Nobody took notes in the meeting. So why does the recap already know who owes what?

Picture the end of a call that ran twenty minutes over. You were the one talking, so you took no notes. Then you open the meeting chat and there it is: a short summary, a list of decisions, and a row of tasks, each one with a name next to it. Nobody typed any of that. So how did it get there?

The short answer is that the meeting was being written down the whole time, just not by a person. In Microsoft Teams, a feature called Intelligent Meeting Recap does this work, and Copilot can answer questions about the result afterwards. Let's follow what actually happens, step by step.

First, the talking becomes text

Before anything can be summarised, the spoken words have to become written ones. This is the same job your phone does when you dictate a message: software listens to the sound of a voice and writes down what it hears. In a meeting, though, there is more than one voice, so the system also has to work out who said which sentence. That is why the recap can label a line with a person's name instead of leaving a wall of anonymous text.

This part is not automatic and it is not silent. Transcription only starts if someone turns it on, and everyone in the call sees a banner and hears a notice telling them it is running. If you would rather not be written down, that warning is your cue to say so.

Then the transcript gets read back and sorted

Once the call ends, a language model reads through that transcript. It is not replaying the audio; it is working with the text. Its job is to spot patterns: which stretches of conversation keep circling the same subject, where someone says a decision has been made, and where a sentence sounds like a promise to do something later.

From those patterns it builds the pieces you see in the recap:

• Topics discussed — the main subjects the conversation returned to, grouped so you can skim them.
• Decisions — moments where the group settled on something, pulled out as short statements.
• Tasks — sentences that sound like an assignment, usually paired with the name of whoever was speaking or was named.
• A timeline — markers showing when your own name came up, so you can jump straight to those moments.
• A search box — you can ask something like 'What did Mariana say about the delivery date?' and get the matching passage.

That last one is worth pausing on. The system is not remembering the meeting the way you do. It is searching the transcript it produced, and handing back the part that best matches your question. If the transcript is wrong, the answer will be wrong in the same way.

Where the names can land on the wrong task

Here is the part that matters most in practice. The recap decides who owns a task by looking at who was speaking, or who was named nearby. That guess is usually fine. But it leans on the audio being clear.

If the connection was choppy, if two people talked over each other, or if several side conversations ran at once, the system can attach a task to the wrong person. Someone who never agreed to anything ends up with their name on a to-do item. Nobody was careless; the sound simply did not give the software enough to work with.

So the recap is a strong first draft, not a verdict. Before you treat the task list as settled, read it once and check that each name really belongs to that job. It takes a minute and saves a lot of quiet confusion later.

What it is good for, and what it is not

The real gain is not that it writes beautifully. It is that it removes the job of scribbling while also trying to think. You can stay in the conversation, and the notes assemble themselves in the background. For anyone who has left a meeting unable to remember the three things they agreed to, that is a genuine relief.

What it cannot do is understand intent the way a colleague does. It does not know that a joke was a joke, that a suggestion was rejected, or that a firm 'maybe' meant no. It reads words and patterns. The meaning, and the final call on what was actually agreed, still belongs to the people in the room.

Similar recaps now exist in other meeting platforms, and standalone transcription apps offer their own versions. The underlying idea is shared: capture the speech, read the text, hand back a summary. The differences lie in how well each one handles messy audio and crowded rooms.

So the next time a recap appears before you have even closed the window, you will know what happened. A machine listened, wrote, and then read its own notes back to you. Useful — as long as you glance at the task list before trusting it.