Customer-Support Agents: Inside the Resolution-Rate Arms Race

Table of Contents

• Why Resolution Rate Became the Only Number That Matters
• The Per-Resolution Pricing Earthquake
• How the Headline Number Gets Gamed
• What Actually Moves Resolution Rate
• Tool access beats model quality
• The knowledge-base tax
• Confidence calibration and the handoff
• The Real Ceiling Nobody Advertises
• Buying Without Getting Burned
• Where This Fits in the GaaS Landscape
• Insights Most People Overlook
• References

Why Resolution Rate Became the Only Number That Matters

For fifteen years, support teams optimized for the wrong things. CSAT was easy to fake with a friendly closing line. First-response time rewarded canned auto-replies. Deflection rate -- the share of customers who never opened a ticket because a help article answered them -- counted people who gave up and churned right alongside people who got genuinely helped. None of these measured the thing a support org exists to do, which is solve the customer's problem.

Agentic support changed the conversation because, for the first time, software could plausibly own a ticket from "hi, my order never arrived" to "here's your refund confirmation" with no human in the loop. Once that was technically possible, the question stopped being "how fast did we respond" and became "did the machine actually finish the job." That is resolution rate: of all the conversations an agent handles, what fraction reach a resolved state the customer accepts, without escalation to a person.

The metric stuck because it aligns with how buyers now think about value. A support VP doesn't want a faster chatbot. They want headcount they don't have to hire, queues that don't back up at 2 a.m., and a cost-per-ticket curve that bends down instead of up as volume grows. Resolution rate is the closest single proxy for all of that. Intercom built its entire Fin product line around publishing the figure prominently, and the rest of the category followed. When Klarna very publicly claimed its OpenAI-powered assistant was doing the work of 700 agents, the underlying brag was a resolution-rate claim dressed up as a headcount story.

The Per-Resolution Pricing Earthquake

Here's what actually made resolution rate non-negotiable: vendors started pricing on it.

The old SaaS model charged per seat or per agent. That worked when humans did the work and software just helped them do it. But an autonomous agent has no seat -- it's not a person sitting at a desk -- so charging per seat made no sense. A handful of vendors made the leap to charging per resolved conversation, typically somewhere in the range of $0.99 per resolution, and the economics of the whole category flipped.

This is the per-outcome pricing model that defines modern Agentic AI-as-a-Service, and customer support is where it went mainstream first. The shift matters more than it looks. Under per-seat pricing, a vendor profits whether or not the software works -- you paid for the license either way. Under per-resolution pricing, the vendor only gets paid when the agent actually closes a ticket. Their revenue is now bolted directly to the metric the buyer cares about. Intercom's decision to charge $0.99 per Fin resolution was widely covered as a turning point precisely because it put the vendor's money where its resolution-rate claims were. a16z has written extensively about how agentic products are moving from per-seat to per-outcome pricing, and support agents are the clearest live example.

But the incentive blade cuts both ways, and that's the catch that buyers underestimate. Once a vendor gets paid per resolution, the vendor has a powerful financial reason to define "resolution" as generously as the contract allows. The pricing innovation that aligned interests on cost simultaneously created a measurement war over what counts.

How the Headline Number Gets Gamed

Ask three vendors for their resolution rate and you will get three numbers that are not comparable, because they are not measuring the same thing. The gaming happens in the denominator and in the definition of "resolved," and it is mostly invisible unless you go looking.

The most common trick is what counts as a "resolved" conversation. Some vendors mark a ticket resolved if the customer doesn't reply within a set window -- say, 24 hours of silence. But silence isn't success. A frustrated customer who gives up and never responds gets logged identically to one whose problem was genuinely fixed. This is the old deflection-rate sin wearing a new coat. The honest measure is confirmed resolution, where the customer explicitly says the issue is handled or a downstream signal (a non-returned refund, an order that ships) verifies it.

