Every other meeting AI product starts from the premise that you should stop listening and let software handle it. Granola starts from the premise that you were going to take notes anyway and your notes are the most valuable signal in the room, because they encode what you thought mattered. It records the audio, you type sparse fragments during the call, and afterwards it merges the two into something that follows your priorities rather than a generic summary of everything said. That is a genuinely better idea and it is the reason this product is worth writing about.
The design consequence people notice first is that no bot joins the meeting. Nobody sees a participant called Notetaker appear and has to decide whether to say something about it. That removes a real social cost, and anyone who has watched a client visibly change their tone when a recording bot arrives will understand why it matters. The help material is short, plainly written and does not pretend there is a methodology to learn.
Correct call. The second order effect is the one I find most interesting. Because your typed fragments steer the output, the tool rewards paying attention rather than punishing it. Most meeting AI quietly encourages disengagement, on the promise that the summary will catch anything important, and the result is meetings full of people who were not really there.
Granola inverts that, and the material is right to lean on it. The flip side is that if you type nothing, you get a fairly ordinary summary and the whole advantage evaporates, and the guidance could be much more forceful that the tool is only as good as your engagement with it. Consent is where I would want considerably more. This records conversations, in jurisdictions with different rules, in meetings with clients, candidates and colleagues who did not necessarily agree to anything.
The absence of a visible bot makes disclosure less automatic, which is a benefit socially and a risk ethically, and the material does not sit with that tension. Telling people plainly that you are recording is the practice, and a help centre for a tool like this should say so at the front rather than leave it to the user's conscience. Retrieval is the weaker half of the product. Capture is excellent and finding a thing you half remember from a call in March is more effort than it should be.
Since the value of a note archive compounds with its searchability, this is where I would want the next year of work to go, and the material has little to offer because there is not much to describe yet. Coverage of setups outside the original Mac focus has improved and still feels secondary. Three point nine. The rare AI product built on a correct insight about human behaviour rather than a capability demonstration, with documentation appropriately brief, marked down only for treating the consent question as somebody else's job.