All notes

· 3 min read

I built the clever half first

The AI feature I spent the longest on had been opened by six people out of a hundred. It needed a full portfolio to be any use, and almost nobody had got that far, because getting that far meant typing.

The number that started this

In October I sat down and actually counted what people did in a SaaS I run, instead of going off the handful of friends who use it and tell me things.

With my own accounts excluded there were 107 of them. 68 had created any data at all. 45 had built up anything I'd call a real portfolio, meaning ten items or more. 23 had got as far as owning a property and parking a vehicle in it, which is the point where the product starts being worth anything.

And 88 of the 107 were only ever active on the day they signed up.

The feature I'd spent the longest on, the one with a language model in it, had been opened by six people.

Why it was six

The feature takes a request in plain English, works out a plan for which vehicle goes in which bay, and applies it. I'm still happy with how it's built. The model only produces an instruction and ordinary deterministic code produces the plan, so it can misunderstand you but it can't propose something that doesn't fit. It took weeks.

But think about who it's for. Arranging your garages is only a problem once you own enough things for the arrangement to matter. Below that there's nothing to organise, and you'd do it by hand in less time than it takes to type a sentence.

So the feature needed people to be at a state that the funnel above says almost nobody reached. 45 out of 107 had ten items. The thing standing between a new account and a full portfolio was a few hundred vehicles to find and tick off, one at a time, in a session that gave you nothing back until you were most of the way through it. That's the part I'd left alone, because it was a list of checkboxes and a search box, and I'd already built it, and it worked.

It worked in the sense that it did what it said. It didn't work in the sense that people stopped there and didn't come back.

What I did about it

I built an import. You paste your vehicles in as a list in whatever form you already have it, a message out of a group chat, a column from a spreadsheet, something you typed, and they get matched to the catalogue and parked for you. That went out at the start of November and it's what the product leads with now.

It's also where the AI budget went. The matching is the hard part, because people write "the new elegy" and mean one specific vehicle, so that's the bit worth a model call. Arranging things afterwards was never the hard part. I'd just assumed it was, because it was the interesting problem.

What I actually got wrong

Not the build. The organiser is good, and it's still there, and I didn't delete it. It costs me almost nothing to keep and the people who use it like it.

What I got wrong was the order. I picked the feature that was interesting to build over the feature that was in the way, and then I didn't check for months which one that was. The six was always there to be counted. I could have run that query in June.

So now I count the funnel before I pick what to build next, and I look for the step where people stop rather than the step I'd enjoy. It's a boring habit and it would have saved me a few weeks.

Published product · measurement · ai