AI Agent for Meetings & Follow-Through
The expensive failure of a meeting is not the missing summary, it is the decision everyone remembers differently and the action nobody owns. This agent reads the transcript against the agenda and returns decisions, owned actions and the topics that were left open.
What is an AI meeting assistant?
An AI meeting assistant is a governed software worker that converts meeting transcripts into structured records. It distinguishes decisions that were actually agreed from discussion that reached none, extracts commitments with the person who accepted them, carries unresolved topics forward, and prepares follow-up tasks for human confirmation.
What it does
What it is not
Everybody left the room with a different decision
Meeting notes record what was said, which is the least useful thing about a meeting. What matters is what was decided, who agreed to do what by when, and which questions were raised and never closed — and those three are exactly what a summary of the conversation flattens into a paragraph of topics covered.
Discussion reads like decision
A proposal that was explored and not agreed appears in the notes in the same form as one that was settled.
Actions have no owner
An action item is recorded in the passive voice, and three weeks later nobody can say who had accepted it.
Open questions vanish
A material objection is raised, parked because of time, and never appears in any record again.
Confidential discussion cannot use tools
Exactly the meetings that most need a record — legal, personnel, commercial — are the ones a hosted transcription service must not see.
Decisions, owners and open questions
Decisions
What Was Actually Agreed
And what was only discussed.
The transcript is read for the point at which a position was accepted rather than explored, and decisions are recorded with who agreed and on what basis, keeping them separate from the proposals that were raised and left unresolved.
- Decisions distinguished from discussion
- Who agreed recorded against each one
- Basis or condition captured where stated
- Reversals later in the meeting honoured
Agreed vs discussed
Actions
Every Action Has A Name On It
Or it is flagged as unowned.
Commitments are extracted with the person who accepted them and the date agreed, and where an action was raised without anyone taking it the agent records that explicitly instead of assigning it to whoever spoke last.
Owner and date
Confidentiality
The Meetings You Cannot Send Away
Processed inside your network.
Board discussions, personnel matters, deal negotiations and legal calls are exactly the meetings where a record is most valuable and a hosted transcription service is least acceptable, so all processing happens on infrastructure you control.
Never sent out
How the AI Meeting Assistant runs a task
- STEP 01
Set the context first
The agenda, attendee list, pre-read documents and the previous meeting in the series are loaded before the transcript is read, because an action referring to the thing we discussed last time is otherwise unresolvable.
Agenda parsingSeries history - STEP 02
Attribute the speech
Statements are attributed to participants where the transcript supports it, and where speaker attribution is uncertain the agent says so rather than guessing, since a decision credited to the wrong person is worse than an unattributed one.
Speaker attributionConfidence marking - STEP 03
Find the decision points
The conversation is read for the moment a proposal moved from being discussed to being accepted, together with any condition attached, and later reversals in the same meeting supersede earlier agreement rather than sitting beside it.
Decision detectionSupersession handling - STEP 04
Extract and test commitments
Each action is captured with the person who accepted it and the date agreed, and anything raised without an explicit acceptance is listed as unowned so the chair can assign it rather than discovering it later.
Commitment extractionOwner validation - STEP 05
Prepare, then wait
Tickets, follow-up messages and record updates are drafted from the accepted actions and held for confirmation, because a meeting record that silently creates work in other systems stops being trustworthy very quickly.
Task draftingConfirmation gate
Systems the AI Meeting Assistant connects to
Meeting input
Analysis
Inputs, outputs and runtime
- Ingests
- Meeting transcript or recordingAgenda and attendee listPre-read documentsPrevious meeting recordLinked tickets
- Produces
- Decision log with agreementAction items with ownersUnowned action listUnresolved topicsDrafted follow-up tasks
- Triggered by
- Meeting endsTranscript uploadedSeries follow-up check
- Human oversight
- The chair confirms actions before they are created
- Models
- Open-weight LLMs you host — Llama, Qwen or Mistral class
- Typical latency
- Minutes after the meeting ends
- Deployment
- On-premise or sovereign cloud with egress control
- Data residency
- Recordings and transcripts never leave
Where the Meeting Assistant pays back
Decision Records
Produce a decision log from a meeting showing what was agreed, by whom, and under what condition.
Action Item Extraction
Pull out every commitment with its owner and date, and flag the ones nobody actually accepted.
Open Question Tracking
Carry forward the topics that were raised and parked so they reach the next agenda rather than disappearing.
Confidential Meeting Records
Produce records for board, legal and personnel discussions without a transcript leaving your network.
Follow-Up Task Creation
Prepare tickets or tasks from accepted actions for a human to confirm before they are created.
Series Continuity
Check a recurring meeting against the previous one and report which actions were never closed.
