[ transcription × sales ]

Notes without a recording.

The real business in sales happens in conversation: at the customer's site, in meetings, on the phone. The documentation afterwards eats up the evenings, and recording would be convenient but is a data-privacy risk. This project reference solved both: conversations are transcribed live in the browser, the audio is never stored, and an AI agent turns the transcript into a visit report, tasks, and CRM entries. The system is in daily use by a Head of Sales. This page shows how it is built, which decisions stand behind it, and what was learned along the way.

No recording

The audio is transcribed live and the audio track discarded immediately. Only text is stored, never the sound.

Speakers separated

The transcript knows its speakers: anonymous markers become names with one click, and the conversation stays readable as a dialogue.

From talk to task

One click sends the transcript to an AI agent. Minutes later, the visit report, tasks, and CRM entry are ready.

Consent built in

Every recording starts with a mandatory consent, stored with a timestamp. And deletion is final, no recycle bin.

[ sales without paperwork ]

The best note is the one nobody has to write.

Anyone in sales spends the day in conversations and the evening documenting them: writing visit reports, noting tasks, maintaining the CRM. Much of it stays undone, and with every day of distance, details are lost that the next conversation would have depended on.

Recording would be the convenient way out, but a recording is a data-privacy risk and changes conversations. The better way: write it down instead of recording it. Transcription runs live, the audio is discarded in the same moment, processing is data-privacy compliant, and consent comes before every recording. What remains is only what a note would have contained anyway: text.

The result: sales holds conversations instead of writing minutes. Minutes after the meeting, the report, tasks, and CRM entry are in place, nothing is lost, and the next conversation starts with full context.

[ architecture ]

Without paperwork: from conversation to result

For the user, it is three steps: obtain consent, hand the transcript to the agent, approve the follow-up.

Audio path

From the microphone straight to text

No audio is stored: the sound is transcribed live and discarded in the same moment; it never touches our own backend. Nothing but text remains.

Provisional text appears immediately, final segments roughly every two seconds, speaker separation in real time, six languages to choose from. No bot has to dial into the meeting: the notes run quietly in the browser, whether on-site, video call, or phone call.

Deliberately not used is the browser's built-in dictation function: it processes speech outside the EU and would therefore not be GDPR-compliant.

Transcript & handover

Only text is stored, the user stays in control

The transcript is editable Markdown with speaker markup, strictly isolated per user and covered by tests. Deletion is hard and final, deliberately without a recycle bin.

The handover is push, not pull: the user sends the transcript actively; the agent fetches nothing on its own. And it receives the full original, no pre-summary. An hour of conversation is roughly 12,000 to 15,000 tokens, about seven percent of the context window.

Traceability without raw data: consent is stored with a timestamp, and accesses are logged, with a hashed IP instead of plain text.

Agent & CRM

From transcript to report, tasks, and CRM entry

The agent is defined through a system instruction: it determines what the visit report looks like, which tasks with owners and follow-ups are derived, and whether a follow-up email draft is produced right away. The business team adapts that itself, without code.

The CRM is connected via OAuth, restricted server-side to the approved tools. HubSpot is in use; the path works the same for every connected system.

Writing actions, creating a note, creating a task, are recognized as mutations and land in the approval inbox. Only the confirmation executes: the human has the last word before anything reaches the customer record.

[ lessons ]

The pitfalls sit in the browser.

The most instructive bugs were invisible: the browser's echo cancellation filtered the speaker out of their own signal, and silence arrived at the transcription service. And browsers occasionally ignore the requested sample rate, turning speech into noise. Anyone processing audio live has to check every browser detail instead of trusting the defaults.

The obvious shortcut would have been the browser's built-in dictation function. It failed the test: processing runs outside the EU and is therefore not GDPR-compliant. The lesson: data privacy works as architecture, not as a footnote. Part of that is the honest limit that no tool replaces informing the participants. That is why consent is mandatory, before every single recording, with a stored timestamp.

[ numbers ]

The system in numbers

audio files are stored: the sound is transcribed live and discarded immediately

100%

of the transcripts belong to the user: only they see them, only they edit them, and deletion is final

languages are currently available, extensible as required, with speaker separation in real time

~2 seconds

is how far the final text trails the conversation: written down while people are still talking

[ technology ]

Technology in use

HubSpot

AG2 / AutoGen

Django

Next.js

React

HubSpot

AG2 / AutoGen

Django

Next.js

React

TypeScript

PostgreSQL

Redis

Docker

TypeScript

PostgreSQL

Redis

Docker

Put the potential of this technology to work.

Do you also spend your days in conversations and meetings, and want technology that supports you in the follow-up? Then let's talk: you bring the challenge, I bring the experience and the tools.

Let's get started