How Aether's AI features handle client information

Aether uses an AI model to draft. It never uses one to decide. This page says exactly what leaves Aether, when, and what a firm can switch off — in enough detail for an IT or privacy reviewer, in plain enough language for an adviser explaining it to a client.

The provider is Anthropic PBC, through the Claude API. It is listed with every other third party on the subprocessors page.

Last reviewed: 3 November 2026.

The three rules that bound everything below

The model has no authority. It never reads the database, never calls a tool, and never writes a record. Aether builds the context in code from queries scoped to one firm (for chat, to one household; for the briefing, to the adviser’s own book; for the migration features, to the firm’s records), sends it, and parses what comes back into a strict, allow-listed shape. Anything a proposal would change is computed in Aether’s own code and applied only after an adviser confirms it.
Nothing is sent until a person does something. Every feature below is triggered by a person — an adviser typing a question, pasting a transcript, selecting a document or requesting a generation; a firm administrator or an Aether operator running a data migration; a client asking in the portal. The one exception is the morning briefing, which runs at 6:30 am Sydney time each day for every adviser who has it turned on (it is on unless the adviser turns it off in their notification preferences).
A firm can switch it all off. One setting, described at the end of this page, disables every feature here. With it off, no request leaves Aether with the firm’s data. (One model call is not the firm’s at all: see product documents below.)

What each feature sends

Feature	What triggers it	What is sent to Anthropic	What Aether keeps
Vesper — adviser chat	An adviser types a message	The message, plus context Aether builds in code: the household’s stage, next review date and fee-consent expiry; each member’s role and fact-find date; funds under management; counts of tasks, goals and vault items; recent document titles; open task titles. The names Aether holds for the household — its name, its clients’ and dependants’ names, and the other people and entities recorded for it — are replaced by role labels such as Client A or Entity 1 — in the context and in the typed message — and put back on the answer before the adviser sees it (see what is pseudonymised for the names that are not). Optionally the adviser’s own writing samples, if that adviser opted in	The conversation text for 365 days, then deleted. A 500-character excerpt of each message the adviser types in the Vesper audit trail, kept with the other audit trails for the life of the tenancy and then 7 years. A usage row (tokens and cost — never content)
Morning briefing	A scheduled run at 6:30 am Sydney time, per adviser	The top eight briefing lines’ titles and detail, which include task text and client-request subjects, plus counts by kind and Aether’s own summary of them, with the names Aether holds for the households in those lines replaced by role labels	The briefing text
Meeting copilot	An adviser pastes a meeting transcript and attests that participants consented	The full transcript (200–120,000 characters), with participants labelled by their names. Optionally the adviser’s own writing samples, if that adviser opted in, as written	The transcript is not kept by default — only its length and a hash. An adviser can choose to keep it through a separate confirmed step. The extraction summary is kept
Fact-find extraction from a document	An adviser selects a stored statement	The extracted page text of that statement (up to 50 pages / 120,000 characters)	A proposal, until an adviser confirms it. A statement that matches an account the fact-find already holds records its balance at once, beside the account’s current figure for the adviser to review
Portfolio import	An adviser uploads a holdings file	The uploaded PDF (up to 20 MiB) or image (up to 5 MiB) as it is, or its extracted text with tax file numbers withheld	A proposal only. Nothing is stored
Adversarial review	An adviser runs a review of draft advice	The draft plan: strategy codes, strategy names, advice areas, statuses, priority scores and rationale text, plus recommendation titles and the client’s internal id number — no name field, though recommendation titles are free text an adviser types	The review memo and its events
Client-portal assistant	A client asks a question in the portal, or sends their adviser a change request	The question or the request’s description, plus sections of that client’s already-published advice documents	For a question, the client’s question and the answer (the answer only if every citation resolves), kept for 365 days like chat, and a 500-character excerpt of the question in the Vesper audit trail for the life of the tenancy and then 7 years. For a change request, the request and a draft reply kept for the adviser to review before anything is sent; no citation check applies to the draft. Off by default for every firm
Document narrative and polish	An adviser requests it for one generation, or starts an opted-in campaign, which then runs on a schedule	Structured section data — client names and figures — or template prose	The document
Data migration: column mapping suggestions	A firm administrator or an Aether operator asks for suggestions during a data migration	Source column names, as they are, and up to three 80-character sample values per column, with the person, entity and household names the firm already holds replaced by role labels — on a first import, none yet	The suggestions
Data migration: archive analysis	A firm administrator or an Aether operator runs analysis of imported archive items	An archive item’s title, body (up to 4,000 characters), source record (up to 2,000 characters), source system and date, and up to ten candidate client matches, with the person, entity and household names the firm already holds replaced by role labels	The analysis
Data migration: exception triage	A firm administrator or an Aether operator asks for triage of a batch’s validation exceptions	Up to 300 exceptions’ severity, category, code, message, field and record type, with the person, entity and household names the firm already holds replaced by role labels	The triage summary

Uploaded files, meeting transcripts and statement text travel as content. That is inherent to what those features do: an extraction feature that did not send the document could not extract anything. It is stated here rather than buried, so a firm can decide which features to use.

