptions about autonomous resolution
TL;DR: Deflection and containment measure something weaker than resolution, a sent reply can hide an unsolved problem, and a higher rate only helps when CSAT holds up.
This is the term that gets stretched the most, so it's worth naming the stretches directly (I see all four of these on demo calls).
Misconception 1: deflection or containment is the same as autonomous resolution
Deflection only tells you a human wasn't reached. Containment only tells you the ticket closed in-channel. Neither tells you the customer's problem got solved.
The trap is subtler than a low number. It's a high number that's been manufactured by making the human hard to reach.
Bury the "talk to a person" option, and customers who weren't helped get logged as resolved because they had no other option (I've watched vendors do this deliberately). Even Intercom admits the risk, noting that a high deflection rate can "mask customer dissatisfaction and eroded trust."
Misconception 2: the AI sent a reply, so the ticket is resolved
A sent reply and an auto-close aren't evidence the problem went away (reopen rate and CSAT are the real tell). This is exactly why a definition like "the system drafts and sends an email" sets the bar too low: it counts the action and stops there.
Misconception 3: a higher autonomous-resolution rate is always better
I'd flag this one the hardest. Not without CSAT, and not for every ticket type: for high-judgment or regulated queries, a fast handoff to a human is the correct outcome, so pushing the rate up in those categories does more harm than good.
Misconception 4: "autonomous" means no human is ever involved in support
It's a per-ticket measure of what the AI handled alone. A healthy setup autonomously resolves the repetitive tail and routes the rest to humans with full context, and the goal was never 100%.
What autonomous resolution is NOT, and how it relates to adjacent terms
TL;DR: Autonomous resolution is the event; resolution rate is the metric; deflection and containment are weaker cousins; handoff is its opposite; agentic workflows are the mechanism.
Autonomous resolution sits in a cluster of metrics that get used interchangeably and shouldn't be. The quickest way to keep them straight: autonomous resolution is an event (one ticket, solved by AI, no human).
Term | What it is | How it differs from autonomous resolution |
AI resolution rate | The percentage of issues an AI fully resolves without a human | The metric that counts how often autonomous resolution happens; the resolution itself is the event. |
Deflection rate | The share of contacts prevented from reaching a human | Measures absence-of-human, not problem-solved. |
Containment rate | The share of tickets closed in-channel without a human | Closed is not the same as solved; it counts give-ups. |
Automation rate | The share of work the AI touches | Touching a ticket is not resolving it end-to-end. |
The structured transfer of a conversation to a human | The opposite event: a handed-off ticket is, by definition, not autonomously resolved. | |
Agentic workflows | Multi-step procedures an AI runs end-to-end | The mechanism by which action-requiring tickets get autonomously resolved. |
If you only take one thing from this section, make it this: a vendor reporting "resolution" and a vendor reporting "containment" can show you the same percentage and mean very different things about whether your customers got help.
How does My AskAI handle autonomous resolution?
TL;DR: My AskAI resolves the repetitive tail autonomously and hands off the rest with full context. Because we bill per ticket, a rising resolution rate never inflates your bill.
We count a resolution in a deliberately basic way: did the conversation get escalated to a human? If it didn't, the AI handled it, and that counts. It isn't the most sophisticated numerator on the list, and that's fine.
It's defensible because we make escalation genuinely easy. A customer can ask for a person in any number of ways, and the AI hands off when it can't answer, when it picks up on frustration, or when a ticket hits a topic you've set for escalation. So a conversation that was never escalated is one the AI genuinely handled, because anyone who needed a person could reach one.
We're also honest about the ceiling. You can't really know you solved someone's problem unless they tell you, so we don't invent ways to pretend we know. Tasks and Tools let the AI actually do things (process a refund, look up an order, update an account), so action-requiring tickets get handled rather than just answered.
The pricing model is what lets us keep the definition honest. We charge per ticket rather than per resolution, so we've no reason to inflate the count, and your bill stays flat as the AI gets better (the opposite of how per-resolution vendors work). RecruitCRM is the proof point: a disciplined weekly review loop took them from about 35% to 68% AI resolution across their Intercom tickets, at 75% AI CSAT, saving 62 hours a month.
