Summaries are usually good. Action items are where things get missed. An AI is reliable at describing what a meeting was about. It is much less reliable at catching the small, casual commitments people make in passing, which are exactly the ones you needed written down.
The usual cause is not the model. It is asking one request to produce a summary, key points, decisions and action items all at once, so the tasks lose out to the prose. Extracting action items as their own separate pass finds far more of them.
Not the obvious things. Nobody's AI misses "we agreed to a June deadline." What goes missing is the soft stuff:
These are precisely the commitments that cause trouble later, because nobody wrote them down and both people remember them differently.
Two reasons, both fixable.
Shared attention. A single request asked to produce a summary, decisions, key points and action items will do the prose well and treat the task list as a byproduct. Splitting it so action items get their own dedicated pass over the whole transcript changes the results substantially.
An over-strict definition. Systems told not to invent tasks from passing remarks end up filtering out real commitments that happened to be said casually. For a notetaker, a missed task is much worse than a spare one. Deleting an unwanted item takes two seconds. A missed one can cost a client.
Jotted was rebuilt around exactly this. Action items are extracted in their own pass, the transcript is scanned in overlapping chunks so nothing falls between them, and the system is deliberately tuned to over-include rather than miss. Testing on a real meeting against a hand-written list took it from six or seven missed items out of fifteen down to zero or one.
Transcription accuracy is generally high on clear speech and drops on accents, crosstalk, technical vocabulary and mixed-language conversation. Two things help.
Using a larger transcription model, which is slower but noticeably better on accented or mixed speech. And wearing headphones, so your microphone is not also picking up the other person through the speakers.
One thing worth knowing from doing that exercise: the AI is sometimes right when you are wrong. In Jotted's own testing, the extraction found three real commitments the founder's from-memory list had left out, and correctly attributed one she had misremembered.
Often, yes, particularly small casual commitments with no stated owner or deadline. Big decisions are usually captured well. The gap is in the soft promises people make in passing.
Usually because summary and action items are produced by the same request, so the task list becomes a byproduct of the writing. Extracting action items in a separate dedicated pass finds far more of them.
High on clear speech, lower on strong accents, crosstalk, technical terms and mixed-language conversation. A larger transcription model and headphones both help noticeably.
For anything that matters, yes, especially the action items and who each one belongs to. Keeping the recording means you can settle a disagreement rather than argue about a summary.
If your calls involve accents or more than one language, usually yes. It takes longer to process but is meaningfully more accurate. In Jotted the model is a setting you can change.
Record free forever, no card and no bot. Join the waitlist and you'll be first in.
Free forever plan · Unlimited recordings · No bot, ever