AI customer support can resolve roughly half to three quarters of conversations end to end in 2026, depending on your queue and on who is doing the counting. Vendor-reported platform averages run from about 52% automated resolution (Ada, across 550+ deployments) to 76% (Fin (now part of Salesforce), across 12,000 customers), and Fin's own guide puts first deployments at 40% to 60%. Lorikeet does not publish a platform average; its customers publish their own figures, such as Carmoola resolving 60% of inbound support end to end.

Every one of those numbers is self-reported under a definition its publisher wrote. This guide lists eight benchmarks, says who reported each one and what it actually counts, and explains how to turn them into a target you can defend for your own queue.

Key takeaways

• Published averages sit between 52% and 76%. Both are vendor-reported across each vendor's own customer base. Best deployments reach the 80s on both vendors' pages.

• Containment runs about 20 points above resolution. Ada reports 72% containment against 52% automated resolution on the same deployments. A rate without a written definition is probably containment.

• Expect 40% to 60% at launch. Fin's benchmark page gives 60% or more after 6 to 12 months, and says teams without structured knowledge stall between 30% and 45%.

• Backend access moves the number most. Ada reports Dott going from 32% to 77%; Chexy's agent went from about 24% to about 72% of inbound in the week its systems were connected.

• Decision rule: write your definition of resolved per ticket type, replay 100 to 300 of your own historical tickets, and treat repeat contact within 24 to 48 hours as the check on any headline rate.

Which AI support platforms publish how they count a resolution?

Most AI agent vendors now bill per outcome, so the billing definition is the fastest way to see what a rate means. Lorikeet charges only for successfully resolved tickets; if you are unhappy with how a ticket was handled you do not pay for it, and escalations to a person are not charged (pricing). Fin (acquired by Salesforce on 10 September 2026) counts an outcome when the customer confirms the issue is resolved, does not ask for more help after Fin responds, or Fin completes a Procedure, including handoffs (Fin pricing). Zendesk bills automated resolutions, requests resolved by an AI agent without escalation to a human (Zendesk help centre). Notch says it charges only for tickets that meet full end-to-end autonomous handling (Notch).

How we chose and checked these benchmarks

We included every resolution figure that a vendor or operator publishes on a public page with a stated sample or a stated definition, plus the two definitions buyers most often meet in contracts. We excluded analyst predictions we could not open on the publisher's own site on the day of checking, and any figure without a visible source page.

Every figure below was checked on the publisher's own page on 30 October 2026, and each is linked. Figures are labelled as vendor-reported or operator-reported because none of them is independently audited. Lorikeet publishes this guide and is one of the tools listed. We have tried to describe every product, including our own, using only what each vendor publishes, and we link to the source for each claim.

Resolution rate benchmarks at a glance

1. Fin (now part of Salesforce): 76% average across 12,000 customers

Fin publishes the largest sample in the market. Its 2026 benchmarks page gives an average resolution rate of 76% across 12,000 customers, improving about 1% a month, with top performers at 80% to 84%. Its resolution rate guide adds that some implementations reach 93%, and its resolution versus deflection guide says Fin started at 23% before climbing to 76%.

What it counts. The billing outcome on Fin's pricing page is broader than a fixed problem: a customer who does not ask for more help after Fin responds counts, and so does a completed Procedure that ends in a handoff. To its credit, Fin charges at most once per conversation and refunds the outcome if the conversation is reopened. Price: $0.99 per outcome, with a $49 monthly minimum that includes 50 outcomes.

Fair read. The strongest published dataset, measured under the vendor's own outcome definition and skewed to Fin's customer base. Use it as a ceiling for well-documented, transactional queues, not as a forecast for a regulated one.