100% was never the target. Handing off the right tickets is part of a good setup, and we don't offer voice today.
FAQs
What is autonomous resolution in customer service?
It's a support request an AI agent handles from start to finish, with no human involvement, where the customer's issue actually gets solved. All three conditions matter: AI-only, no handoff, and genuinely resolved rather than just replied-to.
What's the difference between autonomous resolution and deflection?
Deflection measures whether a contact was prevented from reaching a human. Autonomous resolution measures whether the AI actually solved the issue on its own (this is the distinction I find buyers miss most). You can deflect a ticket without resolving anything, because the customer just didn't get through to a person.
What's the difference between autonomous resolution and containment rate?
Containment counts tickets that closed in-channel without a human, even if the customer gave up unsatisfied. Resolution counts only the ones where the problem got solved. A closed ticket and a solved problem are two different things.
Is autonomous resolution the same as AI resolution rate?
They're tightly linked but not identical. Autonomous resolution is the event (one ticket solved by AI with no human), and AI resolution rate is the metric that counts how often that event happens across your volume.
How is autonomous resolution measured?
We measure it at the conversation level: did the AI handle it alone, did it avoid a handoff, and did the issue actually get resolved? The catch is that vendors define "resolved" differently. Some require a confirmation or no-reopen, others count a 72-hour silence or even a non-reply, so two platforms' numbers aren't directly comparable.
What's a good autonomous resolution rate?
Roughly 40-65% is average, 65-80% is solid, and 80%+ is world-class, but only read alongside CSAT. A high rate with poor satisfaction usually means customers are being shut out rather than helped.
Can an AI resolve a ticket without any human involvement?
Yes, for the repetitive tail: order status, password resets, policy questions, and action-requiring tasks where the AI can call the right tool. The rest should be handed off to a human, which is exactly what should happen.
Does a higher autonomous resolution rate mean better support?
Not on its own, and I'd push back on anyone who treats the headline number as the whole story. A rate without CSAT can hide customers who gave up. Sofar Sounds runs 26% resolution at 85% CSAT on purpose, routing most tickets to humans, and that's a healthy pattern for their team.
What types of tickets can be resolved autonomously?
Answer-only tickets grounded in your knowledge base, and action-requiring tickets where the AI can run a task or call an API. High-judgment, sensitive, or regulated queries usually shouldn't be; those should escalate.
How is autonomous resolution different from agentic workflows?
An agentic workflow is the mechanism, the multi-step procedure the AI runs to get something done. Autonomous resolution is the outcome of that mechanism running successfully on a ticket with no human involved.
Why do vendor autonomous-resolution claims vary so much?
Because each vendor defines the numerator differently (fun fact: that's why two "70%" claims rarely mean the same thing). Intercom Fin can count a configured handoff as an outcome, Zendesk uses a 72-hour quiet period, and Gorgias counts a non-reply, so identical headline percentages can mean very different things.
Does autonomous resolution mean replacing my support team?
No. It's a per-ticket measure of what the AI handles alone, and it was never a plan to remove humans. In practice it frees your team from the repetitive tail so they can spend time on the conversations that genuinely need a person.
How do I improve my autonomous resolution rate?
In our experience, most of the gains come from your own work: filling the questions your AI couldn't answer, connecting APIs and tools so it can take actions, writing custom answers for your top questions, tuning guidance, and reviewing the misses each week. That's how RecruitCRM nearly doubled their rate from 35% to 68%.
How does My AskAI define an autonomous resolution?
We use a deliberately basic definition: a conversation counts as resolved if the AI handled it without escalating to a human. It's defensible because we make escalation easy (an explicit ask, the AI being unable to answer, signs of frustration, or a set escalation topic), so a conversation that stayed with the AI is one it genuinely handled. We don't pretend to know an issue was solved without the customer confirming it, and because we bill per ticket rather than per resolution, we've no reason to inflate the count.
