Can AI work out who owns each action item? It can propose an owner. Only the owner can confirm one.
Short answer. Not reliably yet. A model can propose an owner for each action item, and when the transcript says plainly that Meena will send the draft by Friday, it has little to get wrong. Trouble starts when the owner is implied: can you take that, we should, someone from finance, or a name said by one person to another. The owner is the person who will do the task, not the person who asked, and a transcript without reliable speaker labels hides that difference. The safe rule is that the model proposes, quotes the line it relied on, and the named person confirms. ANVE works that way and logs each confirmation.
01
Why ownership is harder than it looks
Research on action items mostly asks a simpler question: which parts of a meeting contain one? A 2023 ICASSP paper notes that datasets labelled for this are scarce and small. Deciding who owns the item adds two more problems, working out who is speaking and resolving words such as you, that and we. The table lists the patterns that cause most confusion. These are examples.
Pattern in the meeting
Example
Risk
Direct assignment
Meena, please send the draft by Friday.
The owner is Meena, not the person speaking. A system that credits the speaker gets it backwards.
Volunteering
I will send it.
The owner is whoever is speaking, so a wrong speaker label means a wrong owner.
Pronouns and gaps
Can you take that? Sure.
The names and the task are in earlier lines. You and that have to be resolved.
Groups
Finance will confirm the figures.
A department is not a person. Someone has to be named.
Mixed-language instruction
Priya, tum draft bhej dena by Friday.
The assignment is in Hinglish. The transcript must be right first.
Hedged suggestion
We should probably update the policy.
It may not be a commitment at all.
Changed mind
Actually, let Ravi do it instead.
The last statement overrides the first.
02
Why speaker labels matter here
Many ownership statements are first person: I will do it. They are only as reliable as the speaker labels. On AMI, a corpus of recorded meetings, and on AliMeeting, pyannote’s model card reports speaker diarization error rates of 18.8 and 24.4 per cent, counting missed speech, false speech and speech given to the wrong speaker. So a good transcript can still attach I will do it to the wrong person. This is why the safeguard has to be a person confirming, not a better model alone.
03
What a person must still check
That the owner accepted. An item with a name on it is a proposal until that person says yes.
That the owner is a real person on the attendee list or staff directory, not a department or a misheard name.
That the task is what was agreed, in the words of the line quoted.
That a later line did not change the owner or cancel the task.
That a due date, if shown, was said aloud and not filled in by the model.
04
How an office would set this up
Define owner in one line on the minutes template: the person who will do the work and has said so.
Have the model output the owner, the due date only if stated, the exact line quoted and the speaker of that line.
Restrict owners to the attendee list or the staff directory. Anything else is flagged for the minute-taker.
Send each proposed item to its owner to accept, reject or reassign. Record who answered and when.
Review items nobody accepted at the start of the next meeting. An unowned item is a finding about the meeting.
Count corrections for a month to learn how often the first proposal was wrong.
05
What ANVE does today
In ANVE’s meeting minutes maker every action item must quote the transcript line it came from, and its owner needs to be confirmed by a person before the item is approved. Each confirmation is logged. The maker works with transcripts in Hindi, English and Hinglish.
This runs today in ANVE’s development build. No sealed unit exists yet, and ANVE does not list automatic speaker identification as a function.
06
Questions people ask
Why not let the AI assign owners automatically?
Because an owner who never agreed to a task turns minutes into a source of disputes. A person’s confirmation turns a guess into a commitment and leaves a record of who agreed.
What happens if nobody accepts an item?
It stays proposed and unowned, and goes on the next agenda. That is a result about the meeting, not a fault in the software.
Does it work for tasks given in Hindi or Hinglish?
Whatever tool you use, test it on your own meetings. Phrasing such as tum bhej dena is an assignment and must be read as one, and the transcript has to be right before the model can do that.
Can the owner be someone who was not at the meeting?
Yes, if the transcript names them, but they have not agreed yet. Keep such items as proposed until the named person confirms.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