Product documents

One model call carries no firm’s data. When Aether’s operator loads a financial product’s published documents (a product disclosure statement, a fee schedule) into Aether’s shared product library, the document’s text is sent to the model to extract the product’s details. It contains no client information, it runs outside any firm, and so the firm switch does not apply to it. It uses Claude Sonnet 4.6.

What Aether does not send

Passwords, password hashes, session cookies, API keys or any other credential.
Bank account numbers or payment details from the fields Aether builds the context for chat, the briefing or the adversarial review from. Free text still carries what a person wrote into it — an adviser’s typed question, a task or document title, a client’s request — and so do three kinds of feature: a document a firm chooses to send through extraction, which is why extraction is per-document and adviser-triggered; a meeting transcript, which carries whatever was said aloud; and the data-migration features, which send values from the firm’s own source files as they stand.
Tax file numbers, in any text. Before any text leaves Aether, from every feature on this page — statement text, migration source data, meeting transcripts and typed prompts included — Aether withholds every value in a field named as a tax file number and every nine-digit number, whether written as one run or in groups split by spaces, hyphens or dots; an amount with a decimal point is not treated as one. That also withholds some nine-digit numbers that are not TFNs (a member or policy number of that length, or a nine-digit record number), so the model never sees them: it cannot quote one back, and a product document’s excerpt holding a nine-digit run reaches it with that run withheld, which can leave a figure read from it flagged for review instead of confirmed. Two cases still carry one: a portfolio PDF or image an adviser uploads for import is sent as it is, and a TFN in a form Aether does not recognise, such as an eight-digit TFN written without a TFN label, is sent as written.
Any other firm’s data. Context is built under row-level security scoped to the requesting firm.

What is pseudonymised, and what is not

Where Aether builds a request in code from its own records, it replaces names before the request leaves and puts them back on the answer:

Pseudonymised: Vesper adviser chat (the household context, the typed message, and the first call that sorts the question), the morning briefing, and the three data-migration features — column mapping suggestions, archive analysis and exception triage. The names replaced are every person and entity name Aether holds for a household: the household’s name and any former name it is recorded under; its clients’ first, middle, last and preferred names, and their initials with a surname (R. Smith, Smith, R); its dependants’ names; and the other people and entities recorded for it — its trusts, SMSFs and companies and the trustees, appointors, beneficiaries and office-holders named on them, nominated and likely beneficiaries and attorneys from estate records, the parties to an identity check, a complainant, an accountant or controller named on a wholesale certificate, and the contact an SMSF document was requested from. A client becomes Client A or Partner, a dependant Child 1, a household Household A, another person Person 1 and an entity Entity 1. Chat and the briefing use those of the households in that request; the migration features use those of every household the firm holds. Aether also checks each of these requests before it leaves: one that still carries a name it holds is not sent, unless the name is the whole of a column header or target field, which the mapping suggester sends as written (below).
Not pseudonymised, because sending the content is the feature: meeting transcripts, statement extraction, portfolio import, the client-portal assistant (which quotes that client’s own published advice), and document narrative and polish, which sends client names and figures. The adversarial review sends no name field, but its recommendation titles are free text an adviser types and are sent as written.
What the replacement does not catch.
- Names Aether holds but does not map: a lone initial or other one-letter name part, and a name held only in a field Aether does not read for names — a file name, a product or fund name, a bank account’s display name, a provider’s name — or written inside a free-text note in a form Aether does not take for a name.
- People outside the request: chat and the briefing know only the people in the households the request is about, so in chat a name outside them — another client’s, a third party’s, a staff member’s, or any name when no client is open — is sent as typed, and an adviser’s opted-in writing samples can carry another client’s name. The briefing’s free text (a client’s request subject, a task naming someone else) is sent as written. The migration features know only the people the firm already holds, so on a first import nothing is replaced; column headers are never rewritten.
- Spellings the match misses: the match is whole and uses the exact spelling and capitals Aether holds, so a name written in a different case or spelling is sent as written: in other capitals (john, JOHN, Mcdonald for McDonald), without its accents (Zoe for Zoë), joined to other letters or digits (the Smiths, Smith2024.pdf, johnsmith@…), as a nickname, or as part of a longer name Aether holds (Mary for a first name held as Mary Jane, one half of a hyphenated surname held whole). A held name set off by punctuation (John.Smith@…, Smith's) is replaced.
Where the reverse goes wrong. Putting names back can put one where none was meant. A given name that is an ordinary word is replaced where it is capitalised as the word, so “Will the fee consent expire?” can come back with a client’s name in it. And the role labels themselves are put back wherever the model writes them, so an answer that says “the household” or “Partner contributions” in the ordinary sense can be shown with a real name in that place. A surname, or a first name only one client holds, is replaced wherever it appears, so a third party who shares it (“call Smith the accountant”) is sent as the client’s label and comes back named as the client.

