I Was the AI (and My MVP Failed)

Our AI medical scribe - the product that now turns thousands of consultations a day into structured records in about thirty seconds - began as me, pretending to be an AI. And the first version failed completely.

Both halves of that sentence matter. Here’s the whole story.

The setup

We’d spent a decade building Doutore, the system doctors type their records into. Ten years of watching the most highly trained people in any building spend a third of their day as data-entry clerks. If you’d asked me what our customers’ single biggest burden was, I wouldn’t have needed to think: typing.

Then OpenAI released Whisper, and for the first time the burden looked optional. I ran it on my own machine. It took several minutes to transcribe a single consultation audio. Too slow to be a product. More than fast enough to be an experiment.

The wizard of Oz

I didn’t build anything. Instead, I hand-picked a few doctors from our customer base and made them an offer: record your consultations, send me the audio, and within 24 hours you’ll receive a well-structured medical record. Better organized than anything you’d write yourself in the two minutes you have between patients.

Behind the curtain, the “product” was me. Whisper grinding away on my laptop, then me structuring the output into a proper record by hand, every evening, for every consultation. In product circles this is called a Wizard-of-Oz MVP: the user sees a product; the founder is the product.

The records I delivered were good. Genuinely good - higher quality than anything those doctors had seen, and I knew what good looked like after ten years in medical records.

The failure

They didn’t care.

Not “they complained about the delay.” Not “they asked for changes.” They were simply, politely, uninterested - in a free service, hand-crafted by a founder, delivering the best documentation of their careers. That stung. Then it taught me the most valuable product lesson I’ve ever learned.

A medical record that arrives after the patient has left the room might as well not exist. The consultation is over; the doctor’s mind has moved to the next patient; whatever they were going to remember, they’ve already half-forgotten or scribbled somewhere. The 24-hour record wasn’t a slower version of the product. It was a different, useless product.

I had validated quality. Quality was never the question. The product was speed. The record has to exist before the patient stands up - or it doesn’t count.

The rebuild

So we engineered for speed with the single-mindedness of people who had just been told the truth. Every stage of the pipeline - capture, upload, transcription, structuring, delivery - attacked for latency. The goal wasn’t “faster.” The goal was a number with physical meaning: a full consultation should become a finished, structured record in about thirty seconds, because thirty seconds is what the end of a consultation gives you - the patient gathering their things, the doctor’s hand reaching for the next file.

Same underlying model anyone can rent. Same quality bar I’d held by hand. Completely different product.

The addiction

After it got fast, the reaction changed in kind, not degree. Doctors didn’t “adopt” it - they latched onto it. The same population that had shrugged at my hand-crafted 24-hour records now couldn’t go back to typing.

One doctor put it in words I’ll never improve on. He told me - his framing, not mine - that we were handing out free cocaine: one taste and there was no way back to the old life.

What everyone building with AI gets backwards

I see teams everywhere making my 2023 mistake in 2026. They obsess over output quality - better prompts, bigger models, eval suites - while their product takes ninety seconds to do something the user needed in ten. They’re hand-delivering beautiful records 24 hours late.

Latency isn’t a performance metric. In AI products, latency decides what your product is. A 30-second scribe is a different species from a 24-hour scribe, the way a conversation is a different species from a letter.

Quality got us politely ignored. Speed got us addiction. If your AI product isn’t landing, check your latency before your prompts.