Trust / The AI analyst
The AI analyst
On the accounting platform the analyst answers questions about a client's own reconciled numbers. On WealthVault the assistant lives on the advisory firm's own AI workspace and reads a tokenised view.
The AI sees read-only data, and stores nothing at rest.
The AI reads from a read-only database role. It cannot write, and it retains nothing after a query; the model processes data at query time only. Where an engagement puts personal information in its reach, that information is tokenised before it reaches the model.
No training on client data.
Client data is not used to train models. On the accounting platform the AI layer (Anthropic) operates under a signed Data Processing Addendum incorporating EU Standard Contractual Clauses for the cross-border leg; its SOC 2 Type 2 report is on file. On WealthVault the platform makes no model call of its own: its tasks have no internet route and hold no AI-provider credential, and the firm's own workspace agreement governs training terms.
The analyst does not make decisions or change records.
It answers questions and surfaces what is in your reconciled numbers. It does not move money or change records. The working stays open to inspection.
Guardrails are enforced in CI.
The read-never-write boundary, output validation, and (where an engagement carries it) the tokenisation step are covered by automated tests that run before every deployment, so a regression fails the build before it reaches a client. On WealthVault a connector response that matches a direct-identifier pattern is rejected, and behaviour under instructions embedded in uploaded documents is tested against an adversarial document set.
What the analyst is not.
The AI explains and reconciles. It is not a forecaster or a financial adviser. Where a figure depends on a judgement, that judgement is the client's or their adviser's.