Vendors will quote you a resolution rate. The honest answer is that there is no single number, because resolution rate depends on what you count, who you serve, and whether the AI is allowed to take actions or only answer questions.
Resolution rate is the share of customer support tickets that an AI agent closes end-to-end without a human stepping in. In 2026, published and realistic figures range widely: simple FAQ deflection can look like 80-90% on paper, while genuine end-to-end resolution of complex, regulated tickets sits closer to 30-60% for most teams and climbs into the 70-85% range only where the AI is deeply integrated, allowed to act, and tightly guardrailed. Treat any one number with suspicion until you know how it was measured.
Resolution rate is not deflection rate. Deflection counts conversations a human did not touch; resolution counts problems actually solved. The two are routinely conflated in marketing.
Realistic 2026 ranges, as an industry observation: 30-50% for early deployments, 50-70% as workflows mature, and 70-85% for deeply integrated, action-taking agents on well-scoped use cases.
Gartner predicts agentic AI will autonomously resolve 80% of common customer service issues by 2029, up from low double-digits in 2024.
The biggest driver of higher rates is action-taking: an agent that can look up an account, run a check, and update a system resolves far more than one that only retrieves answers.
Rates vary enormously by use case. A billing FAQ and a KYC unlock are not the same problem, and a single blended number hides that.
Last updated: June 2026
If you are evaluating AI customer support, resolution rate is the first number every vendor puts in front of you and the easiest one to misread. A platform can advertise 90% by counting any conversation a human did not join, including the ones where the customer gave up. Another can report 45% because it counts only tickets where the underlying problem was genuinely solved and verified. The second number is lower and far more useful. This guide explains what resolution rate actually measures, what realistic 2026 ranges look like, what separates a 40% deployment from an 80% one, and how a few representative platforms talk about their own resolution posture. We avoid inventing precise figures: where the industry has not published a defensible number, we describe ranges and say so.
What Does Resolution Rate Actually Mean?
Resolution rate is the percentage of inbound support tickets an AI agent resolves completely, without escalation to a human, over a given period. The definition sounds simple. The disagreement is entirely in the word "resolves."
There are at least three numbers that get called resolution rate, and they are not interchangeable:
Deflection rate: the share of conversations a human agent never touched. This is the most generous and least meaningful metric. A customer who asks a question, gets an unhelpful answer, and abandons the chat counts as deflected. So does a customer who was helped perfectly. Deflection tells you about human workload, not customer outcomes.
Containment rate: the share of conversations that stayed inside the AI channel and were not transferred. Better than deflection, because it excludes obvious handoffs, but it still counts a customer who quietly gave up as contained.
Resolution rate (true): the share of tickets where the customer's actual problem was solved end-to-end. This is the number that matters, and the hardest to measure honestly, because it requires defining resolution and verifying it after the fact rather than assuming it.
The gap between these three can be 30 percentage points or more on the same set of tickets. When a vendor quotes a single high number, the first question is which of the three they mean and how they verify it.
Who decides what counts as resolved?
In the strongest measurement setups, the customer (the business buying the AI) defines what counts as a resolution, and the platform verifies it rather than self-reporting. Some platforms run automated post-ticket quality assurance on every conversation to check whether the stated outcome actually happened, instead of trusting that a closed ticket equals a solved problem. Lorikeet, for example, lets the customer hold a veto on what counts as a resolution and runs 100% automated QA through its Coach agent, so the resolution number is the verified one rather than the optimistic one. That distinction matters more than the headline figure.
Realistic Resolution Rate Ranges in 2026
The honest framing is a range, not a point. Based on how platforms and analysts describe results across the category in 2026, the following bands are a reasonable industry observation rather than a precise measured figure. Your own rate will land somewhere inside them depending on use case, integration depth, and how strictly you define resolution.
Early deployments: roughly 30-50%
A newly launched AI agent, working from a knowledge base with limited integrations, typically resolves a minority of tickets on its own. It handles the repetitive, answerable questions and escalates anything that requires looking something up or taking an action. This is not a failure; it is the expected starting point. The teams that succeed treat this number as a baseline to grow from, not a verdict.
Mature workflows: roughly 50-70%
As you connect the agent to your core systems, give it the ability to take actions, and tune its workflows against real ticket data, the rate climbs. Most well-run deployments that have been live for several months and invested in integration land in this band. The ceiling here is usually set by how many ticket types still require an action the agent cannot yet take.
