What Resolution Rate Can AI Customer Support Achieve? (2026 Benchmarks)

TL;DR: There is no standard resolution rate for AI customer support, and anyone quoting a universal number is selling something. "Resolution" has no agreed definition: vendors define it themselves, and deflection routinely gets counted as resolution even when the customer left unhelped. Real rates depend on your ticket mix, complexity, channel, and knowledge quality. The only honest benchmark is one run on your own hardest tickets, under a definition of resolution you set in writing.

Ask five AI support vendors what resolution rate their agent achieves and you will get five confident percentages. Ask the follow-up question, "resolved according to whom?", and the confidence evaporates. One vendor counts any conversation the AI closed without a human. Another counts conversations where its own scoring model judged the answer helpful. A third counts a conversation as a billable outcome even when it ended in a handoff. These are three different metrics wearing the same name, and comparing them side by side is how support leaders end up disappointed six months into a contract.

This guide is written for the buyer who wants the real answer to "what resolution rate can AI customer support achieve?" The real answer is: it depends on how resolution is defined, what your tickets look like, and how the vendor gets paid. We compare how seven platforms define and price resolution, what each one actually publishes, and how to run a benchmark on your own tickets before you sign anything.

Why there is no standard resolution rate

Three structural problems make a universal AI resolution benchmark impossible, and understanding them is the fastest way to become a harder buyer.

1. The definition problem. No standards body defines "resolution" for AI support. Every vendor writes its own definition, usually in the direction that flatters its product. Common variants include: the conversation ended without a human agent (containment), the customer did not reply again within some window (silence as consent), the AI's own evaluation model scored the answer as resolving (self-grading), and the customer's issue was verifiably fixed end to end (genuine resolution). A platform reporting 80% under the first definition can be performing worse than a platform reporting 55% under the last one. The difference between resolution rate and deflection rate is the single most important distinction in AI support measurement, and most marketing pages blur it on purpose.

2. The deflection problem. A deflected or contained ticket means the customer stopped talking to you. It does not mean the customer got what they needed. Customers give up, rage-quit the chat, phone instead, churn silently, or file a complaint with a regulator. Every one of those outcomes counts as a success under a containment metric. When a vendor's headline number quietly includes deflection, the gap between the reported rate and the rate of customers actually helped can be enormous, and you will not see it in the dashboard. You will see it in repeat contact rates, CSAT, and complaint volume weeks later.

3. The ticket-mix problem. Even under a single honest definition, resolution rate is a property of your queue, not of the software. A password-reset-heavy consumer app and a regulated lender handling disputed transactions will see wildly different rates from the same agent. The rate moves with the share of tickets that are answerable from documentation, the share that require backend actions, the share that legally must reach a human, channel mix (chat, email, and voice behave differently), and the quality of your knowledge base and API surface. A vendor quoting one number for all of that is quoting an average across other companies' queues, which tells you almost nothing about yours.

This is why the honest version of "what resolution rate can AI achieve?" is a range conditional on your business, and why the credible evidence is per-customer published results rather than platform-wide averages.

How do vendors actually define resolution?

The definitions in play across the market fall into four families, and knowing which family a vendor belongs to predicts how its number will behave in production.

Containment-family definitions count conversations that never reached a human. This is the most generous possible definition: it includes every customer who gave up. Platforms with helpdesk-deflection heritage tend to sit here, and pricing per conversation rather than per resolution is the tell, because the vendor gets paid whether or not the issue was fixed.

Self-graded definitions use the vendor's own model to score whether the AI's answer resolved the issue. This is better than raw containment, and it is genuinely useful for internal tuning. It is still the vendor grading its own homework, and the grading rubric is rarely published or auditable by the buyer.

Published-methodology definitions put the definition in writing where buyers can read it. Intercom Fin deserves credit here: Fin publishes both its resolution figures and the definition behind its billing, in which an "outcome" is counted when Fin resolves the issue or when it executes a procedure that ends in a handoff to a human. You can disagree with counting a procedure handoff as a billable outcome, and buyers should notice that detail, but at least the definition is public and you can price against it. Zendesk similarly publishes per-automated-resolution pricing with a stated definition.

