Plans is an AI itinerary tool I've been building for travel agents. It started as a weekend project about sixteen months ago and a small group of agents now use it to put client itineraries together. Here's what I've picked up shipping it.
The problem it started from was small and specific. A travel agent I know was spending Sunday afternoons assembling day-by-day itineraries from scratch, flight times and hotel descriptions and transfers and restaurant notes, for clients she'd already had a discovery call with. All the information already existed. What didn't exist was the shape of it, and that turned out to be something AI is unusually good at.
The first version was one prompt over a Google Doc. Paste in a brief, get an itinerary back. It worked badly enough that I was embarrassed by it and well enough that two of her colleagues started using it within a month. That's roughly the bar I keep watching for.
Almost none of the work since then has been about the model. It picks the right hotel most of the time, the right restaurant some of the time, and the right tone nearly always. Everything else gets edited by an agent who knows the destination far better than the model ever will, and that edit step is where most of the design work has gone.
The main thing I got wrong at the start was thinking of the agent as someone reviewing AI output. They're not. They're the author and the AI writes their first draft. Once I understood that, most of the interface decisions followed from it. The diff view, the inline edits, the saved variants, the tone settings, all of it treats the person as the writer.
The biggest actual mistake was a multi-step agent I built early on, meant to chain flight lookup into hotel matching into a daily plan. It demoed beautifully and broke constantly in real use. Every step inherited the errors of the step before it and explaining what went wrong took longer than doing the task by hand would have. I pulled it apart into three separate single-shot calls and the whole thing became reliable more or less overnight. I now bring that lesson to every AI build I take on.
The numbers I look at each week are how many agents are using it daily, how many itineraries got generated, how many edits each one needed, and how long it takes to go from a brief to something sent to the client. The last one is the number that tells me whether it's actually useful, because that's the one agents care about.
The reason any of this matters for client work is that everything I've learned building Plans goes into the builds I take on for other people. I've already made the mistakes on my own product, which feels like the right order to do it in.
Plans is the long bet. Tee Sheet and Whip are quietly in progress behind it. If next year goes anything like this one, the writing here should get more specific and less abstract.