> ## Documentation Index
> Fetch the complete documentation index at: https://docs.pete.vet/llms.txt
> Use this file to discover all available pages before exploring further.

# AI data handling

> AI-assisted case work stays source-backed, reviewable, and protected for sensitive firm, client, claimant, and veteran information.

AI-assisted features support case review. Sensitive data handling depends on approved vendor paths, configuration, and firm policy.

Customer case data, veteran records, uploaded files, prompts, and case-specific outputs are not used to train third-party large language models. Sensitive AI workflows run only through provider paths approved for the data being processed.

## Grounded outputs

Answers come from the case materials available in the file. Substantive answers point back to case materials so the attorney can verify the source.

If the answer is not supported by the record, the response says so. Missing citation links, missing page or location metadata, and missing quoted source text are marked so unsupported output is not treated as verified.

## AI-assisted case updates

Important case changes stay reviewable. Safe source-backed factual or operational updates may be reflected in the case when they are supported by the record. Uncertain, deadline-related, legal, strategy-sensitive, or conflicting updates appear as attorney review work, with source context preserved for verification.

When new source material creates reviewable changes, the case digest shows what changed and who needs to review it. Accepted attorney corrections stay attached to the case, while correction-dependent legal or strategy issues remain attorney-reviewable.

## Data minimization

Keep AI inputs limited to the information needed for the task. Do not paste unrelated client, claimant, or veteran records; medical details; disability facts; or legal strategy into unsupported tools.

Analytics and operational logs do not contain raw person PII, medical facts, document contents, prompts, model outputs, or filenames. Processing retry and classification events use safe metadata such as case ID, firm ID, source type, status, and provider request IDs.

If a user opts in while connecting email, a bounded sample of recent sent emails supports that user's email drafting style profile. Only the compact profile is stored, not copies of those sent emails. The profile helps future email drafts match the user's normal tone, grammar, punctuation, greeting and closing habits, and structure. It does not override case facts, attorney review, or final send approval.

## Attorney responsibility

AI output is support for attorney work. Attorneys remain responsible for client advice, legal strategy, filings, and final work product.

Do not rely on AI output as the only source for a legal deadline, filing decision, client communication, medical conclusion, VA form, or final work product. Verify source citations, page context, record provenance, and any generated draft before using it in a matter.

## Vendor review

Before processing sensitive data at scale, confirm that the firm's vendor review, data retention expectations, and client notice obligations are satisfied.

If a workflow requires a data processing addendum, data protection agreement, security review, or firm approval, complete that review before using the workflow with sensitive production data.