What Anthropic says it does with it

Quoted from the provider’s own published terms and policies, retrieved 3 November 2026:

Training. Anthropic’s Commercial Terms of Service (effective 17 June 2025) state: “Anthropic may not train models on Customer Content from Services.” Its privacy centre adds: “By default, we will not use your inputs or outputs from our commercial products (e.g. Claude for Work, Anthropic API, Claude Gov, etc.) to train our models.” The exception it names is content a customer explicitly submits as feedback through Anthropic’s own thumbs-up/down control. Aether sends no such feedback.
Ownership. “Customer (a) retains all rights to its Inputs, and (b) owns its Outputs.”
Retention. “For Anthropic API users, we automatically delete inputs and outputs on our backend within 30 days of receipt or generation”, with exceptions for longer-retention services, agreed arrangements, policy enforcement and legal compliance. On flagged content, its privacy centre says: “We retain inputs and outputs for up to 2 years and trust and safety classification scores for up to 7 years if your chat is flagged by our automated trust and safety systems as violating our Usage Policy.”
Zero data retention. Anthropic offers a zero-data-retention arrangement on request, per organisation, for eligible API features. Aether has not requested one. The models Aether uses — Claude Haiku 4.5, with a Sonnet 4.6 fallback — are not among those Anthropic excludes from the arrangement. If a firm’s obligations require zero retention, say so before signing: it is a change Aether would pursue with the provider, not a setting in the product.

Sources: Commercial Terms · How long do you store my organization’s data? · Is my data used for model training? · API and data retention

Anthropic is a United States provider. Aether does not pin a processing region for these calls.

What Aether has not done yet

Stated plainly, because a reviewer will find it either way:

Names are replaced only where Aether builds the request. The features that send content send the names in it, and the replacement has the gaps listed above.
Numeric grounding of chat answers is audited, not enforced. Figures in a Vesper chat answer are checked against the context and the mismatch is logged; they are not yet blocked. Every money amount and percentage that reaches a document is gated: the document layer reverts any polished or narrated text that adds one the source does not hold. Counts, ages and calendar years are not checked.

The controls a firm has

The master switch. Settings → AI assistance turns every feature on this page off for the whole firm. A firm administrator sets it, and Aether records who acknowledged this disclosure and when. With the switch off, features that can work without a model fall back to their deterministic behaviour and the rest are unavailable — no request leaves Aether.

Per-feature controls, on top of the switch.

The client-portal assistant is off by default and must be turned on per firm.
Document narrative is off unless requested for a specific generation.
Adviser writing-style learning is per-adviser opt-in.
Meeting copilot requires an explicit attestation that participants consented, and does not keep the transcript unless the adviser separately chooses to.
Portfolio-import AI, adversarial review and voice learning each have their own operator-level off switch.

Spend and abuse bounds. Four flows have a per-user hourly limit: Vesper chat 120, meeting extraction 12, AI document generation 20 and portfolio-import AI calls 10. Statement extraction, the adversarial review, the client-portal assistant, the migration features and campaign document generation, which runs on a schedule, have none of their own yet; every flow has per-request timeouts, and document generation a per-document call cap. Each call made for a firm writes a usage row recording firm, user, flow, model, token counts and estimated cost (never content); a call with no firm (product documents) is logged instead, and a row that fails to write is logged and not retried. Morning-briefing calls are recorded under the chat flow’s name.

Disclosure. Each adviser sees a short version of this page inside Vesper before their first use of an AI feature and acknowledges it; the acknowledgement is recorded with their name and the time. Every feature on this page refuses to send anything for an adviser until it exists, and refuses too when Aether cannot read whether it exists, so it is a gate, not only a prompt. The client-portal assistant runs for no adviser, so it needs a firm administrator’s acknowledgement for the firm instead; campaign generation needs the acknowledgement of the adviser who starts it and the firm’s, and so does the morning briefing. When the short version changes, every adviser acknowledges the new edition, and a firm administrator acknowledges it for the firm from Settings → AI assistance; until then the features that need it send nothing. The meeting copilot also carries its own disclosure and consent attestation on the page itself.

Questions

Email reid@aetherforge.au. Aether’s Security Review Brief covers the controls behind this page in full, and is available to a firm’s IT reviewer on request.