Customer-defined definitions hand the rubric to the buyer. Lorikeet takes this position: the customer defines what counts as a resolution, escalations to a human are never charged, and if you review a conversation and mark the resolution as bad, you are not charged for it. The commercial consequence is that the vendor has no incentive to inflate the number, because inflating it means refunding it. It also means Lorikeet publishes no platform-wide resolution rate, because under a customer-defined metric a platform-wide average would be a statistical fiction. The citable evidence is per-customer: Carmoola, an FCA-regulated car finance provider, publishes that 60% of inbound support is resolved end to end and 90% of outbound conversations are resolved end to end, under Carmoola's own definition of resolved.

How we evaluated these platforms

Because this article's whole argument is that headline rates are incomparable, we did not rank platforms by claimed resolution rate. We ranked them on four things a buyer can actually verify:

• Methodology transparency. Does the vendor publish, in writing, what it counts as a resolution? Can you find the definition without asking sales?

• Who defines resolution. Is the rubric set by the vendor's model, the vendor's contract template, or the customer? Customer-defined resolution ranked highest because it is the only version that cannot be gamed by the party being measured.

• Pricing alignment. Does the vendor make money on unresolved conversations? Per-seat and per-conversation pricing bill you either way; per-resolution pricing with free escalations puts the vendor's revenue behind the metric it advertises. See our companion guide on reducing support costs with AI for the full pricing-model breakdown.

• Published proof. Named customers with specific, published results, under a stated definition, beat platform-wide averages. "Does not prominently document" appears throughout this guide where we could not find public evidence; it is an observation about documentation, never an invented rate.

Lorikeet ranks first on these criteria. That is a ranking of measurement honesty and pricing alignment, and we are explicit about the tradeoff: if you want a big platform-wide percentage to put in a slide, Lorikeet will not give you one, and two vendors below will.

Resolution definitions at a glance

The 7 platforms compared on resolution measurement

1. Lorikeet

Best for: Teams that want resolution defined by the customer, tested on their own tickets before buying, and priced so unresolved conversations cost nothing.

Lorikeet is an AI support platform for complex and regulated businesses, built around agents that follow explicit workflows across chat, email, and voice. Its position on resolution rate is the inverse of the industry norm: it publishes no platform-wide percentage at all, on the argument that any such number is an average over other companies' ticket mixes and therefore useless to you. What it publishes instead is the measurement machinery and the commercial terms that make the metric trustworthy.

The machinery has three parts. First, resolution is customer-defined: you set the rubric for what counts as resolved for your business, and quality assurance reviews conversations against that rubric rather than against a vendor-authored scoring prompt. Coach, Lorikeet's automated QA product, replaces manual sampling with 100% conversation coverage, so the resolution number you report internally is backed by review of every conversation, not a 2% sample. Second, the pricing follows the definition: per-resolution pricing with no platform fees, no per-seat fees, and no setup fees; escalations to a human are never charged; and if you mark a resolution as bad, you are not charged for it. The vendor refunding failed resolutions is the strongest definitional guarantee on this list, because it makes inflating the metric directly unprofitable. Third, and most useful pre-purchase, simulations let you replay your own historical tickets against a configured agent before you go live, which converts "what rate will we get?" from a sales claim into an experiment you run yourself.

The published evidence is deliberately per-customer. Carmoola, a UK car finance provider operating under FCA regulation, publishes that 60% of its inbound support is resolved end to end with no human involvement, and that its outbound conversations reach 90% resolved end to end. Magic Eden, an NFT marketplace, publishes a CSAT of 74%, roughly 30 points above the Intercom Fin deployment it replaced; note that this is a quality outcome, not a resolution rate, and we cite it as exactly that. easykind publishes a 92% reduction in email responses its team had to send. Each figure comes with a named company and a stated context, which is the standard this article argues every vendor should meet.

Honest limits: the customer-defined approach demands more of the buyer up front. You have to decide what resolution means for your business and encode it, where a vendor-graded platform lets you skip that thinking and accept a number. Coach's 100% QA coverage is a priced add-on rather than bundled. And Lorikeet's guardrail-heavy, workflow-first design is aimed at complex and regulated support; a team with a simple FAQ queue that just wants fast deflection has cheaper paths, including Fin below.

