Free Users Burn Three Quarters of the Tokens

By Georg Malahov

I finally got back to Call Copilot. It hadn't been updated since March — it just sat there working on its own while I was busy with other things. And the first thing I did wasn't open the code; I sat down to go through the statistics on every install, ever. The timing was good, too: I'm not running or buying any ads, so everything in there is pure organic traffic, without distortions.

I had one very specific question: should I switch off the free mode entirely. Along the way I wanted to see the average call length, the activation stats, how people behave in general. And the answer came out worse than expected: seventy-five percent of all the tokens I'm burning are burned by free and anonymous users. Three quarters of the costs go to people who not only don't pay — they don't even sign up.

Then it got even better. A quarter of all installs are anonymous folks who are quite openly abusing me: they install the extension, burn through the limit, uninstall, immediately reinstall, and get another sixty minutes. And they do this quite actively, not as a one-off. In other words, there's a hole in my unit economics that I cut open myself with my own generosity.

The decision I made: the free plan gets switched off entirely. What stays is a demo mode — the first call, a demo video, you can poke around and roughly see how it all works, and then a hard paywall shows up. Real calls require registration. The trial gets cut down to two ten-minute calls. Sounds harsh, I know. But paying for three quarters of the tokens out of my own pocket sounds worse.

Over the weekend I closed most of the tasks. I did what I'd long wanted to do: a review prompt — people are already leaving me stars — collecting feedback after calls, fixed every bug I knew about, rewrote the paywall logic: what's behind it, when it shows, to whom. The only thing I didn't do was email sending — all the email communication is still just plans.

And here is my biggest failure. I got a lot done, but I never managed to properly test what came out of it and give the code real feedback. I started testing on Monday — everything sort of works, but it needs polishing: the copy on the site, the copy in the app, the logic has flaws in places. And I just can't carve out the time to sit down and walk through all the scenarios by hand, one after another. So the release is getting postponed. Not because it isn't written, but because it isn't verified. Doubly frustrating.

Out of that came an idea I want to try: give agents devops access, so that full E2E testing can be handed over end to end instead of doing QA in the evenings. The plans for these changes, by the way, I wrote together with Claude — and working from a ready-made plan went much better than my old approach of prompting task by task.

Whether all of this was for nothing, the next release will show — it's meant to be a big one: freemium with a paywall, emails, review asks. After it, the share of tokens burned by free and anonymous users should fall well below those seventy-five percent. If the picture hasn't changed in a couple of weeks, it means I didn't cut off the abuse — I just scared off normal people, and I'll have to roll it back.

And until the release is out — I'm off to test by hand, since I haven't automated that process yet.