Trust Center / AI Privacy

Controlled AI processing for permission-scoped case work.

Caseflow AI features are designed to operate inside workspace, case, document, and permission boundaries. The goal is useful assistance with clear human review, not autonomous professional decision-making.

AI architecture

The AI path is backend-mediated. The browser asks Caseflow for an AI-assisted task; Caseflow validates access, prepares limited context, calls the configured AI provider, and returns an assistive response.

  • AI requests are initiated by signed-in users from supported workspace features.
  • Requests are routed through Caseflow backend controls rather than direct browser calls to the AI provider.
  • Relevant case, task, and document context is selected according to current access permissions.
  • The AI response is returned as an assistive output that requires human verification.

What may be processed

Depending on the feature and the user action, AI processing may include selected prompts, case fields, task records, document chunks, generated summaries, and operational metadata.

Case context

Case titles, notes, task summaries, timelines, selected records, and workflow metadata relevant to the request.

Document context

Selected text chunks or extracted document signals from documents the user is allowed to access.

Operational metadata

Tenant, case, user, feature, model, token counts, redaction status, timestamps, and retention-expiry metadata.

What AI is not

AI output is not a final decision layer. Caseflow positions AI as a review aid for professionals who remain responsible for source verification and final use.

  • AI can produce incomplete, outdated, or incorrect outputs.
  • AI output is not legal, tax, financial, medical, investigative, or professional advice.
  • Users must verify generated summaries, extracted facts, citations, and proposed actions against source records.

Permission-aware retrieval and vector safeguards

  • Tenant and case metadata are filtered with hashed identifiers where provider metadata is used.
  • Accessible document sets are rebuilt from current case, task, shared-document, and inheritance records.
  • Retrieved vector results are post-validated before they are treated as usable context.
  • Superseded or stale vector artifacts are rejected on the hardened retrieval path.
  • Context budgets limit how much record and document material is included in prompts.

Human oversight and confirmation controls

  • High-risk or write-capable AI actions require explicit user confirmation.
  • Confirmation nonces are server issued, short lived, and bound to the case, session, message, user, action, and risk category.
  • Destructive or unsupported actions are blocked rather than inferred from free-form text.
  • Users must review AI-generated records, document outputs, and proposed updates before relying on them.

Retention, deletion, and auditability

Audit metadata

AI audit records focus on operational traceability, request purpose, retention expiry, and safety status rather than broad raw-content logging.

Deletion

Case deletion is designed to clean up related AI chats, audit metadata, temporary processing records, and index references.

Saved outputs

AI outputs saved into cases, reports, notes, or documents follow the retention lifecycle of those workspace records.

Controls and honest limitations

  • Raw document uploads to AI providers are blocked in production unless an approved exception is configured.
  • Pattern-based redaction reduces common personal-data exposure but is not marketed as guaranteed anonymization.
  • AI write actions are allowlisted, risk classified, and confirmation-gated with short-lived server nonces.
  • Audit metadata records the operational context of AI use without treating logs as a complete transcript store.
  • Case deletion workflows are designed to remove related AI sessions, audit metadata, temporary artifacts, and index references.
  • Legacy document URL and attachment surfaces may still exist outside migrated controlled-access flows.
  • Sensitive AI workflows should be reviewed against your workspace settings and support requirements.