Almost every AI product has the same shape on the retention curve: a sharp novelty spike when people try the magic, then a cliff. The model is impressive, people poke at it, and they leave. Better model quality does not fix this. Habit design does. Here is how I think about building loops into AI products, drawn from getting Listen2RE to a 71 percent session-completion rate.
Novelty is not retention
The first time an AI product does something surprising, users feel delight. Delight does not repeat. The second time, the magic is expected; the third time, it is infrastructure. If your retention plan is "the AI is amazing," your retention plan is a decay function. You have to convert the novelty spike into a loop before it fades.
The loop, applied to AI
The classic loop is trigger, action, reward, investment. For an AI product:
- Trigger. What brings the user back without you nagging? The strongest triggers attach to something already in the user's day. Listen2RE's trigger was a session delivered at 6:30 AM, anchored to the existing commute. We did not invent a new habit; we attached to one that existed.
- Action. The smallest thing the user does to get value. Make it nearly free. For us: tap play. Not "configure your learning plan" — tap play.
- Reward. Value delivered, ideally with a little variability. The session itself, plus a streak that acknowledges the return.
- Investment. The user puts something in that makes tomorrow better — a rating, a topic override, progress accrued. Investment is what makes the next loop stronger than the last.
Consistency beats choice
The biggest habit lever on Listen2RE was counterintuitive: we chose a single daily push over an on-demand library. Less user control, by design. For habit formation, the paradox of choice is real — a library creates decision fatigue and skipped sessions, while one well-timed session creates a routine. Completion hit 71 percent. The feature that would have given users "more" would have given us less retention.
Variable reward, used carefully
A little unpredictability in the reward strengthens the loop — but with AI products there is a trap. Variability in quality is not delightful, it is corrosive. The variability that works is in content (a different topic, a fresh angle), not in whether the output is good. Your quality floor has to be rock solid before you add any variability on top. This is exactly why the eval harness and the habit loop are connected: inconsistent quality destroys the trust the loop depends on.
Measure the loop, not the magic
The metrics that matter for an AI habit product are loop metrics: completed actions per user per week, return rate, streak length. Not "time in app" — for a commute product, less time for the same progress is better. Not downloads. The north star should be the loop firing, because that is the thing that compounds.
The takeaway
The model gets users in the door once. The loop is what brings them back. Spend your novelty budget buying the first action, then design relentlessly for the second, third, and fourth. Retention in AI products is a design problem wearing a technology costume.