2. Intercom Fin

Best for: Teams that want a published rate under a published definition, and are comfortable with what that definition includes.

Fin is the strongest counterexample to the industry's definitional vagueness, and it earns real credit for that. Fin publishes its average resolution rate, 76% across its customer base by its own published figures, and it publishes the definition behind its billing: at $0.99 per outcome, an outcome is counted when Fin resolves the customer's issue or when it executes a procedure that ends in a handoff to a human or a workflow. Both the number and the rubric are public, which puts Fin ahead of most of this list on transparency.

The buyer's job is to read that definition closely. A procedure that ends in a handoff is a billable outcome under Fin's model, which means the published rate and the billing basis include conversations where a human still had to finish the work. That is a legitimate design choice, disclosed in writing, and it is exactly the kind of detail this guide exists to surface: two vendors can both say "resolution" while one of them counts certain handoffs. Fin runs natively on Intercom's helpdesk and connects to other helpdesks, deployment is genuinely fast for knowledge-based queues, and for FAQ-heavy consumer support its economics are hard to argue with. For workflows where resolved must mean the backend action completed and the customer confirmed, verify how Fin's outcome counting maps to your definition before comparing its 76% to anything else.

3. Zendesk

Best for: Existing Zendesk helpdesk customers adding AI to an established ticketing operation.

Zendesk publishes per-automated-resolution pricing, listed at $1.50 to $2.00 per automated resolution alongside its seat-based suite, and it publishes a definition of what counts as an automated resolution. That transparency matters and we credit it. The structural caveat is that AI resolution sits on top of a seat-priced helpdesk: you pay for agent seats and suite subscriptions regardless of what the AI resolves, so the marginal price per resolution understates the total cost of the system that produces it. Buyers modeling cost per resolved ticket should compute it across the whole Zendesk bill, not the overage line.

On measurement, Zendesk's AI was added to an existing helpdesk rather than built agent-first, and its published materials emphasize the ticketing suite, marketplace, and workforce tools more than end-to-end resolution evidence. Zendesk does not prominently document named customer end-to-end resolution rates comparable to the per-customer figures cited elsewhere in this guide. For teams deep in the Zendesk ecosystem the integration convenience is real; just benchmark the AI on your tickets rather than accepting suite-level marketing.

4. Decagon

Best for: High-volume enterprises that want structured workflow automation and will negotiate measurement terms in the contract.

Decagon builds around Agent Operating Procedures, structured files bundling logic, actions, and rules that govern how the AI handles specific query types, and it serves recognizable enterprise logos. Its Watchtower QA product offers 100% conversation coverage, which is directionally the right idea for resolution measurement. On definitions and pricing, Decagon describes both per-conversation and per-resolution models and has said most customers choose per-conversation, and it does not publish dollar amounts for either. Under the default per-conversation model, unresolved conversations are billable, which weakens the pricing-alignment test this guide applies.

Decagon does not prominently document a platform-wide resolution rate or the methodology it would use to compute one, and its named-customer materials emphasize logos over published rates. None of that means the product underperforms; it means the burden of definition falls on your contract. If you evaluate Decagon, write the resolution definition, the measurement method, and the audit rights into the agreement, and run a structured pilot on your own historical tickets before committing volume.

5. Sierra

Best for: Large enterprises that want heavily customized agents and accept contract-defined outcomes and engineering-led deployment.

Sierra, founded by former Salesforce co-CEO Bret Taylor, is the most explicitly bespoke platform on this list: agents are built with an engineering-oriented SDK, deployments are customized per enterprise, and pricing is outcome-based, with the customer and Sierra defining together which business outcomes count and blended models available for interactions that fall short of a full outcome. In one sense this is the same philosophy as customer-defined resolution: the definition is negotiated rather than imposed. The difference is transparency and testability. Sierra publishes no standard rate card, no platform-wide resolution figures, and no self-serve way to benchmark before an engagement, so everything depends on what your team negotiates and how well you specify outcomes up front.

Sierra does not prominently document per-customer end-to-end resolution rates in the way this guide would like to cite, and its third-party review footprint remains comparatively small for a company of its valuation. For enterprises with the negotiating weight to define outcomes precisely and verify them independently, Sierra is a credible partner. For buyers without that negotiating power, contract-defined outcomes without published benchmarks concentrate a lot of trust in the sales process.

