How to evaluate an AI note taker

Every tool in this category demonstrates well, and the reason is structural rather than dishonest. A demonstration is one clean meeting: known audio, three speakers taking turns, a subject the thing has seen a thousand versions of. Almost anything looks capable under those conditions. Your week is not those conditions.

So the useful move is not watching more demonstrations. It is arriving with questions specific enough that the answers differ, and then running one test you can actually mark. Both parts are below, and so is the part where the questions come back at us.

Start with where the recording comes from

This is the question people ask last and regret asking last, because it is the one that decides how much of your life the thing is present for. Everything about summaries and answers is downstream of whether the conversation was captured at all.

• Where does a recording start, and what happens to the conversations that never appear on a calendar?
• Does anything capture on your behalf when you are not there, and what do the other people see when it does?
• Can the person opposite tell it is running without being told, and can they stop it without an argument?
• What does it do when there is no meeting at all — a corridor, a car, a stand at a conference, a kitchen table?

None of those has a right answer. They have answers that suit a week shaped one way and not another, which is why the specific wording matters: you are not asking whether a tool is good, you are asking whether it will be there when you need it.

Then ask what comes back

The second group is about the output, and it is the group where a general question returns a general answer. Precision earns the most here.

• When it answers a question, can you get from the answer to the moment it came from in one move?
• Does it answer across everything you have, or one recording at a time?
• What does it produce from a four-minute conversation? A tool that returns a page of structure for something that had none is telling you what it does to everything else.
• When two conversations disagree, does the answer say so, or does it pick one and sound certain?
• Give it your worst twenty minutes — a bad line, three accents, two product names nobody outside your company has heard of. What comes back, and what does it do with the words it did not catch?

The first of those is the one to lead with. An answer that sounds right is not evidence of anything, and the whole difference between a tool you can rely on and a tool you have to double-check by hand is whether there is a route from a claim back to the sentence underneath it. Ona's own answer here is Ask Ona, and the part we would ask you to test is not the paragraph it writes but whether you can land on the transcript under it in a single move.

Then ask about leaving

Nobody asks these in month one and everybody wants them answered in year two. Three are enough. What stops working the day you stop paying — does the memory go read-only, or does it go? When you delete one recording, what else disappears with it, and how would you know? And who inside the company can open a recording of yours, under what circumstances, and is that a policy or a habit?

Judge the shape of the reply as much as its content. Precision is the signal. Somebody who can describe exactly what deletion removes, in two sentences, without going away to check, works at a company that knows its own mechanics. Somebody who answers a question about staff access by talking about encryption has changed the subject, and that is an answer too.

Which of these matters most is not the same for everybody, and trying to weigh all twelve equally is how people end up choosing on the strength of a landing page instead.

A test you can mark

Answers in an email are cheap. The fastest way to find out whether any of it is true is a test with a right answer, and the trick is to run it on conversations you already know the outcome of, so that every response is markable rather than merely plausible.

Our own answers, weak ones included

We are one of the things being evaluated here, which makes a buyer's guide an awkward document to publish. The only useful way to write one is to run the list over ourselves in public. Here are four of the twelve, answered: the capture question we come off worst on, and then all three about leaving, which is the section a guide like this poses and then tends not to sit still for.

Does anything capture on your behalf when you are not there? No, and it is the worst answer on this page. Nothing joins a call for you and nothing records unless you started it, so a conversation you forgot is not thinly captured, it is missing, and an hour that happened while you were somewhere else leaves nothing at all. The same answer covers whatever you are arriving with: there is no import route for an archive built elsewhere, so anything recorded with us begins on the day you begin.

Who inside the company can open a transcript of yours? Take the two halves separately, because they are different kinds of assurance. What you record is private to your account unless you share it, and inside a team workspace you decide per meeting what gets shared, so an admin sees what people have shared and not the rest. That half is how the product is built. The other half is policy: production access is held to the few engineers who need it, scoped to the task in front of them, and logged, with alerts on any access that looks unusual. Ona processes your conversations in the cloud, which is how it works today, and deletes the audio once it is transcribed, so what that adds up to is a short list of people and a set of rules — not a system in which we could not look.

When you delete one meeting, what else goes with it? The transcript and the summary leave your account straight away (there is no audio to remove, because Ona deletes it once it is transcribed), and the data is purged from our systems on a short, fixed schedule after that; deleting the account takes the lot, exports and settings included. How would you know? You would not, past the description, which is why the description is as specific as we can make it. And what stops working the day you stop paying? Not the archive. Dropping to a smaller plan limits how much new recording you can do; everything already recorded stays searchable, and you can take all of it out whenever you like. That is the strongest answer on this page, so it belongs next to the weakest: SOC 2 is under way rather than held, and work under way is not a certificate.

A guide from a seller that never produces a weak answer about the seller is a brochure with numbered paragraphs. Put the same four to anybody else you are weighing up, and listen for whether the awkward half arrives without having to be dug for.

One more thing worth deciding before any of this: whether the conversations you want back are work at all. A good deal of what people end up recording is not, and Ona for personal life and productivity is about that half of it.

Where this guide runs out

A list of questions is a filter and not a decision. Everything in this category can answer well in writing, and the answers only become facts at the point you check one of them, which is what the five conversations are for.

The thing you would most like to know is also the thing no test contains: whether you will still be pressing the button in eight months. That is a fact about you and your week rather than about any tool, and the honest way to find out is to notice how often you currently wish you had a conversation back. If the answer is twice a year, none of this is worth buying.

This guide is shaped by what we happen to think matters, which is not a neutral place to write from. Somebody who built a differently shaped tool would hand you a different list, and theirs would also be reasonable. And nothing on this page asks what it costs, which is not because cost does not matter — it is because price is the one thing in this category you can already read for yourself, on our pricing page and on everybody else's.