AI Meeting Assistant vs chatbots and SaaS copilots
Meeting tools converged on the summary because it demos well, but nobody has ever been harmed by not having a summary — they have been harmed by an action with no owner and a decision two people remember differently.
| Generic chatbot | SaaS copilot | VDF AI | |
|---|---|---|---|
| Primary output | A summary | A summary | Decisions, actions, open topics |
| Decision vs discussion | Merged | Merged | Kept separate |
| Action ownership | Passive voice | Sometimes named | Named or flagged unowned |
| Unresolved topics | Dropped | Dropped | Carried to the next agenda |
| Creates tasks | No | Automatically | Drafted, confirmed by the chair |
| Confidential meetings | Unsuitable | Vendor cloud | Processed on your own hardware |
| Where audio is processed | Vendor service | Vendor cloud | Inside your own network |
Governance and controls
A meeting record can become evidence in a dispute, a grievance or a regulatory enquiry, so what it claims was decided needs to be traceable to the point in the transcript where that happened.
Decisions cite the transcript
Each record links to where it was said
Uncertain attribution flagged
Unclear speakers are never guessed
No unconfirmed task creation
Actions wait for the chair to approve
Participant visibility respected
Records shared only with attendees
Retention under your policy
Transcripts kept for your stated period
No performance inference
The agent does not assess individuals
Evidence it leaves behind
What changes after rollout
Who runs the AI Meeting Assistant
Programme chair
Closes a meeting with the decision log and the unowned actions already visible, so assignment happens while everyone is still in the room instead of by email four days later.
Company secretary
Produces board records where each decision points to the passage that establishes it, without a recording of a confidential discussion ever leaving the organisation.
Delivery manager
Sees which actions from the previous three meetings in a series were never closed, which is a more honest status signal than anything reported in the meeting itself.
Questions about the AI Meeting Assistant
What is an AI meeting assistant?
It is an agent that turns a meeting transcript into a usable record: separating decisions from discussion, extracting commitments with their owner and date, carrying forward unresolved topics, and preparing follow-up tasks for confirmation.
How is an AI meeting assistant different from a generic chatbot?
A chatbot summarises what was said. This agent distinguishes what was agreed from what was merely raised, names who accepted each action, and flags the commitments nobody actually took.
Can an AI meeting assistant run on-premise on meeting transcript data?
Yes, and it is usually the reason for deploying it. Board, legal, personnel and negotiation discussions cannot go to a hosted transcription service, and those are the meetings where a record matters most.
What does an AI meeting assistant produce, and in what format?
A decision log with who agreed and any condition, action items with owners and dates, an unowned-action list, unresolved topics carried forward, and drafted follow-up tasks.
Where does an AI meeting assistant fit in a governed AI programme?
It records and prepares; it does not act. Creating tickets, sending follow-ups and updating records happen only after a person confirms, and scheduling belongs to the scheduling assistant.
Does it join meetings and record them?
It works from transcripts and recordings your own platform produces rather than joining as a participant. That is deliberate: a bot appearing in a confidential meeting changes the conversation, and in several jurisdictions recording requires notice and sometimes consent from everyone present. Your existing recording policy stays in force and the agent processes what that policy already permits to exist.
How does it tell a decision from a discussion?
By looking for acceptance rather than for topic. A proposal that was raised, explored and left without anyone agreeing is recorded as an open topic; one that a named participant accepted, with any condition attached, is a decision. Where the transcript is genuinely ambiguous it is recorded as unclear rather than resolved, because a meeting record that overstates agreement is how two teams proceed on incompatible assumptions.
Will it create Jira tickets automatically?
It drafts them and the chair confirms. Automatic creation seems efficient until a hypothetical raised in discussion becomes a ticket, or an action assigned to the wrong person starts appearing in their workload. One confirmation step keeps the record trustworthy, which is the property that makes anyone use it at all.
Can it be used for personnel or grievance meetings?
Technically yes, and that is a policy decision rather than a technical one. Processing happens entirely inside your network, which removes the third-party exposure, but recording and note-taking in personnel matters is usually governed by HR policy, employment law and often works council agreement. The agent is deliberately constrained to what was said and decided, and does not assess participants or infer anything about performance or conduct.
How does it relate to the executive assistant agent?
This one goes deep on a single meeting and produces its record. The executive assistant works across the whole day — mail, calendar, documents, commitments — and uses meeting records as one input among several, for instance to prepare a brief before the next meeting or to surface a commitment that is now overdue. Most deployments run this agent for everyone and the executive assistant for a small number of people.
Leave the meeting with the decisions written down
See the AI Meeting Assistant turn a transcript into decisions and owned actions.