6. Ada

Best for: Enterprise teams that want no-code automation breadth and multilingual coverage, and will independently audit the resolution scoring.

Ada is one of the most established AI-native support platforms, with a mature no-code builder, 50+ language support, and wide brand awareness. Its measurement approach is the self-graded family: Ada reports an automated resolution metric scored by its own evaluation model, judging whether conversations were resolved without human involvement. Self-grading is better than raw containment counting, and Ada has invested visibly in the measurement problem. The limitation is auditability: the rubric belongs to the vendor, and a buyer cannot independently verify why a given conversation was scored resolved.

The pricing model compounds the concern: Ada's published positioning is conversation-based pricing, so the bill accrues whether or not the issue was fixed, and buyers should confirm in the contract whether escalated or unresolved conversations remain billable. Ada does not prominently document named-customer end-to-end resolution rates with stated definitions. If Ada's breadth fits your needs, pair it with your own QA sampling of conversations Ada scored as resolved; the divergence between the model's grade and your customers' reality is the number that matters.

7. Forethought

Best for: Teams focused on triage, routing, and deflection in front of an existing helpdesk.

Forethought built its reputation on Solve, Triage, and Assist: AI that sits in front of helpdesks like Zendesk and Salesforce, classifying tickets with fine-grained intent detection, routing them to the right team, and deflecting the answerable ones. That heritage is the important context for resolution claims, because triage and deflection are exactly the metrics this guide distinguishes from resolution: a perfectly routed ticket and a deflected ticket are both still unresolved from the customer's perspective. Forethought's proof-of-value engagement model, demonstrating results on your data before you commit, is a genuinely good practice that aligns with this article's test-on-your-tickets thesis.

Forethought does not prominently document a standard public resolution definition or rate card; pricing is enterprise-quoted and usage-based. It is comparatively lighter on live voice and native end-to-end action execution than the agent-first platforms above. For a team whose bottleneck is misrouted tickets and repetitive FAQ load, Forethought is a sensible fit; for a team whose goal is verified end-to-end resolution of complex tickets, measure it under that definition explicitly during the proof of value.

How do you benchmark resolution rate on your own tickets?

Everything above argues that the only resolution number worth trusting is one produced from your queue, under your definition. Here is how to produce it before signing a contract.

1. Write your definition of resolved first. Before any demo, define resolution for your top ticket categories in one sentence each. A refund ticket is resolved when the refund is issued and confirmed, not when the refund policy has been explained. A card dispute is resolved when the dispute is filed and the customer knows the timeline. If a definition feels hard to write, that category is where vendor numbers will mislead you most.

2. Pull a hostile sample of historical tickets. Take 100 to 300 real tickets weighted toward your hardest categories: multi-message threads, backend-action tickets, angry customers, edge cases. A sample of password resets will tell you nothing except that every vendor can reset passwords.

3. Run simulations, not demos. Replay those tickets against the configured agent and score the transcripts against your written definitions. Lorikeet exposes this directly: simulations run the agent against your historical tickets so you can read every transcript before a single live customer is exposed. Whatever platform you evaluate, insist on an equivalent; a vendor that will not let you test on your own tickets is asking you to buy the average of other people's queues. This matters double for seasonal and peak-volume operations, where the ticket mix itself shifts under load.

4. Score quality, not closure. For every simulated conversation the agent "resolved," ask whether the customer would agree. Closure without correctness is the failure mode that containment metrics hide. This is also the standard to hold after go-live: 100% QA coverage against your rubric, not a sampled dashboard number, is what keeps the production rate honest as your volume scales.

5. Put the definition in the contract. Whatever definition you tested against goes into the agreement, along with the measurement method, your audit rights, and the commercial consequence of a bad resolution. If the vendor prices per resolution, confirm in writing that escalations are free and that resolutions you mark as bad are not charged. Lorikeet's public pricing already works this way; other vendors may agree to it if asked, and their reaction to the request is itself useful data.

What should you ask every vendor about resolution rate?

1. What exactly do you count as a resolution, in writing? Get the definition as a document, not a verbal answer. If the definition includes any form of handoff or containment, you want to know before you compare numbers.

