Analyze the retention mechanics of an AI product with high initial excitement.

Larchnote Notes listens during a vet visit and writes the clinical note afterward, so the vet can just talk to the animal and the owner. Cassius Drammeh runs Oldfarrow Veterinary Clinic, five exam rooms, three vets, a waiting room that's loud by 9am.

How to answer this, stage by stage

Seven moves, and the fourth one is the whole answer. The rest is proof.

Let's learn

Larchnote Notes runs on a small mic clipped to a vet's coat. It listens through the exam, then writes the visit note by the time the vet reaches the next room.

Before Larchnote, Cassius typed his own notes between patients, roughly seven minutes a visit, twelve visits a day. That's about 84 minutes a day gone to typing, most of it after the owner had already left.

With Larchnote, that dropped to about six minutes a day, a quick skim of each note before signing it. For six straight weeks, Cassius told anyone who'd listen that it was the best thing that had happened to his practice in years.

The turn: the extra wrong notes were never really the problem. The real problem was what Cassius did the first time one showed up with no way to fix just that one line.

At its worst: three months in, Oldfarrow's clinic looks like it never adopted Larchnote at all, except now the vets type slower than they used to, out of practice, and the clinic is still paying for a license nobody opens.

What I would leave alone: the free mic and the ambient listening itself. Vets loved not fumbling with a phone or a keyboard mid-exam, and that part of the product never caused a single complaint.

The lesson: a burst of early excitement tells you the product feels good on a clean day. It tells you nothing about what a person does the first time it hands them a mess with no easy way to clean it up.

Now here is the same thing as a story

Use the short version above when you're being timed. Read the story below when you want to feel why the flip was real, not just plausible.

Cassius has run Oldfarrow for eleven years. He can read a limping dog's gait from across the room before the owner's finished describing it.

The good months were genuinely good. Every visit, the mic caught the whole exam, and by the time Cassius reached the next room, a clean note was already sitting there waiting for a quick nod.

Everyone assumed trust would fade like a dial turning down. It didn't. It clicked off.

The habit thinned in three beats nobody wrote down at the time. First, Cassius stopped reading every note word for word, just the first line. Then he stopped reading most of them at all, just signing. Then, one Tuesday, he stopped opening the app before a visit even ended.

Nobody filed a ticket at any of these four points. That's exactly why nobody noticed.

The trigger was small: a healthy dog's slight limp, described in the note as "likely a hairline fracture, recommend imaging." Cassius caught it before the owner saw it, since he remembered the exam. But the note offered no way to fix just that line. Edit meant rewrite the whole thing.

Only one of these four branches keeps a vet using the tool. Larchnote's design pointed straight down the other three.

He mentioned it at a regional vet meetup, half a joke, and two other clinic owners said the same thing had happened to them within a week of each other, unrelated clinics, same wrong branch of the tree.

The weekly usage digest Oldfarrow's clinic manager got never named any of these four things.

By week eight, Cassius was back to typing every note by hand, the same 84 minutes a day he'd had before Larchnote ever showed up, except now the app sat unopened on his phone and nobody at the company had any idea why.

The old note-taking flow asked Cassius to trust an entire finished note or rewrite an entire note. The new one, once fixed, asked him to trust most of a note and fix the one line that was wrong.

I built the correction flow around what was fast to ship, a single rewrite box, not around what a vet would actually do at 4pm with four more patients waiting. It took hearing the same story from two other clinics, unprompted, to see it wasn't Cassius being picky. It was the design, working exactly as built.

FLIPS, five letters for a habit that clicked instead of fadedNot a dial slowly turning down. FLIPS is what finds the exact click.

Five letters, and the third one, the flip itself, is the only genuinely hard step.

The recap, one line per letter: find the person is Cassius at Oldfarrow, locate the habit is the 84 minutes of typing he stopped doing, identify the flip is opens-every-visit versus never-opens-again, pinpoint the old decision is the all-or-nothing rewrite, and show the replay is a same-day fix that keeps the app open instead of abandoned.

And if you want to be sure it really works, try it somewhere elseSame five letters, a library instead of a clinic. A completely different flip family this time.

Kestrelwing is an AI tool that suggests subject headings and call numbers for new books at a small public library system. Amara Ffolkes heads cataloguing there, six branches, a backlog that never quite empties.

Mapped onto FLIPS, with a different flip family this time, substitution instead of abandonment: find the person is Amara. Locate the habit is trusting Kestrelwing's suggestion on every new title without a second look. Identify the flip: once the system added a small per-item review fee to fund a "priority queue," Amara didn't stop using Kestrelwing, she started saving it only for the easy, obvious titles and went back to cataloguing hard, ambiguous ones by hand, the exact cases Kestrelwing was built to help with most. Pinpoint the old decision: pricing per item, which made sense when volume was low and nobody rationed anything. Show the replay: flip the pricing to a flat monthly fee instead, and Amara runs every title through Kestrelwing again, hard ones included, since there's no per-use cost steering her away from the cases that need it.

A different flip family, but the same shape: a small pricing change, not a worse model, is what rationed her trust toward the wrong cases.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "no way to fix one wrong line, so vets stop opening it, add a one-tap flag routed to a same-day fix," and stop.
Cost: there's no budget this quarter for a same-day correction queue. Say so honestly, and start with a lighter version, a flag button that at least tells the team which notes are wrong, even without same-day fixes yet.
The model gets better, for real: if Larchnote's transcription accuracy genuinely improves, that's still not a reason to skip the repair affordance, since even a rare wrong note with no fix path can end a habit that took months to build.

Where people run it wrong.
They treat fading excitement as inevitable novelty wearing off, instead of looking for the actual moment usage clicked to zero.
They watch total notes generated instead of opens per vet, which hides a slow abandonment inside a flat-looking average.
They fix accuracy first, assuming a better model will bring the habit back, when the real gap is that there was never a way to fix one line.

How to use it live. When someone describes excitement fading, ask yourself: what's the exact action the person stopped doing, and is there a day you could point to. If you can't name a day, you don't have the flip yet, keep digging before you answer.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0