Pitch Brief

The two-page memo a reader takes into a first meeting: the problem, the mechanism, the evidence and what happens next.

Someone walked out of a meeting this morning with a 159 US dollar recording card stuck to the back of their phone. Plaud reported more than 1.5 million sold by January 2026 across its line. The phone under it already has the battery and the data plan, and Recapp does the recorder's job as an app: tap start, tap stop, and the transcript, the summary and the action items are waiting before you reach your desk. The company is in the part nobody advertises: the phone can stop recording mid-meeting and nobody has measured how often.

Start with what is not true. Voicenotes does this exact interaction on a free tier; Fathom and Granola give the recording and the transcript away, and Granola is a defendant in one of four pending United States class actions against note-takers; Apple's Notes and Google's Pixel Recorder ship it pre-installed, though neither is documented as shipping action items as a separate list, and Pixel Recorder is Pixel-only. The app version exists, much of it free, the hardware sold anyway, and the reason to pay us is not a missing feature: they lose meetings.

We asked why. The only primary evidence is 64 stated purchase reasons from retailer reviews and owner forums; the largest block, 28.1 percent, is capturing a cellular phone call, which no third-party app may do, and on that source's own capability matrix not one of the 64 is a job an app supplies outright. The counter is selection: those forums record why buyers stayed, not what someone still choosing would do. We no longer say that removing the device wins that customer. The buyer we would go after is the one already recording on an app and losing meetings.

The failure modes are documented and the rate is not. Accepting a phone call suspends the app, and a backgrounded app that tries to restart capture is refused by the privacy subsystem: the meeting carries on, the recording does not, and nobody notices until the file is opened. On Android the adversary is the handset maker, not the operating system: sleeping-app tiers, memory eviction and freeze policies end a correctly declared recording service, behind settings paths an update reverts. One of our own lines died here: neither platform releases the raw per-microphone channels to any app, so we claim nothing about the phone's acoustics against the device.

What we would build

You tap start, say out loud that the room is being recorded, and tap stop. Announcing is the default, not a regional mode: one out-of-state participant pulls the meeting under stricter law. The category is being sued over quiet capture, and local processing is no shield. The app checks second by second that audio is arriving, writes to storage in short increments so a kill costs seconds not the meeting, and arms a notification that every healthy heartbeat cancels: a frozen process warns nobody, so the alarm fires when cancelling stops. Restarting is a third tap. Recognition runs during the meeting; the number we would publish is the interval from tap stop to all three outputs for a one-hour meeting, on the phone. Action items need owners, and neither platform gives a third-party app speaker labels, so speaker naming is undecided.

Why now:

• Both platforms now put speech recognition and summarisation on the phone for third-party apps: Apple in iOS 26, Google in alpha, basic mode on Android 12 and above, generative mode gated to Pixel 10 and 11. Apple documents a 4,096-token session limit covering prompt, transcript and output, against 11,000 to 13,000 for an hour of speech, and Google publishes no equivalent figure, so a meeting is summarised in chunks, never in one pass.

There is no technical moat in a wrapper over rented models and we will not claim one. What we would own is the number nobody has: capture survival across handsets, skins and operating-system versions, a publication rather than a patent, re-earned every release. The threat is not a platform shipping a warning like ours, it is a platform fixing the kill, which deletes the product.

Our cost follows the architecture, so we name it: capture and recognition on the phone, summary on the phone inside the token window and in the cloud past it. That is near 0.05 dollars a recorded hour against 0.21 to 0.62 for the all-cloud pipeline the category runs, under a 15 to 30 percent store commission at an unsettled rate, and it is paid in battery: 10 to 15 percent an hour for recognition, 25 to 35 for a full offline pipeline, near nothing for recording. The 159 dollars buys the object and 300 transcription minutes a month; the bill outlives the hardware. Distribution is the open hole: the device is cash-positive at the point of sale and pays to find the customer, which is why the company shipping it raised under 6 million. The position nobody holds is the first run: a multi-step setup on every rival, a per-brand settings walk on Android.

Nothing is decided until the first test returns. On shipping recorder apps, across handset models and skins rather than people, force an incoming call, a sleeping-app tier, a memory reclaim and a competing microphone request, and count the complete recordings. If the incumbents pass cleanly there is nothing to sell here and we will say so. Our instrumented build follows and logs metadata only: timestamps, buffer counts, interruptions and watchdog firings, audio never leaving the handset. Completion alone is not a bar, since a session can run to the end and still be unreadable, so a word error rate sits beside it, against the 22.2 percent the best single-microphone system reaches on a far-field meeting benchmark. One gate comes first: App Store review guideline 2.5.4 makes a background-audio declaration with no visible feature a rejection risk, which the in-room warning is meant to answer.

Nothing here works unless the phone keeps recording to the end of the meeting, and nobody can tell you how often it does.