2. Lorikeet customers: per-company rates, no platform average

Lorikeet does not publish a platform-wide resolution rate. Its evidence is per customer, printed on live story pages, and each customer uses its own wording:

• Carmoola, a UK car finance lender regulated by the FCA: previous automation on a knowledge base and form-driven prompts resolved about 30% of inbound. With Lorikeet it resolved 40% end to end from day one, and now resolves 60% of all inbound on WhatsApp, chat and email end to end.

• Hnry, whose customers are self-employed: live from mid-May 2026 with zero prior automation, 17,000+ conversations in the first month, and around 70% automated in the peak week of Australian end of financial year, from roughly 58% across the full period.

• Chexy, a Canadian bill payments app: the agent's share of inbound handled moved from about 24% to about 72% in the week it was opened to the full queue, after payment systems were connected.

• Breeze: 40% of complex support volume resolved independently within 30 days. Wonderschool (voice): 41% of genuine parent calls fully resolved, with a further 44% handed off cleanly.

Honest limits. These are different measures. Hnry's story says "automated" and Chexy's says "handled", which are not the same claim as Carmoola's "resolved end to end", so do not average them or set them beside Fin's 76%. If you want one platform number for a slide, Lorikeet will not give you one, and customer-defined resolution means you write the rubric before launch. Pricing follows the same rule: only successfully resolved tickets are charged, at $0.99 (Start) or $0.90 (Scale) for chat, email and SMS, on top of a plan fee of $2,100 or $5,100 a month paid annually.

3. Ada: 52% resolution against 72% containment

Ada's automated resolution guide is the most useful page for understanding the gap between metrics, because it publishes both numbers on the same base. Across more than 550 deployments, containment averages around 72% while automated resolution averages around 52%. Best-in-class deployments reach 84% or higher, its best fintech customers have reached 84%, and most enterprise deployments launch in the 50% to 65% range.

Ada's customer Dott shows what action access does: its automated resolution rate went from 32% to 77%, more than doubling within months of launch.

Fair read. Vendor-reported, and the grader is Ada's own. The 20-point gap is the single most important number in this list, because it is the difference between a conversation ending and a problem being fixed.

4. Notch: 77% autonomous resolution within 12 months

Notch reports that across more than 20 million conversations in insurance, ecommerce, SaaS and regulated industries, its customers reach 77% autonomous resolution within 12 months (Notch benchmarks). It defines resolved as full end-to-end autonomous handling and excludes deflection, containment and interactions that need a customer follow-up.

The same page offers a four-tier framework: legacy chatbots at 10% to 25%, standard AI assistants at 40% to 60%, AI-native platforms at 55% to 70% first contact resolution, and agentic platforms with deep backend integration at 70% to 85%. Notch presents those tiers as a capability guide rather than measured data, and it is right to: they are orientation points, not survey results.

Fair read. A strict published definition, a large sample, vendor-reported.

5. Klarna: two-thirds of chats handled in the first month

Klarna's February 2024 announcement said its AI assistant held 2.3 million conversations in its first month, two-thirds of its customer service chats, doing the equivalent work of 700 full-time agents, with a 25% drop in repeat inquiries and errands resolved in under 2 minutes against 11 previously.

Note the verb. "Handled" is a share of conversations the assistant took, not a resolution rate. The repeat-inquiry drop is the closer signal of genuine resolution. This is an operator's own launch announcement, now more than two years old, and it is still one of the most quoted figures in the category.

6. Zendesk: a strict definition, no published average

Zendesk's AI resolution rate explainer defines the metric as the share of issues an AI system fully resolves end to end without human intervention, and gives a worked example: 150 of 200 requests fully resolved is 75%. It adds the test that matters: a resolution is successful when the issue is solved with no escalation or repeat contact within 24 to 48 hours.

Commercially, Zendesk bills automated resolutions, requests resolved by an AI agent without escalation to a human, with an allowance included in each plan. Neither page publishes a customer-base average. If you already run Zendesk and most of your tickets are answerable from help content, its native AI agent is the simplest place to start measuring.

