Resolution Rate Means Three Minutes at One Vendor and 72 Hours at Another

In June 2026 Intercom changed how it calculates Fin's resolution rate. Conversations where Fin never got the chance to answer stopped counting in the denominator.

Its own worked example: 1,000 conversations, 250 resolved. The old calculation gave 250 of 750, or 33%. The new one gives 250 of 500, or 50%. Intercom states plainly in the same document that the number of resolutions is not changing and billing is not affected.

Seventeen points of improvement, the same 250 resolutions, the same invoice. Both figures are still queryable side by side in the product, one of them labelled legacy.

That single fact retires the entire "our resolution rate is X%" genre, and it is the most honest thing any vendor in this category has published about its own metric.

The word is not a measurement

Nearly every AI support platform sells on a resolution, deflection or containment rate. The rates are quoted against each other constantly, in comparison posts and in procurement decks.

They are not comparable. Not because vendors are dishonest, but because each one has picked a different rule for the same word, published it, and moved on. Read the rules next to each other and the comparisons stop meaning anything.

How long silence has to last

The most common mechanism is assumed resolution: the customer stops replying, and after some period the conversation is recorded as resolved. The period is where the vendors diverge wildly.

Three minutes against seventy-two hours is a factor of fourteen hundred, applied to the same nominal metric. A customer who wanders off to make coffee is a resolution at one vendor and an open conversation at another.

Does giving up count as success?

This is the question that separates the definitions, and vendors answer it in opposite directions.

Tidio counts it as resolved. Its analytics documentation defines resolved conversations as those handled without an agent, in a bracket that includes where the visitor simply left the conversation.

Intercom counts it as resolved for billing and says so directly: a customer leaving after receiving Fin's answer without requesting further help is counted as an assumed resolution and will be charged. It also publishes a deflection rate that explicitly includes customers who left before receiving an answer and customers who left after reacting negatively.

Ada goes the other way. When a conversation is handed to a human, its documentation says the AI Agent marks it Not Resolved immediately following the handoff. Zendesk is equally firm: as soon as a conversation is escalated to a human agent, it can no longer be counted as an automated resolution.

So the same customer behaviour — reads the answer, is unhappy, walks away — improves the reported rate at some vendors and does nothing at others.

Where billing and reporting disagree

The more expensive question is whether the number on your invoice matches the number on your dashboard. At two vendors it does not.

Intercom bills $0.99 for a procedure handoff. Its own documentation on Automation Rate answers the question directly: does the metric include procedure handoffs? No — Automation Rate only counts conversations fully resolved by Fin. You are charged for an outcome that the headline performance metric excludes.

Freshworks does not measure outcomes at all. Its billing unit is a session — every interaction between a customer and the agent within a 24-hour window — and it charges $49 per hundred. Its documentation also confirms something worth knowing before you build: previewing an agent while you are building it consumes sessions, published or not. You pay to test your own bot.

Zendesk publishes a resolution allowance in dollars per agent seat and, across two cross-linked articles, contradicts itself on whether one tier draws down that allowance. It publishes no per-tier unit price anywhere, so a buyer cannot compute how many resolutions the allowance actually buys.

The vendors criticising the metric they sell on

Decagon is worth crediting because it argues against its own pricing model in public. Its glossary says defining a resolution can be tricky, that a customer may leave mid-conversation or get an answer that does not solve the issue, and that grey areas lead to billing disagreements.

It goes further, and this is the sentence a buyer should carry into every demo: vendors predisposed to push issues toward resolved can present risks to customers.

It also names the failure mode precisely — that an operation can post high deflection while resolution quietly declines, because customers abandon rather than escalate and technically count as deflected. And it puts a number on the correction: adjusting for customers who re-contact within 24 hours can lower reported containment by five to fifteen percentage points.

Ada asks the same question more briefly: can you tell the difference between a customer who ended a conversation because they were satisfied and one who got frustrated and gave up?

And the vendors who publish nothing

Sierra sells outcome-based pricing and its public disclosure is that if a case needs to be escalated, in most cases there is no charge, with clear agreed-upon criteria provided upfront. In most cases is doing the work in that sentence. The definition is a contract term rather than a documented product behaviour, which means it is negotiable and also that no two Sierra customers necessarily buy the same thing.

Chatbase markets resolution metrics and, when we searched its full documentation index, has no resolution, deflection or containment metric documented anywhere in the product.

What to do instead

Do not compare vendors' published rates. They are not the same measurement and the exercise produces a confident wrong answer.

Ask four things instead. How long does silence have to last before this counts. Does a handoff to a human count as resolved. Does a customer who abandons count as resolved. And is the number you bill me the same number you report to me.

Then define your own. Pick a rule stricter than any vendor's — resolved means no repeat contact on the same issue within seven days, matched on the person rather than the ticket — and measure everything against that. It will produce a lower number than any vendor quotes, and it will be the only one that means the same thing next quarter.

Definitions read from each vendor's own documentation in August 2026 and dated. Where a vendor's billing rule and its reporting rule differ, both are given. Where a vendor publishes no definition we say so rather than inferring one from a pricing page.

WRITTEN BY AR · UPDATED 2026-08-20

I read the fine print. Vendor pricing pages, billing definitions, terms, funding filings and acquisition notices — then I do the arithmetic nobody publishes: what a platform actually costs at your volume, what its headline metric is really counting, and who owns it now. I do not run benchmarks, and no page here pretends otherwise.

Editorial policy