The second trick lives in the denominator. A vendor can quote an 80% resolution rate by quietly excluding every conversation the agent never attempted -- the ones it routed straight to a human because they looked hard. Filter out the difficult tickets before you calculate, and your rate on the easy remainder looks spectacular. The number that matters to a buyer is resolution rate over total inbound volume, not over the curated slice the agent chose to engage.

A third, subtler distortion: counting a resolution every time the agent answers a sub-question inside a longer conversation. Split one messy ticket into five "resolved" micro-interactions and the rate inflates while the customer is still stuck. Reddit threads from support managers evaluating these tools are full of stories about dashboards that looked great until someone reconciled them against actual refund volumes and reopened tickets.

None of this means the vendors are lying, exactly. It means "resolution rate" is a category of measurement, not a standardized one, and the absence of a shared definition is itself the story.

What Actually Moves Resolution Rate

Strip away the marketing and the real determinants of how many tickets an agent can close are surprisingly mundane. The model is rarely the bottleneck.

Tool access beats model quality

The single biggest lever is whether the agent can do things, not just say things. An agent that can read your order management system, issue a refund through Stripe, update a shipping address in your commerce backend, and cancel a subscription will resolve a vastly higher share of tickets than a smarter model that can only talk. Most support tickets are not knowledge questions -- they are action requests. "Where's my order," "cancel my plan," "change my address," "I was charged twice." These resolve only if the agent has write-access to the systems of record. This is why depth of integration, not raw model intelligence, is the real moat in vertical agents, a theme that runs through this whole cluster. The vendor with twelve native, write-capable integrations into the tools your business already runs will beat the one with a marginally better language model and read-only access every time.

The knowledge-base tax

For the tickets that are knowledge questions, the agent is only as good as the documentation behind it. Garbage in, garbage out is brutally literal here: an agent pointed at a stale, contradictory, half-migrated help center will hallucinate or escalate, and there is no prompt clever enough to fix bad source material. Vendors quietly know this, which is why onboarding for a serious deployment is mostly a knowledge-cleanup project wearing an AI costume. The uncomfortable truth buyers discover is that they're not buying an agent; they're buying a forcing function to finally fix the documentation they've neglected for years.

Confidence calibration and the handoff

The agents that post sustainable resolution rates are the ones that know what they don't know. A well-calibrated agent escalates the 20% of tickets it would botch and confidently closes the 80% it can handle. A poorly calibrated one barrels ahead on ambiguous cases, gives a wrong refund or a confidently wrong policy answer, and generates a reopened ticket plus an angry customer -- which is worse than no agent at all. The handoff to a human is not a failure of the agent; it is the feature that makes the high resolution rate trustworthy. Anthropic's published guidance on building effective agents emphasizes exactly this -- knowing when to act autonomously versus when to defer is a design decision, not an emergent property.

The Real Ceiling Nobody Advertises

There is a number vendors will not put on a billboard: the resolution rate ceiling for a given business is set by the business, not the agent.

Some share of every support queue is genuinely unresolvable by software -- and not because the AI is weak. A customer who needs an exception to policy that only a manager can grant. A retention case where the right move is a human empathizing and offering a save. A legal or safety-flagged situation where regulation requires a person. A billing dispute that hinges on facts living in someone's email, outside any system the agent can read. These tickets have a structural ceiling. Push an agent past it and resolution rate goes up while customer trust goes down, because the machine is now "resolving" things that should have escalated.

This means the right target resolution rate is industry- and account-specific. A high-volume e-commerce store with clean policies and good integrations might genuinely hit 70-80% autonomous resolution. A wealth-management firm or a healthcare provider, where liability and nuance dominate, might top out at 30% and be right to -- the same dynamic that constrains agents in regulated verticals across this cluster. A buyer who imports a SaaS vendor's e-commerce benchmark into a regulated business is setting themselves up to either miss the target or, worse, hit it by over-automating cases that needed a human. McKinsey's research on generative AI in customer care lands on a similar point: the value shows up unevenly, and the highest-leverage gains cluster in specific, automatable ticket types rather than across the board.

