AI customer support agent · private beta
It answers your customers from your help centre, cites the document behind every reply, and hands over what it cannot answer. Then it shows you the conversations that went wrong, next to the ones that went right.
For the person who has to defend the number, not just present it.
27 points the usual dashboard never shows you
The number you are sold
Many AI support tools score every conversation nobody complained about as a win. Someone who left in frustration counts the same as someone who got their answer, and the vendor defines, measures and is often paid on that number.
Support Proof counts resolved, abandoned and handed-over conversations separately, and never folds abandoned into resolved. It shows how often it answered without a source. And when your own settings are what kept it quiet, it says so.
How the figures are countedHow it works
Publish your articles. Every later edit re-runs the questions an article used to answer, and a change that breaks one is blocked.
Add the tag and the origins allowed to use your key. No integration project, and it never touches a card number.
Resolved, abandoned, handed over and declined are counted separately, and answers without a source show up as a number.
The publish gate
Publish a change to an article and the questions that article used to answer are asked again against the new text. If one of them stops having an answer, the publish is blocked and you are told which question broke.
This is the part most tools leave to hope. An article is edited on a Tuesday, the answer quietly degrades, and nobody finds out until a customer does. Here the check runs at the moment of the edit, against the questions you already know people ask.
How the gate decidesPublish blocked — one question lost its answer
What it does differently
And the revision of it. When an article changes, you can see which answers came from the old text.
Can I get a refund after 30 days?
Refunds are available within 60 days of purchase for annual plans. Monthly plans can be cancelled at any time and are not refunded.
Refund policy · revision 12rev 11 retired
A question your knowledge base cannot answer gets a handover, not a confident guess stitched from nearby paragraphs.
Editing an article re-runs the questions it used to answer. If the edit breaks one, the publish is blocked.
When your thresholds kept the agent quiet, it tells you, working against its own numbers on purpose.
Speech to text, a text model, then speech: about $0.012–0.020 a minute against $0.08 for speech-to-speech, by our dated cost notes.
From the moment it promises a person, the customer's messages go to your queue and the agent stops replying.
Honest accounting
| The usual pitch | What Support Proof reports |
|---|---|
| “72% resolution rate” | Resolved, abandoned and escalated counted separately. Abandoned is never folded into resolved. |
| “AI-powered insights” | An overlap, labelled as an overlap. It will not claim a silence cost you three customers, because it cannot know that. |
| Dashboards that only go up | A page you can publish showing what it declined and what it answered without a source. |
The trust page is opt-in and revocable, and carries counts, rates and dates, never a customer's words.
Free tool · no sign-up
Enter your ticket volume, handling time and hourly cost. The calculator works out the share of tickets an agent would have to actually resolve before a plan covers its cost: the rate you would need, not a rate we claim.
Work out your break-evenThey report a resolution rate they define, measure and are paid on. Intercom counts a handoff to a human as an outcome. We count resolved, abandoned and escalated separately, never fold abandoned into resolved, and publish the same data scored their way next to ours so you can see the gap.
A public page you choose to publish, built from the same ledger your own dashboard reads. It shows the last 30 days: what the agent answered, what it declined to answer, how often it answered without a source document, how long people waited, and what your own alert thresholds suppressed. Counts, rates and dates only - never a customer's words. It is opt-in and you can withdraw it at any time.
A rate needs a denominator worth dividing by. Below 30 conversations we publish the count and say what it is out of, rather than a percentage that would swing on a single case.
A public comparison built only from accounts whose trust page is live, computed by the same code that renders those pages. Cohorts are role plus volume band, and we refuse to report on a cohort smaller than 5 accounts. It is self-selected - accounts that publish are more likely to be pleased with their numbers - and the page says so.
Every answer cites a document and the revision of it that was live at the time. When the knowledge base cannot answer, it hands over rather than assembling a confident guess from adjacent paragraphs - and how often it answered without a source is one of the numbers on the trust page.
Publishing replays the past conversations whose answers cited that document. If an answer that used to be supported stops being supported, the publish is blocked and you are told which answers broke. Live traffic never sees a candidate revision.
200 messages, one seat, no card. It stops at the cap and tells you. It does not begin charging on day fifteen.
One script tag on your site. It does not live inside your application, and it never holds a card number - checkout is the payment provider's.
Private beta
Join the list and we will invite you when there is a place. You get the product as it is today, including the parts it reports going wrong.