2. Who grades a conversation as resolved: your model, your team, or ours? Self-graded metrics need independent audit; customer-graded metrics need tooling that makes grading cheap.

3. Do we pay anything for conversations that are not resolved? Per-seat, per-conversation, and usage-committed pricing all bill for failure. Ask for the price of an escalated conversation specifically.

4. Can we run your agent against 200 of our historical tickets before we sign? This is the single highest-signal question on the list. The answer separates platforms confident in their fit from platforms confident in their averages.

5. What happens commercially when we mark a resolution as bad? A refund or credit aligns incentives; a dashboard annotation does not.

6. Which named customers publish their resolution results, and under what definition? Platform-wide averages are marketing; named, contextualized, published results are evidence. Ask for the definition behind each number offered.

What are the red flags in resolution rate marketing?

• A universal benchmark quoted without a definition. "AI resolves X% of tickets" with no stated rubric and no named customers is the industry's most common claim and its least meaningful. This is the first and biggest flag.

• Containment theater. Metrics that count silence as success: the customer stopped replying, the chat window closed, the ticket auto-closed after 48 hours. If abandonment is indistinguishable from resolution in the vendor's analytics, the headline number is measuring patience, not performance.

• Deflection sold as resolution. Watch for the words contained, deflected, and automated being swapped in and out of sentences where resolved appeared in the headline. See the full breakdown of that distinction.

• Pricing that profits from failure. A vendor billing per conversation or per seat earns the same revenue whether the customer was helped or abandoned. Incentives are not everything, and they are never nothing.

• No pre-purchase testing path. A platform that cannot replay your historical tickets before go-live is asking you to accept its average as your forecast.

• Rates without recency or context. A number from a single easy quarter, a single FAQ-heavy segment, or an unnamed "leading customer" is an anecdote wearing a percentage sign.

Why Lorikeet

Lorikeet's answer to the resolution-rate question is structural rather than numerical, and it is the reason this guide can be blunt about the industry's definitional games without hypocrisy. There is no Lorikeet platform-wide resolution percentage in this article because no such published number exists, and the company's position is that publishing one would repeat the exact error this guide documents: averaging over ticket mixes that have nothing to do with yours.

What exists instead is verifiable machinery. Resolution is customer-defined and QA-enforced, with Coach providing 100% conversation coverage against your rubric instead of sampled spot checks. Pricing is per resolution with no platform, seat, or setup fees; escalations to humans are never charged; and resolutions you mark as bad are not charged, which means the metric that appears on your invoice is one you control and can dispute. Before you buy anything, simulations run the agent against your own historical tickets so the forecast comes from your queue, not a brochure. The guardrail and audit architecture underneath is built for regulated support, where a wrong "resolution" is a compliance event rather than an inconvenience, which is why the published proof points come from businesses like FCA-regulated Carmoola. Ask for the simulation run before you ask for the rate; that order of operations is the whole point.

Verdict: what resolution rate should you actually expect?

Expect a range, conditional on your queue, and distrust anyone who offers a number before seeing your tickets. The published evidence in this guide spans Fin's self-reported 76% average under a definition that includes procedure handoffs, and Carmoola's published 60% of inbound resolved end to end under a customer-defined, FCA-scrutinized rubric with 90% on outbound. Those numbers are not comparable to each other, and that incomparability is the honest state of the market.

By buyer type: if your queue is dominated by documentation-answerable questions and you want the best published economics under a published definition, Intercom Fin is the strongest pick, with eyes open about what an outcome includes. If you live inside Zendesk, its per-automated-resolution add-on is the path of least resistance, costed across the whole suite. If you are a large enterprise negotiating bespoke outcomes with engineering support, Sierra and Decagon are credible, provided your contract does the definitional work their public materials do not. If you want breadth and multilingual no-code automation with independent auditing, Ada fits. If your problem is triage in front of an existing helpdesk, Forethought fits. If your support is complex or regulated and you want resolution defined by you, tested on your own hardest tickets via simulation, verified by 100% QA, and priced so failure costs nothing, Lorikeet is the platform built around exactly that position, with helpdesk integrations that let it run alongside what you already have. Whichever you choose, the benchmark that matters is the one you run yourself.