The analyst agent explained why the user left

By Georg Malahov

Overnight came a perfect and very painful case. A user went through the entire funnel: watched the demo, signed up, got all the way to recording a call — and there the transcription service (Deepgram) wouldn't let him record. And instead of a human "looks like your network is blocking us, try turning off your antivirus or switching connections", the extension showed him a raw technical error that says nothing at all. He tried once more, gave up, and uninstalled the extension. Fair enough — I would have uninstalled it too.

I didn't dig into it by hand. I asked my analyst agent — a Claude session with access to Supabase, Stripe, the repository, and the server, the kind of thing commonly called a harness — and within a few minutes had an answer: it's almost certainly the client's network, something on his side won't let audio be recorded. Not a guess out of thin air, but a walkthrough of this specific user's specific events.

And then I realized I had been using such an analyst agent all year — just thrown together on the fly and without a name. This is exactly how the year's biggest bug was found in August: Stripe had changed the webhook response signature back in March, I'd stopped granting credits, and the client from Brazil who disputed the charge through his bank turned out to be right about the substance of his complaint. The same way cohorts and churn got counted. The same way the two of us went through all my bills and cut just under 700 euros a year in costs. Half of this year's decisions were made with its help — I just never treated it as a tool of its own.

And it seems worth calling it one. If a product is wired with enough telemetry — what a person clicked, where they stumbled, what they paid for — and you hand all of that to a decent model together with the code that produces it, what you get isn't a dashboard with charts, it's an actual investigation. An answer to the question "why did this specific person leave". That, I think, is the real value — and it isn't about call transcription.

The blind spots haven't gone anywhere, though: users who turn on the local model in the browser are almost invisible to analytics — those sessions have to be pieced together from fragments. My telemetry only covers how features are used: what was clicked, where someone got stuck, what was paid for — the content of the calls never enters it. I only started seeing a visit's country a few days ago — before that, all I had was an IP hash. And it works only because I personally granted it access and wired it into my infrastructure — SSH, browser, GitHub. Selling this as a standalone product would be hard: every client would need their own configuration, and that scales badly.

On top of that, I have no infrastructure for using this in everyday life. I saw the case in Telegram in the morning on my way to work and could do nothing until the evening. I need a skill: I send a message into a mobile session, it investigates on its own, proposes the email on its own, I review and send. And along the way, automatically gift tokens to everyone who uninstalls after getting deep into the funnel — they're warm, they've already understood everything, a random error cut them off.

There is exactly one way to check this: the user who came back has to become a paying one. Then direct communication with two or three warm leads a week pays for itself, and the analyst agent that finds and explains them for me deserves to be called a product of its own. If he doesn't convert — then all I have is a convenient way to explain my failures to myself nicely, in hindsight.

For now, I'm sitting down to fix the error messages — the very small thing that cost me the person who made it all the way to the end.