The honest framing is that resolution rate has a numerator the vendor can influence and a denominator the business defines. The arms race is really about who can responsibly expand the numerator without quietly poisoning the denominator.

Buying Without Getting Burned

If you're evaluating support agents, the headline resolution rate on the pitch deck is the least useful number in the room. A few questions cut through it fast.

Ask exactly how "resolved" is defined in the contract, and whether it's confirmed resolution or no-reply timeout. Insist the metric be measured over total inbound volume, not the agent's self-selected slice. Run a paid pilot on your ticket data -- not the vendor's demo -- because resolution rate is so dependent on your integrations and knowledge base that any benchmark from another company is close to meaningless. Watch the reopen rate and the post-agent CSAT alongside resolution, because an agent optimizing purely for closures will happily close tickets badly. And under per-resolution pricing, audit what you're being billed for; a vendor paid per resolution and trusted to count its own resolutions has an obvious incentive you should verify against your own logs.

The teams that win with these tools treat the vendor's number as a hypothesis, not a fact, and the pilot as the experiment that tests it.

Where This Fits in the GaaS Landscape

Customer support is the proving ground for Agentic AI-as-a-Service, not a sideshow. It went first because the work is high-volume, well-bounded, and sits on top of clean digital systems of record -- the ideal conditions for an agent. The pricing model that emerged here, per-resolution and per-outcome, is now propagating across every vertical agent category, from IT-helpdesk agents handling tier-1 support to HR agents fielding employee questions. The measurement debate playing out over resolution rate is a preview of the same fight other categories will have over their own outcome metrics. And the lesson that depth of integration -- not model horsepower -- drives real-world performance is the central thesis of vertical agents generally. Whoever masters resolution rate honestly in support is writing the playbook the rest of the GaaS market is about to copy.

Insights Most People Overlook

• Per-resolution pricing quietly made the vendor your adversary on measurement. Everyone celebrated outcome pricing for aligning incentives on cost -- you only pay when it works. Almost no one noticed it gave the vendor a direct financial motive to define "works" as loosely as the contract permits. The pricing model that fixed one misalignment created another, and the second one is harder to see because it hides in the metric definition.

• A high resolution rate can be a symptom of over-automation, not success. Past a certain point, rising resolution rate means the agent is closing tickets that should have escalated -- denying an exception, mishandling a retention case, "resolving" a dispute the customer didn't accept. The healthiest deployments sometimes cap their resolution rate deliberately. A number that only goes up is a red flag, not a trophy.

• You're not buying an agent, you're buying a documentation-cleanup mandate. The biggest determinant of knowledge-question resolution is the quality of your help center, which most companies have neglected for a decade. Deploying a support agent forces the cleanup that should have happened years ago. Half the realized value often comes from the docs you're finally forced to fix, not the AI itself.

• The escalation is the product, not the failure. Buyers instinctively read every handoff to a human as the agent falling short. It's the opposite. A well-calibrated handoff on the cases the agent would botch is precisely what makes the autonomous resolutions trustworthy. Vendors that minimize escalation to inflate their headline rate are degrading the very number they're advertising.

• Tool access is the moat, and it's why a "worse" model often wins. The market keeps benchmarking these agents on language quality when most support tickets are action requests, not knowledge questions. An agent with write-access to your refund, shipping, and subscription systems beats a smarter model stuck in read-only mode on almost every real queue. The defensibility lives in the integrations, which is also why these vendors are stickier than they look.

References

• The Cost of Compute: A $7 Trillion Race to Scale Data Centers / AI pricing -- a16z on the shift from per-seat to per-outcome AI pricing
• Building Effective Agents -- Anthropic on agent design, autonomy, and when to defer to humans
• Where is customer care in 2024? -- McKinsey on generative AI's uneven value across customer-care ticket types