Pocket hits 300K devices sold — founder Akshay Narisetti on why dedicated AI hardware beats phone apps

Pocket: 300K devices, $129, and a bet that dedicated hardware beats your phone

Akshay Narisetti built Pocket backwards. The company launched a meeting notetaker app first — the obvious move, the easy path to scale — and saw almost no traction. Then Narisetti and his team handed people a physical recording device and watched usage jump tenfold. That gap between app and hardware is the entire business thesis.

Pocket is a clip-on voice recorder that retails for $129, paired with a freemium subscription. The base tier includes unlimited summaries and transcripts, with three AI queries per day. The $20/month pro plan unlocks speaker tagging, advanced AI reasoning, and a library of 300-plus professionally sourced templates — SOAP notes for doctors, case notes for lawyers, CRM entries for salespeople. The company has sold 300,000 devices to date, a figure Narisetti corrected upward mid-conversation from the 200,000 cited in the show introduction.

“It's 300,000 now. We have the highest margins in industry for AI — roughly 75% margin in software subscriptions. We use our own version of OpenRouter where we essentially route to different kinds of providers and models... Y Combinator actually currently uses Pocket to record all their interviews.”

Why hardware over an app

The case for a separate device is less philosophical than practical. Recording on a phone has real friction: you can't record a Zoom call while you're on the same phone, incoming calls interrupt active recordings, and pulling out a phone to start recording in a client meeting reads as awkward in the first eight seconds. The people who most need this product — consultants, field salespeople, real estate agents, doctors moving patient to patient — take ten to fifteen back-to-back meetings a day and can't afford that friction.

Narisetti says one early user bought a second phone purely for recording and strapped an external battery to it. Pocket is essentially that workaround turned into a product.

The failed hardware wave — Rabbit, Humane, Friend — tried to replace the smartphone. Pocket does one thing and fits into an existing workflow. Narisetti argues that's why those devices died and Pocket didn't: "convenience beats intentions."

Customer mix and use cases

DoorDash uses Pocket to let salespeople record restaurant pitches in the field, then sync notes back to a shared knowledge base. Y Combinator records all its founder interviews through Pocket's API. The core buyer is a field worker — anyone whose most important conversations happen away from a desk.

Narisetti makes a broader point about AI context: every conversation recorded through Pocket can be piped via MCP or OpenAI's API directly into ChatGPT or other tools, giving the model a running record of the user's professional life. A VC meeting that sounded like a soft yes can be parsed later for what was and wasn't actually said.

Margins and model routing

Pocket claims the highest gross margins in its category for AI software, around 75%, achieved through its own model routing layer — effectively a private version of OpenRouter. Transcription runs on fine-tuned versions of OpenAI's Whisper, adapted for offline environments. The public Whisper models were trained on clean YouTube-style audio; they degrade badly when a train passes or ambient noise spikes. Pocket's fine-tuned versions handle that.

Users choose their own summarization model, which Narisetti says is a meaningful product feature rather than a technical footnote. That optionality also gives Pocket flexibility to shift workloads based on cost and quality without changing the user experience.

What's coming

Within three months, Narisetti plans to ship agent functionality — workflows that take conversation context and automatically generate reports, decks, and documents. The pitch to consultants and salespeople is that the post-meeting synthesis step, currently manual, becomes automatic. That's a straightforward extension of what the device already captures; the execution risk is whether the outputs are good enough to replace the work rather than just approximate it.