Try the experiment. Open any general-purpose chatbot and ask it to write a performance review for a software engineer who is technically strong but struggles to bring others along. You will get something back in seconds, and it will be good: fluent, balanced, the right shape, the tactful sandwich of strength and growth area. It will also be completely untethered from the actual person, because the model has never met them. It is writing the average of every management blog ever indexed, dressed up to sound like it knows your team.
This is the part worth sitting with. The output is not wrong, exactly. It is plausible, which is more dangerous. A wrong answer you can catch. A plausible one slides past, and you find yourself nodding along to sentences about a colleague that no one who has worked with them would recognize.
The model is not the asset
There is a quiet assumption baked into most excitement about AI at work: that the intelligence lives in the model, and a better model means better answers about anything you point it at. For most knowledge tasks that holds up well enough. Summarize this contract, rewrite this paragraph, explain this error: the raw material is right there in the prompt, and a sharper model does a sharper job.
Managing a specific human breaks that assumption, because the raw material is not in the prompt. It is in your head, scattered across nine months of moments the model never saw. The time she stayed late to unblock someone else’s release. The pattern where his best work happens right up until a deadline and then quietly slips. The disagreement in March that you handled badly and have been meaning to revisit. None of that is on the internet. A model trained on everything humanity has written still knows nothing about the one person you are responsible for.
So when you ask it to assess them anyway, it does the only thing it can: it reaches for the average. The composite struggling engineer. The generic high performer. And it delivers that composite with total confidence, because confidence is cheap when you have no stake in being right. Generic input gives generic output. No amount of model quality fixes a missing subject.
Assistance versus outsourcing
Here is the line I think actually matters, and it has nothing to do with how advanced the AI is.
An assistant works from what you already know. You have been watching this person all year. You noticed things. The trouble is not that you lack judgment. It is that judgment, left in your head, decays. By review season the year has blurred into a feeling: I think she did well. A grounded assistant takes the notes you actually wrote, in your own words, about things you actually saw, and hands them back to you organized, recalled, and ready. It does the remembering so you can do the thinking. The judgment stays yours. The work it removes is clerical, not moral.
Outsourcing is the opposite move. You skip the watching and the noticing, hand the AI a thin prompt, and let it manufacture the assessment you were supposed to form yourself. Then comes the quiet betrayal. You take its fluent paragraph, put your name on it, and deliver it as your considered view. You have laundered a stranger’s confidence into your own authority. The person across the table assumes they are hearing what their manager observed. They are hearing what a language model guessed about someone like them.
Same technology, both times. The difference is entirely in what sits underneath the request. Real memory, or no memory.
Why grounding changes the answer
Give the same model something true to stand on and it becomes genuinely useful. Not because it got smarter, but because the question changed. “Write feedback for a struggling engineer” asks it to invent a person. “Here are my notes on Dana from the last two quarters. Pull out the recurring themes and remind me what I said I wanted to raise” asks it to organize a person who already exists in your record. The first is fiction. The second is recall, and recall is exactly the thing humans are worst at and machines are good at.
This is also why grounded AI is honest in a way the general kind cannot be. When the source is your own notes, you can see where every claim came from. If the assistant says she stepped up during the outage, there is a note from the week of the outage saying so, in your own words. The AI is not the witness. You were the witness; it is the filing clerk that can find what you saw faster than you can. Nothing gets asserted that you did not first observe and write down.
And when the notes are thin, a grounded assistant has the decency to be thin too. It cannot tell you how someone is doing if you never wrote anything down, and that honest blank is far more useful than a confident paragraph spun from nothing. The general chatbot never says “I don’t know enough about this person.” A tool tied to your own memory says it constantly, which is the correct answer most of the time.
The discipline this asks of you
All of which puts the real demand back where it belongs: on the manager, not the model. Grounded AI is only as good as the observing you actually did. It cannot rescue a year you spent not paying attention. If you want an assistant that prepares you for a hard conversation using what you genuinely know, you have to have genuinely known something, and written enough of it down that there is anything to work from. The tool rewards attention and exposes its absence. That is a feature.
This is the bet behind Notivo. It is a private notebook where a manager jots the small things as they happen (a line about a person after a standup, a tag so it lands in their timeline), and when a 1:1 or a review comes around, the assistant prepares you from those notes and only those notes. Never generic advice, never a composite, never a stranger’s guess wearing your name. It gives your judgment a memory. It does not try to have the judgment for you, because that was never the model’s to have.
The most useful thing AI can do about a person is help you remember what you already knew about them. Everything past that, it is just making up: confidently, fluently, and about someone it has never met.
Frequently asked questions
Will AI write my performance reviews for me?
No, and that is the point. A general chatbot will happily produce a confident review about a person it has never met, which is exactly the trap. Notivo's assistant works only from the notes you wrote about that specific person, so it helps you recall and organize what you saw. It prepares you to write the review. The judgment, and the words you stand behind, stay yours.
How is a grounded AI assistant different from ChatGPT?
A general chatbot answers from the average of the internet, so when you ask about your engineer it invents a composite stranger. Notivo's assistant is grounded in your own per-person notes, organized by month. It surfaces what you actually observed and dated, and when you can see where every line came from, you can trust it. When your notes are thin, it stays thin too rather than making something up.
Is the AI assistant free, and is my data used to train it?
The AI assistant is part of Notivo Plus, the paid plan, and there is a 14-day free trial so you can decide before paying. Capturing and organizing notes by person and topic works without it. Your notes are private by default and account-scoped. They are not a public training set; they exist to prepare you, not to teach a model about strangers.
What happens if I have not taken many notes?
Then the assistant has little to work from, and it will tell you so instead of filling the gap with invention. That honest blank is more useful than a fluent paragraph about no one. Grounded AI rewards the attention you paid and exposes the months you did not. The fix is small: write one line when something happens, tagged to the person, so next time there is a real record to read.