Deeply integrated, action-taking agents: roughly 70-85%
The highest defensible rates come from agents that are integrated into the systems where work actually happens, allowed to take real actions (issue a refund, update an account, file a dispute), and constrained by guardrails strong enough that the business trusts them to act unsupervised. Lorikeet has described regulated customers reaching automation in the mid-80s percent while holding customer satisfaction equal or better than their human baseline, which is at the top end of what is realistic for complex workflows. Numbers above that, especially on regulated tickets, deserve scrutiny about how resolution was defined.
A caution on the 90%+ claims
When you see resolution or accuracy claims in the 90s, check three things: whether the figure is deflection rather than true resolution, whether it covers a narrow ticket type rather than the full inbound mix, and whether it is self-published on an internal benchmark rather than measured in production against a customer-defined standard. A 95% number on a billing FAQ tells you little about how the agent handles a failed transfer or a fraud claim.
Why resolution rate should never travel alone
A resolution rate is only meaningful when you pair it with at least two other numbers: the customer satisfaction (CSAT) of the resolved tickets, and the escalation quality on the ones the agent did not resolve. An agent that resolves 80% of tickets but tanks CSAT is closing conversations, not solving problems. An agent that escalates cleanly, with full context handed to the human, is worth more than one with a marginally higher rate that dumps confused customers into a queue. The strongest deployments report resolution alongside CSAT held equal or better than the human baseline, which is a much harder and more honest bar than resolution alone. If a vendor will only show you the resolution number and not the satisfaction of the tickets behind it, assume the two numbers move in opposite directions.
What Drives a Higher Resolution Rate
Resolution rate is not a property of the model. It is a property of how the agent is built, integrated, and governed. Four things move it more than anything else.
Deep integrations
An agent can only resolve what it can reach. If it can read your knowledge base but cannot query the payment system, the CRM, or the core banking platform, it is limited to answering questions and will plateau early. Native, write-capable integrations (the agent can update a record, rather than only read it) are the single biggest lever. The difference between "we integrate with Stripe" meaning read-only invoice lookups and meaning issue-refunds-with-idempotency-keys is the difference between a 45% and a 70% deployment on payment-related tickets.
Action-taking, beyond answering
Most real support tickets are not questions, they are requests: change my address, why did my transfer fail, lock my card, dispute this charge. An agent that can only retrieve and reply will escalate all of these. An agent that can chain several tool calls in order (verify identity, check status, take the action, confirm) resolves them. This is the line between a chatbot and an agent, and it is where most of the resolution-rate gap between platforms comes from.
Guardrails that let you trust unsupervised action
Counterintuitively, strong guardrails raise resolution rate rather than lowering it. A business will only let an AI act unsupervised on a ticket type if it trusts the agent not to do something harmful. That trust comes from pre-launch adversarial testing, runtime checks on inbound and outbound messages, and post-ticket QA. Lorikeet describes this as defense in depth: simulation and red-teaming before launch, message checks during, and 100% automated QA after. The more provable the behavior, the more ticket types a compliance or risk team will sign off on for full automation, and the higher the resolution rate climbs.
Scoping and continuous tuning
Resolution rate rises fastest when teams analyze where the agent escalates, identify the ticket types responsible, and either add the missing integration or refine the workflow. Treating the agent as a system that improves with data, rather than a fixed product, is what moves a deployment from the 50% band into the 70% band over a few months.
Channel coverage and shared memory
Resolution rate is usually quoted per channel, but customers do not stay in one channel. A dispute that starts on chat continues by email and finishes on a phone call. If the agent runs voice on a separate stack from chat and bolts them together with a transcript handoff, the customer repeats themselves and the conversation effectively restarts, which depresses resolution on anything that crosses channels. An agent that handles chat, email, voice, and SMS on one engine with shared memory resolves multi-channel issues as a single ticket rather than three disconnected ones. For businesses with meaningful phone volume, voice that can take real actions on the call (rather than read out an answer and route to a human) is often the difference between a respectable blended rate and a poor one.
Representative Platforms and Their Resolution Posture
The table below summarizes how a few well-known platforms tend to talk about resolution, and the realistic read behind the marketing. These are characterizations of public positioning, not audited benchmarks, and actual rates depend entirely on your use case.
