Meeting Copilot

What this screen is for

Paste or load the text of a meeting the participants agreed to have transcribed, and Vesper (Aether’s AI assistant) proposes fact-find changes, one file note, follow-up tasks and advice prompts, each tied to an exact quote from the text. You accept, edit or reject every proposal on its own. Nothing reaches the client record without a confirmation, and there is no accept-all.

Aether does not record audio, and the page takes text only. The page is for adviser logins; a portal login is refused.

The household picker offers only the firm’s Active households, which is the same rule the submit path applies, so a household you can pick is one you can analyse. A Prospect household is not offered. A household carried in from another page that fails the rule is dropped with a message. Choosing a household here also makes it the app-wide active household, as the other pickers do.

Tick the household members who took part; up to twenty participants in all. External participants are names only: they never receive a proposed fact, and the model is told so. If a client was in scope when you arrived, they are pre-ticked and must stay among the participants or the run is refused.

Consent needs a method (verbal or written), the time it was obtained, and the attestation box. The time is sent with its timezone and may not be more than five minutes in the future.

The transcript is pasted or loaded from a .txt or .md file, and must run between 200 and 120,000 characters. Windows line endings are normalised before anything is counted or hashed. Text holding a null character, or more than one control character per hundred, is refused, and the whole request may not exceed one megabyte. Analyse stays disabled until household, participants, consent and transcript are all present.

The retention tick adds a separate “Retain source transcript” proposal to the review stage. The text is stored only if you later confirm that proposal.

Each adviser may run twelve analyses an hour by default; past that the page reports too many meeting extractions this hour. If the firm’s AI switch is off, the run fails as provider unavailable. The first-use AI disclosure that Vesper chat asks for does not apply here, because this page carries its own notice and attestation.

What is sent, and what is kept

Sent to Anthropic’s Claude (Haiku 4.5, with Sonnet 4.6 as the fallback), at temperature zero, with a thirty-second timeout and no retry: the participant list with the names exactly as you labelled them and an alias for each (member_1, external_1), the transcript in chunks of 12,000 characters overlapping by 500, the list of seven fact-find fields it may propose, and, when Vesper voice is on, up to three of your earlier file notes as style examples. Names are not replaced with role tokens on this path, and the published AI data-handling page says the same.

Kept before the model is called: the consent record (notice version, method, time, who attested and when), the participants, the file name if any, a fingerprint (a hash) and character count of the transcript, and an audit entry that consent was attested. Not the text.

Kept after: the quotes that back each proposal, each at most 500 characters and all together at most 8,000 characters and thirty per cent of the transcript; the advice prompts; the model’s stated limitations; counts of candidates rejected and why; and the model usage figures. The run reads Review when it produced proposals and Complete when it produced only prompts.

The full text is stored only when you confirm the retention proposal, once per run, and only the adviser who ran the meeting can read it back, through a route that tells the browser not to cache it. Reloading the page after the run drops the transcript from the browser, and the retention proposal’s Accept button disables with a message saying so, so retain before you leave.

The call is counted against the firm’s AI usage under the meeting extraction flow.

What comes back

Kind	What Vesper may propose	Refused when
Fact-find update	One of seven fields for one household member: date of birth, email, mobile, employment status, occupation, base salary, employer. One change per proposal, showing the current value, the proposed value and a confidence	the participant is not a member; the field is not one of the seven; the value does not parse for the field; the quote is not found exactly once in its chunk; confidence is below 70 per cent
File note	One draft of up to 12,000 characters, with at least one cited quote	it has no valid citation, or its body is the whole transcript
Follow-up task	A description, an optional due date, a priority of low, medium or high, and one quote	the due date is not a real date; the description is the whole transcript; no quote
Advice trigger	A label, an explanation, a quote and a confidence. Read only, marked “adviser assessment required”	no quote

The model is capped at 30 fact candidates, 15 tasks, 15 triggers, 10 limitations and 48 proposals in all; an answer over any cap fails the run with “output limit exceeded”. An answer with no valid proposal and no trigger fails it with “invalid model output”. A failed run keeps its consent record and shows an error code; the transcript is not stored.

Stage 3: review

Each proposal is a card with its kind, the current and proposed value for a fact, the quote with its character positions, and Accept, Edit and Reject. The counter reads “x of n reviewed”.

Edit changes only the value of a fact; the field and the member cannot change, and an attempt is refused as an identity edit. For a file note you edit the body, and for a task the description, due date, priority, assignee and notes. An edit that pastes the whole transcript into a note or task is refused with a message pointing at the retention proposal.

Reject takes a reason (Incorrect, Wrong speaker, Outdated, Duplicate, Not a fact-find field, Other) and an optional note of up to 240 characters. Both are recorded against the proposal and in the audit log.

Accept goes through the same confirmation route as Vesper’s chat proposals and the fact-find’s “before we meet” proposals. What confirming writes:

A fact: the one field into the household’s fact-find for that member, with the same follow-through as a manual save (audit rows, fact-find dates, the resync that feeds the engine).
A file note: a note of type File note on the household, and on the client in scope if there was one, authored by you.
A task: a task on the household with the due date and priority you set, its source recorded as Vesper.
Retention: the text, after a check that it matches the fingerprint and character count recorded at the start. A second attempt answers that the transcript is already retained.

Every confirmation and rejection is audited, a proposal can be confirmed once, and the run returns to Complete when every proposal is resolved. Only the adviser who ran the meeting can open its run; the page adds the run number to its address so you can come back to it.

Vesper voice

Voice learning is off by default and set per adviser. When it is on, up to three of your earlier approved file notes, chosen for topic from your most recent fifty, travel with the request as style examples, and the model is told to copy tone only and never a fact. The library fills from a job that runs at 5:00 AM Sydney time, taking file notes you confirmed here and file notes you wrote by hand, up to two hundred per adviser. The metrics line reports the change in how much you edit drafts once drafts written with your voice have been confirmed. Delete voice library removes every entry.

What it does not do

Record audio or accept an audio file.
Write anything without a confirmation, or accept proposals as a batch.
Propose a fact outside the seven fields, or for anyone who is not a household member.
Store the transcript unless you tick retention and then confirm it.
Replace names with role tokens before sending.
Offer a Prospect household.