Can AI write meeting minutes with action items? Yes, as a draft. The risk is in who gets the work.
Short answer. Yes, as a draft. A language model can turn a transcript into a summary, decisions and action items in seconds, and the result reads well. The risk is substance, not style. Published studies of meeting summaries still find hallucinated and missing points, and action items are where a mistake costs most, because someone is handed work they never agreed to. The workable pattern is to make the model quote the transcript line behind every item, and to have a person confirm owners and dates before anything is approved. ANVE applies that rule to its own minutes, and never uses a model smaller than 4B-class for them.
01
What can go wrong in each part of the minutes
Researchers who study meeting summaries sort the mistakes into types, including omission, irrelevance and hallucination, the last meaning content that is not in the transcript at all. The table shows where each is likely to bite. These are examples.
Part of the minutes
What can go wrong
How to catch it
Summary of the discussion
Points left out, points invented, or a side remark given the weight of a decision.
Compare with the transcript, and check that every agenda item has a line.
Decisions
A proposal recorded as agreed, or a deferred item recorded as decided.
Require a quote of the line where the decision was made.
Action items
Tasks nobody accepted, owners guessed from who spoke last, dates that were never said.
Require a quote, then have the named person confirm.
Numbers and names
Figures misheard, names spelled differently from the attendance list.
Check against the attendance list and the papers tabled.
02
What the research says
In 2024 Kirstein, Ruas and Gipp released QMSum Mistake: 200 automatically generated meeting summaries, annotated by people for nine error types including omission and irrelevance. They report that large language models still struggle to stay relevant and to avoid hallucination. A second model can flag many of the errors, since GPT-4 Turbo found them with about 89 per cent accuracy overall, but it did worst on hallucinations at about 72 per cent. A second model pass can help, but it does not remove the need for a person.
Public evidence on action items specifically is thin. A 2023 ICASSP paper notes that meeting datasets labelled for action-item detection are scarce and small, and tests on the English AMI corpus, which holds 100 hours of recorded meetings, alongside a new Chinese corpus. So there is little published evidence on how any model does on your kind of meeting, which is why a trial on your own recordings matters.
03
What a person must still check
Is each action item a real commitment? A suggestion such as we should probably update the policy may not be one.
The owner is the person who will do the work and has agreed to, not the person who asked.
Due dates appear only where someone said one. Flag any date the model supplied.
Decisions are decisions, not proposals. Check the wording against the transcript.
Sensitive remarks about people, legal advice or commercial terms. Someone must decide what belongs in the record.
Names, amounts and dates against the attendance list and the papers tabled.
04
How an office would set this up
Fix the layout of the minutes once: attendees, decisions, actions with owner, due date and source line, then open questions.
Run the model on a transcript, not on raw audio, and keep the transcript with the minutes.
Require a quote for every decision and every action. Reject any item without one.
Send each draft action to its named owner for a yes or no within a day, and record who confirmed and when.
For a month, count how many owners and dates people correct per meeting. If corrections are frequent, change the prompt or the model before you widen use.
Circulate approved minutes only. Keep drafts private to the minute-taker.
05
What ANVE does today
ANVE’s meeting minutes maker works from transcripts in Hindi, English and Hinglish. Every action item must quote the transcript line it came from and needs its owner confirmed by a person before it is approved. Each confirmation is logged.
ANVE never uses a model smaller than 4B-class for minutes. A 3B-class model failed ANVE’s test harness, which the engineering note The 4B question describes. The minutes maker runs today in the development build; no sealed unit exists yet.
06
Questions people ask
How accurate are AI meeting minutes?
There is no single figure. Studies find omissions and invented points in model-written summaries, so accuracy depends on the model, the audio and the review step. The separate question on accuracy goes into more detail.
Can AI write the minutes in Hindi or Hinglish?
Open models can work from Hindi and Hinglish transcripts, but the minutes are only as good as the transcript. Test transcription on your own recordings before you test the minutes.
Who is responsible for the minutes?
The people who approve and sign them. A model draft changes how quickly the minutes are written, not who answers for them.
Does it need the internet?
Not when a local model runs on your own machine. A cloud service does need it, and sends the transcript to the provider.
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.
