Free tool
Measure your AI's resolution rate yourself, from your own conversations
Bring an export or paste a few dozen conversations. In a couple of minutes you'll see how many your AI actually resolved, how many customers left mid-way, and how many can't be told either way, each with the reason. No signup, and no login to anything.
- In the Inbox, select the conversations Fin handled (200 to 300 at a time).
- Open the "..." menu and choose Export as text.
- The transcripts download as separate files. Put them in one zip (on a Mac: select them, right-click, Compress) and upload the zip.
A CSV, JSON or text file, or a zip of them, up to 20 MB. We read the latest 200 conversations.
Deleted after the check. Never used to train AI. Privacy
How it counts
Every conversation goes through the same four questions, in order. Did the AI fail to deliver what was asked, or hand it to a person? Not resolved. Did it deliver, and did the customer's last message show it settled? Resolved. Did the customer leave while the AI was waiting on them? Abandoned. Did the AI answer and the customer say nothing more? Unclear.
That last one is where vendors differ most: some count a customer who goes quiet as a resolution. So the result is a range, from counting every unclear conversation against the AI to counting every one for it. I run DeskClone, an AI support company, and built this to hold my own agent to account. It works the same on anyone's.
Questions
What does the Resolution Auditor count as resolved?
A conversation is resolved when the customer got what they came for (the refund, the change, a specific and correct answer) and their last message shows it settled. If the customer went quiet after an answer, we count it as unclear, not resolved, and the result shows a range from counting every unclear one against the AI to counting every one for it.
Why do you only count conversations the AI handled?
Because that is what an AI resolution rate is about. Where the file shows a person replied, the conversation is your team's work and is left out, and the result says how many were. Where the file doesn't say who replied, those conversations are included and labelled.
How accurate is it?
We tested it on 120 real chats from an online store (the public ABCD research dataset). Two AI models, Claude and GPT-4o, labelled each one independently and agreed on 104. The auditor matched them on 89% of those. Every row in your result shows its reasoning, so you can check it yourself.
What happens to my conversations?
Your file is deleted when the check finishes (within a day if it fails), and your result an hour later. The text is never kept; what stays is the totals and basic visit data, for an industry benchmark. OpenAI's model reads them on our behalf and doesn't train on them. Email addresses and phone numbers are hidden from it. Details are in our privacy policy.
Do you need my helpdesk login or an API key?
No, and we never will. You upload an export or paste conversations, and that's all we see.