Lorikeet · Positioning: end-to-end resolution for complex, regulated industries (fintech, healthtech, gaming). Realistic posture: emphasizes verified resolution with customer-held veto on what counts, 100% automated QA, and regulated customers reaching mid-80s percent automation at equal-or-better CSAT on well-scoped workflows. Differentiator: defense-in-depth guardrails and action-taking across chat, email, voice, and SMS on one engine.
Fin by Intercom · Positioning: outcome-based, pay-per-resolution AI agent layered on a helpdesk. Realistic posture: strong on high-volume, well-documented support; resolution rate varies by knowledge-base quality and integration depth. Differentiator: low published per-outcome price and fast time to launch.
Decagon · Positioning: enterprise AI agent with white-glove deployment. Realistic posture: high rates achievable on enterprise volume with embedded engineering during launch; figures are negotiated and not publicly benchmarked. Differentiator: per-conversation or per-resolution pricing and heavy implementation support.
Sierra · Positioning: outcome-only billing, you pay when the AI fully resolves. Realistic posture: the pricing model itself signals confidence in resolution, but outcome-only incentives can bias a vendor toward easier tickets. Differentiator: pure outcome pricing and enterprise procurement strength.
Zendesk AI · Positioning: AI agent and automation layered on the Zendesk Suite. Realistic posture: solid on simpler, well-contained ticket types inside an existing Zendesk footprint; resolution depth limited by a ticketing-first architecture. Differentiator: native to the Suite and a per-resolution add-on.
The pattern across all five: the published number is always the optimistic one. The useful number is the one measured in your stack, on your ticket mix, against your definition of resolved.
If you want to see what verified end-to-end resolution looks like on your own hardest tickets, book a Lorikeet demo and bring the ticket types you assume can't be automated.
How to Lift Your Own Resolution Rate
If your rate is lower than you want, the fix is usually mechanical rather than mysterious. Work through these in order.
1. Measure the right number first
Before chasing a higher rate, make sure you are measuring true resolution and not deflection or containment. Define what "resolved" means for each major ticket type, then verify it after the fact rather than assuming a closed ticket was solved. If you cannot trust the number, you cannot improve it.
2. Find where the agent escalates and why
Group escalations by ticket type. You will usually find that a handful of categories account for most of the misses, and most of those misses are because the agent could not take an action or reach a system. That list is your roadmap.
3. Close the integration gaps
For each high-volume escalation category, ask whether the agent is missing read access, write access, or a workflow. Adding a write-capable integration to the system behind your top escalation category is typically the highest-impact change you can make.
4. Strengthen guardrails so you can automate more
For ticket types your risk or compliance team is nervous about, the unlock is provable behavior. Pre-launch simulation, runtime checks, and post-ticket QA give stakeholders the evidence to approve fuller automation. Every ticket type you move from "escalate to be safe" to "automate with guardrails" raises your rate.
5. Tie the rate to cost, not only volume
A higher resolution rate is only valuable if it lowers your cost to serve. The math is straightforward: a human-handled ticket costs roughly $1.25 to $4 depending on complexity and region, while a resolution handled by an AI agent can cost on the order of a dollar or less, depending on channel and pricing model. Watch the pricing structure as closely as the rate. Some vendors charge per deflected conversation, which rewards them for the vanity number rather than the solved problem. Outcome-based or per-resolution pricing aligns better, and pricing where escalations are not charged and the customer defines what counts as a resolution aligns best of all, because it stops you paying for tickets the AI did not actually solve. When you model ROI, multiply your real resolution rate (not the deflection rate) by your blended human ticket cost to get the defensible savings figure.
6. Treat it as continuous, not one-time
Resolution rate is a curve, not a launch metric. The teams in the 70-85% band got there by reviewing escalations monthly, adding integrations, and tuning workflows, not by buying a better model. Expect to start lower and climb.
One honest caveat to end on: there is a point past which chasing a higher rate is counterproductive. Forcing the agent to attempt tickets it should escalate trades a better headline number for worse outcomes on the cases that matter most. In regulated industries especially, a slightly lower resolution rate with verified correctness beats a higher one with hidden failures. The goal is the highest rate you can reach without compromising the answers you cannot afford to get wrong.
