ANVE: the private AI office, on the record. See what that means

ANVE

Private AI glossary · updated 2026-10-07

Action items
Who will do what, by when, and how to make sure an AI list is right.

Short answer. An action item is a task agreed in a meeting, assigned to one or more people and expected to be done within a short time after the meeting. Researchers who build detectors for them describe each item through four elements: a task description, an owner, a timeframe and the group’s agreement. The owner is whoever will do the work, not whoever asked for it. AI can find candidate action items in a transcript, but a draft is only useful if each item points to the line it came from and a person confirms the owner.

01

The four parts

The research definition, used in a 2023 paper on detecting action items, is a task discussed in the meeting and assigned to participants, expected to be completed within a short time window after it. The paper notes that some work detects each element separately.

PartThe question it answersExample
TaskWhat exactly is to be done?Send the revised draft to Mr Rao
OwnerWho will do it?Meera
TimeframeBy when?Friday, 5 p.m.
AgreementDid the owner accept it?“Yes, I will do it”
02

Why detection is hard

The parts are often spread across the conversation. The paper’s own example has the task in one place and the timeframe several lines away, so a detector needs both nearby and distant context. Assignment is often implicit: “let’s get this done by Friday” names nobody.

Data is limited too. The authors report 101 annotated meetings with 381 action items in the English AMI corpus, and say the ICSI corpus has 75 meetings without public action-item annotations. They built a Chinese corpus of 424 meetings to help. A tool tested on these is not tested on a Hindi-English board discussion.

03

Good action-item practice

  • One owner per item, so that no one assumes someone else will do it.
  • Start with a verb and name the deliverable: “Send the revised draft to Mr Rao”.
  • Give a date, not “next week”: a relative date loses its meaning when the list is read later.
  • Read the list back before the meeting ends, so owners hear and accept their items.
  • Link each item to the point in the discussion it came from.
  • Review the list at the start of the next meeting.
04

Checking an AI-drafted list

  • Compare each item with the transcript line it cites. An item with no supporting line should be dropped.
  • Check the owner is the person doing the work, not the person who asked.
  • Turn relative dates such as “by Friday” into calendar dates.
  • Look for missing items: AI drafts miss implicit assignments.
  • Look for duplicates and for decisions mislabelled as actions.
  • Confirm each owner with the owner, and keep a record that you did.
05

Where ANVE stands

In ANVE’s minutes maker, every action item must quote the transcript line it came from, its owner must be confirmed by a person before it is approved, and each confirmation is logged. ANVE never uses a model smaller than 4B-class for minutes, because a 3B-class model failed its test harness; the method is in ANVE’s note on the 4B question.

06

Questions people ask

What is the difference between an action item and a decision?
A decision is what the meeting agreed. An action item is who will do what by when to carry it out. The research literature treats summaries, decisions and action items as separate things to extract.
Can AI work out who owns each action item?
Only when the transcript makes it clear. When it does not, a good tool flags the item for a person to assign. See our question page on owners.
Who should confirm action items?
The owner, and whoever chairs the meeting. Confirmation should be recorded, so that an item can be traced to a person’s agreement.
Where do action items live after the meeting?
In the minutes or in a tracker. What matters is a single list with owners and dates that is reviewed at the next meeting.
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.