Every buyer of enterprise software has lived this moment. The vendor's on the call, screen-share up, and someone asks: "What's this going to cost us — really?" The answer comes back as a percentage. Ninety-eight percent accuracy. Thirty percent deflection. A number that sounds like precision and functions like a curtain.
That conversation is happening right now, on repeat, across every fintech, health, and SaaS procurement team evaluating agentic AI for customer service. Almost none of them ask the one question that would end the meeting early: what, exactly, counts as "resolved"?
The published number that sets the floor for honesty
Fini Labs processed more than two million customer service inquiries across fintech, health, and SaaS clients, publishing a headline of 98% accuracy and a price of $0.69 per resolution, according to Fini Labs. That number is real, published, and specific enough to serve as a floor for the category rather than its ceiling. Compare it to what the rest of the market discloses, and the gap stops looking like competition. It starts looking like a tell.
Here's the mechanism. Cost per resolution is calculated by dividing the unit price a vendor charges by the resolution rate they report. Simple math, as long as "resolution" means the same thing everywhere. It doesn't. Some vendors define resolution as a confirmed outcome: the customer's problem is fixed, verified, closed. Others define it as the absence of a response. Silence.
Silence is cheap to manufacture and expensive to detect. That makes it the perfect lever for a vendor who wants their per-unit cost to look lower without doing the harder work of actually lowering it.
Silence gets billed at the same rate as success
According to Intercom's own documentation of its Fin product, cited by BuildMVPFast, a conversation gets marked resolved if the customer goes silent for 24 hours after Fin's last response and never returns. Intercom bills $0.99 for it, the same price as a resolution the customer actually confirmed. There is no distinction on the invoice. A frustrated customer who gave up and a satisfied customer whose problem got solved generate identical revenue for the vendor. That is not a rounding error in a KPI. That is the KPI being restructured to reward disengagement.
Run that forward and the incentive gets worse, not better. Gartner projects that by 2030, generative AI's cost per resolution for customer service will exceed $3, higher than many offshore human agents cost B2C, according to Gartner. If "resolution" keeps absorbing silence as a countable outcome, vendors facing that price ceiling have a straightforward way to protect margins: let more conversations go quiet, and count the quiet as success. The metric meant to measure whether the AI worked becomes a lever for making the AI look like it's working while doing less.
This isn't isolated to one vendor's fine print. A separate pattern shows up in how the category talks about itself in public. An analysis by Progressive Robot found that across seven pages of marketing material totaling 9,566 words, Cresta's public disclosures contained 27 distinct percentage figures and exactly two dollar amounts, both revenue figures, neither one a price, according to Progressive Robot.
Twenty-seven ways to describe how good the system feels. Zero ways to know what a resolution costs.
The gap between what a benchmark reports and what a customer actually experiences is not a rounding error.
Meanwhile the rest of the category has already converged on cost per resolution as the standard unit of price: Zendesk at $1.50, Salesforce at $2, Intercom starting at $0.99. When most of a market agrees to disclose a number and one player hasn't, the silence isn't neutral. It's a choice, and it's usually made because the number, once published, invites the exact question this piece is asking.
A fifty-fold price gap has nothing to do with efficiency
The CLEAR framework's own findings sharpen the stakes further: a 37% gap between lab benchmark scores and real-world deployment performance, and a 50x cost variance between approaches achieving similar accuracy on the same tasks.
Fifty times. Not fifty percent. Fifty times.
That variance doesn't exist because some vendors are twice as efficient. It exists because "accuracy" and "resolution" get measured against different, undisclosed definitions, and buyers have no instrument to normalize them before signing.
The three questions a vendor cannot dodge in writing
So the operator's task isn't finding the vendor with the best-looking percentage. It's asking, before any contract, three things. How is resolution defined in the SLA: confirmed outcome or elapsed silence? What happens to the invoice when a customer disengages without confirming? Will the vendor publish resolution rate alongside price, in the same sentence, the way Fini did? A vendor who can't or won't answer those questions in writing has already told you which side of the metric they're optimizing for.
This is the layer where infrastructure earns its name instead of borrowing it. A brand running its own conversational agents inside a system built on its own Blueprint, its actual customer data, its actual resolution definitions specified and verifiable before deployment, never has to ask a vendor to define success on its behalf. The specification lives inside the brand's own architecture, not inside someone else's invoice logic. That's the difference between buying a percentage and owning the definition behind it.
The gap between what a benchmark reports and what a customer actually experiences is not a rounding error.
It's the business model.
❯ Unrealville Studio
