The demand-generation tax

OpenAI ran the cleanest retention experiment in consumer AI, by accident. The numbers are from a16z’s consumer report. ChatGPT keeps about 36% of its monthly users active daily, and about half its desktop users are still there a year in. Sora - same company, same models, same distribution - kept fewer than 8% of its users past day 30. It isn’t the model and it isn’t reach. One product sits inside routines people already have. The other waits to be remembered.

That waiting has a price, and I think it deserves a name: the demand-generation tax. Before an open-ended AI product does anything, you have to remember it exists. Then frame your need as a prompt. Then decide it’s worth the detour. The tax is small per session and fatal in aggregate, because it’s charged every session, forever.

My n=1

I’ve been running the opposite design on myself. A cron job on a DigitalOcean droplet checks my Google Calendar every 30 minutes. When a meeting is an hour out, an agent looks up every attendee, digs through Gmail, Slack, and my Fathom call transcripts, and drops a briefing into iMessage. I never ask for it. I never prompt anything. The brief is just there, an hour before, in the channel I already read.

I walked into an investor meeting with a researched brief on everyone in the room that I hadn’t requested and couldn’t have built in the time I had. That’s when it stopped being a demo. Not because the research was brilliant - because it showed up without me generating the demand. The first version was noisier. It briefed me on recurring internal calls I could run in my sleep, and I nearly muted it before narrowing the trigger. An agent that pings you five times a day gets muted faster than a chatbot gets forgotten. This one survives because it rides an event that already exists - the meeting - and stays silent otherwise.

Chat does retain - people come back to ChatGPT because the answers keep getting better. But that’s a different mechanism. Model-driven retention goes to whoever has the best model this quarter. Routine-driven retention goes to whoever is wired into the user’s day. Only the second one is available to people who don’t own a frontier lab.

Here’s the part the switching-cost story usually gets wrong. The moat people cite is data. Mine would move in an afternoon - my agent’s memory is mostly markdown files, and I could carry them anywhere. What doesn’t move is the behavior. The trigger is tuned, the noise is pruned, the brief arrives at the hour I trust it to. Nobody wants to start fresh with an agent that has to relearn all of that.

Microsoft is seeing the same thing at enterprise scale. Nadella told investors in April that M365 Copilot’s weekly engagement now matches Outlook’s - Outlook, an app nobody decides to open. People just live in it. Copilot didn’t build a new habit. It moved into one that existed. And people do keep paying for chat - ChatGPT Plus holds around 71% of subscribers at six months, by Earnest’s card data. They pay for it. They just don’t build their day around it.

If I were building a consumer agent today I’d skip “ask me anything” entirely. Find one moment that already recurs in the user’s week. Show up just before it, carrying the thing they’d otherwise have had to assemble, and stay quiet the rest of the time. That’s the whole design. Mine fits in a cron job.

⁂