7. Market-wide ranges: 40% to 60% at launch

Fin's benchmarks page gives an industry range of 40% to 60% on initial deployment, rising to 60% or more within 6 to 12 months with optimisation, and says teams that launch without structured, comprehensive knowledge content stall between 30% and 45%. Its resolution rate guide lists industry ranges too, from 60% to 84% in ecommerce down to 40% to 65% in financial services and 35% to 55% in healthcare.

These ranges do not state their method, so treat them as one vendor's estimate. They are still directionally useful: regulated queues run lower because more contacts need identity checks, an action, or a human by rule.

8. Your own historical tickets: the benchmark that transfers

The only rate you can plan around is one produced on your own tickets under your own definition. Replay 100 to 300 real tickets, weighted toward multi-message threads, backend-action requests and unhappy customers, and score every transcript. In Lorikeet, imported tickets generate scenarios and test customers, batches run against a draft agent with tool responses mocked, and teams agree a pass-rate threshold before publishing (simulations). Fin's benchmarks page also says simulations let teams validate changes before they go live; whichever vendor you test, insist the run uses your data rather than a curated demo.

What is the difference between resolution, containment and deflection?

Resolution means the customer's problem was fixed; containment and deflection only mean a human was not involved. The three sit on a ladder from loosest to strictest:

1. Deflection. The customer was pointed at an article or portal and did not open a ticket. Fin's comparison guide puts it plainly: a 70% resolution rate that includes 30 points of article views is really a 40% resolution rate.

2. Containment. A real conversation ended without a handoff. It cannot tell a satisfied customer from one who gave up, and it rises if the escalation path is hidden.

3. Automated resolution. A contained conversation that a grading model judged relevant and accurate. Better, but the grader usually belongs to the vendor that bills for the result.

4. Verified resolution. Fixed under a rubric the buyer wrote, and confirmed by behaviour: no escalation and no repeat contact within 24 to 48 hours.

The longer version is in resolution rate versus deflection rate, and the human-team baseline is in first contact resolution benchmarks.

What resolution rate can AI customer support achieve on your queue?

Expect 40% to 60% early and 60% or more once knowledge and actions are in place, with the ceiling set by four things:

• Knowledge coverage. List your top 50 intents and check each has one current, unambiguous answer. Every gap is a guaranteed escalation.

• Backend actions. Explaining a refund policy is an answer; issuing the refund is a resolution. Chexy's jump came from connecting payment records. See how to safely let AI take actions in backend systems.

• Channel mix. Chat, email and voice behave differently. Ask for the rate by channel and set separate targets.

• Regulated share. Complaints, vulnerability and decisions that must reach a person by policy cap the rate before model quality matters.

Peaks change the picture too: Hnry's rate rose in its busiest week, which is covered in handling seasonal support spikes with AI. To turn a staged target into a cost model, see AI customer support ROI for high-volume teams.

Questions to ask before you trust a resolution rate

1. What exactly counts as resolved, in writing, and is it the same definition the invoice uses?

2. Are deflections, article views or handoffs at the end of a workflow included?

3. What is the repeat-contact rate within 24 to 48 hours on conversations marked resolved?

4. Can we audit every conversation, and what happens commercially when we mark a resolution as bad?

5. What is the rate by channel and by ticket type, on a sample of our own tickets?

What still needs a human

A good deployment keeps its rate lower on purpose in four places. Complaints, vulnerability and hardship go to trained people; the agent recognises the signal early and hands over with context. Decisions rather than lookups, such as whether a dispute or claim succeeds, stay with your team; the agent gathers facts and executes, and never forms the opinion. Gaps in the knowledge base should escalate, and each one is a content task, not a model failure. Customers who ask for a person should reach one quickly, because a visible path to a human is what keeps the resolved share genuine.

If you want to measure your own rate rather than borrow a vendor's, you can start a 30-day Lorikeet free trial and run simulations on your own tickets before anything goes live.