Meeting Assistant

A meeting assistant is a program that listens in on a meeting and takes notes. It automatically converts the spoken word into text. From this text, it then creates a short summary. It usually also lists who took on which task. In video conferences, such an assistant often appears as an additional participant in the list. It thereby replaces the person who used to write the minutes.

Why minutes suddenly cost no one any time

Meetings are expensive. If eight people sit together for an hour, eight work-hours are consumed. A clean set of minutes used to cost another half hour on top. That is exactly why, in many companies, it simply wasn’t written. Results got lost, and afterward no one knew exactly what had been decided.

A meeting assistant reduces this effort to almost zero. The summary is often already in the inbox a minute after the conversation. Anyone who wasn’t there can read up in two minutes instead of watching an hour of recording. For large companies, this is the reason such features are showing up in almost every office software package.

But there is also a downside. When everything is being noted down, a permanent record of everything employees say is created. In Germany, such tools are therefore usually only allowed to run with the consent of everyone involved. Often the works council must also give its approval. Otherwise, a spontaneous, careless remark ends up in a file forever.

From microphone to task list

The first step is called speech recognition: a model converts the audio track into written words. These models were trained on many thousands of hours of audio for which the corresponding text was known. With clear pronunciation, they achieve an accuracy rate well above ninety percent. With strong dialects, background noise, or people talking over each other, it drops noticeably.

Afterward, the program tries to distinguish between speakers. It compares voice characteristics and assigns each sentence to a person. This produces a text that shows who said what. In video conferences, this is easier because the system knows which microphone was active at any given moment.

In the final step, a language model is presented with the finished text. A language model is a program that processes text and generates text itself. It is meant to shorten the conversation and extract tasks. This is exactly where the typical errors occur: the model occasionally invents decisions that were never actually made. Experts call this hallucination. A summary should therefore be briefly skimmed before being forwarded.

Built into Teams, Zoom, and Meet

Meeting assistants are most commonly found in video conferencing programs. Microsoft offers the feature in Teams under the name Copilot, Google in Meet, and Zoom has its own Companion. Alongside these, there are specialized providers like Otter or Fireflies, which dial into other people’s conferences as a guest. They are usually recognizable by a participant that consists of nothing but a name and a recording icon.

The technology is also widespread outside of video conferences. Doctors have patient conversations transcribed to shorten documentation. Sales teams log customer phone calls and analyze them later. Some smartphones record calls and summarize them directly on the device.

The term comes up in business news for two reasons. On the one hand, the meeting assistant is considered an example of an AI use case that can actually generate revenue. On the other hand, companies and data protection authorities are arguing over how long the transcripts may be stored. Anyone using such a tool themselves should therefore know where the data is kept.