AI meeting notes that are actually usable

You're sitting in a call, listening and typing at the same time. At the end you have three keywords you no longer understand two days later, and the actual commitment from the client got lost somewhere. This is exactly where AI helps, but only if you set it up right. A raw transcript isn't a note, it's just more text to read through. In this playbook you'll build a workflow in ten steps that produces three things you'll actually use: a short summary, a list of tasks, and the decisions with names attached.

I've been doing it this way for months for every client call and would never go back. One thing up front: the moment you record a conversation, people are involved, and that has legal sides. There's a dedicated step for it further down, don't skip it.

Step 1, decide what you actually need

Before you sign up for any tool, get clear about what should come out the other end. Most people don't want the full word-for-word protocol, nobody reads that anyway. As a rule you want three things: a summary in five to ten sentences, a list of open tasks with the person responsible, and the decisions that were made. Once you know that's what you need, the tool choice gets easy, because you test every tool against exactly those three outputs.

Jot it down on a piece of paper. Sounds trivial, but it saves you the hour where you get lost in features you'll never use.

Step 2, pick the right tool

There are two routes. Either a specialized meeting tool that dials into your call and transcribes automatically, or you use what's already built into your video software. Specialized services are, for example, Otter.ai, Fireflies.ai, tl;dv or Fathom. They join Zoom, Google Meet or Teams as a participant and hand you a transcript plus summary afterwards. If, on the other hand, you're already using Microsoft 365 or Google Workspace, Teams and Google Meet now have built-in AI notes, and for many cases those are perfectly enough.

My advice: start with what you already have. Only reach for an extra tool once the built-in version feels too thin. With everything, pay attention to where the data lives, more on that shortly.

Step 3, sort out consent to record before you start

This is the step most people skip and the one that can cause you real trouble. In Germany and the EU you're not allowed to simply record a conversation secretly or have an AI transcribe it. The others on the call have to know about it and agree. In practice that means: at the start you say a sentence like "I'm having an AI assistant transcribe this meeting, is that okay with everyone?" and you wait for the answer. If someone says no, you turn it off and take classic notes.

Many tools display a visible notice for everyone that a bot is transcribing. But don't rely on that alone, say it actively. With sensitive topics, such as personnel or health, be cautious about recording as a matter of principle. How this connects to GDPR and AI tools is covered in more detail in the playbook DACH Legal, the EU AI Act and GDPR.

Step 4, prepare the recording properly

An AI transcript is only as good as the audio. Make sure everyone speaks reasonably clearly and that you don't have constant noise in the background. A headset almost always beats the built-in laptop mic. If several people are sitting at one computer in the same room, the AI will struggle to tell them apart, so it's worth having everyone join the call on their own device. It sounds like a small thing, but it makes the difference between a note that's accurate and one that's half guessed.

Step 5, run the conversation with a structure

The AI transcribes better when the conversation itself has a structure. If at the end you take two minutes and say out loud "okay, the decision is X, the open task for Anna is Y, I'll take care of Z by Friday", then every tool picks that up reliably and puts it into the summary correctly. So you help the machine by summarizing cleanly as a human. Side effect: the people on the call also end up on the same page, which is often worth more than the note itself.

Step 6, turn the transcript into a real summary

Many tools generate a summary automatically, but it's often generic. This is the part where a chat model like ChatGPT, Claude or Gemini makes the difference. You take the transcript and give it a clear instruction. For example:

Here is the transcript of a client call. Give me three blocks:
1. Summary in a maximum of 8 sentences.
2. Open tasks as a list, each with the person responsible and a deadline if mentioned.
3. Decisions made, short and clear.
Don't make anything up. If something is unclear, write "unclear".

The last sentence is the most important. Without it, the model fills gaps with plausible-sounding nonsense. Why that happens and how to spot it is covered in the lesson When the AI hallucinates.

Step 7, always check against the transcript

Never trust the summary blindly. The most common source of error is a commitment or a number that the model attributes incorrectly. "We'll review a budget of 5,000" quickly becomes "Budget confirmed, 5,000". That's a different sentence with different consequences. Read the summary through once and cross-check the critical points, meaning money, deadlines and commitments, against the transcript. For a half-hour call that's two minutes, and it protects you from embarrassing follow-up questions.

Step 8, get the tasks to where you'll see them

A task list in a notes tool you never open is wasted time. The point is to get the tasks into your normal system. If you work with a tool like Notion, Trello or a to-do app, copy the tasks straight there, ideally right after the call while it's fresh. Some meeting tools offer ready-made connections for this, but even copy-paste into your task tool beats a perfectly formatted list gathering dust in the meeting tool.

Step 9, watch where the data is stored

With every one of these tools, the content of your conversations ends up on a third party's server. For an internal brainstorm that's usually not a concern, for a conversation about client figures or personnel it gets sensitive. When choosing a tool, look at three things: where the servers are, EU is preferable. Whether you can delete recordings, and whether they're actually deleted. And whether your data is used for training, which for business conversations you'll usually want to switch off. Most reputable providers have a setting for this, you just have to find it and set it once.

Step 10, turn the workflow into a habit

The whole effort only pays off once it runs automatically. Set yourself a fixed chain: get consent before the call, generate the summary with your tested prompt after the call, tasks into the system, delete the recording again for sensitive topics. Once you've done that three or four times, it sticks. From then on you easily save twenty minutes of follow-up work per call, and more importantly, nothing slips through the cracks anymore.

What's next

Once you start using AI for meetings, you quickly notice that the same pattern works for other everyday tasks, emails, proposals, research. A good overview for that is in the lesson AI in everyday life. And if you want to know which chat model is best suited for your summaries, Models compared will help you further.

Sources

• The EU AI Act and GDPR in the DACH context, our playbook: https://studiomeyer.academy/playbooks/dach-legal-eu-ai-act-und-dsgvo
• Basics on consent for recordings, the German Federal Commissioner for Data Protection (BfDI): https://www.bfdi.bund.de