Data handling

ADO Pilot sends only the data that an AI reviewer needs to comment on a pull request: the diff content of changed files and their paths. Your full source tree is never read, copied, or indexed. The data is held in memory for the duration of the review and is not persisted in our systems after the review completes.

What we send to Anthropic

For each review, ADO Pilot constructs a single prompt and sends it to Anthropic's Claude API (the Batch API for the first pass, the Messages API for the second pass and any fallbacks). The prompt contains exactly three pieces of content, concatenated into the user-facing portion of the request:

• The unified diff of files changed in the pull request — added, removed, and context lines as Azure DevOps returns them, including the file paths embedded in the diff headers so Claude can reason about language and structure.
• Tree-sitter file-structure skeletons of the changed files, when our enrichment pipeline can extract them, so the reviewer can see types, signatures, and surrounding declarations beyond the diff hunks.
• Semgrep SAST findings produced from the same diff, when any are produced, so the reviewer can correlate static-analysis hits with the code under review.

Your exclusion patterns and other review configuration are applied on our side to decide which file changes reach the diff in the first place; the rules themselves are never sent to Anthropic. Pull request metadata such as the PR title, description, and source and target branch names is used by our orchestrator to route the review and create review records on our side — it is not included in the prompt.

The system instructions we send Claude are not customer-specific. They describe how to review code; they do not contain any tenant data.

What we don't send

• Files outside the diff. ADO Pilot does not clone your repository or walk your source tree.
• Files matched by your exclusion patterns or by the built-in defaults (binaries, lock files, generated code, minified bundles).
• Build artifacts, test fixtures, or pipeline outputs.
• Customer credentials. Personal access tokens, OAuth refresh tokens, and service-principal certificates are encrypted at rest, used only to call the Azure DevOps REST API on your behalf, and never sent to Anthropic.
• Identity data beyond what already appears in PR metadata. We do not enrich the prompt with user directory data or commit history outside the PR.

Prompt caching and how long content lives

ADO Pilot uses Anthropic's prompt-caching feature to keep review costs predictable. Cached content lives inside Anthropic's infrastructure for a bounded TTL and is scoped to ADO Pilot's API account, not shared across customers.

• System instructions: 1-hour TTL. Our review instructions are identical across every review, so we mark them with a 1-hour cache lifetime. They contain no tenant data.
• PR-specific content: 5-minute TTL. The diff, tree-sitter skeletons, and Semgrep findings are cached only long enough to bridge the first review pass and the self-critique pass. After 5 minutes of inactivity the cached entry is evicted.

ADO Pilot itself does not retain the diff after the review completes. Findings, the verdict, and usage metadata are stored in our database; the prompt content that produced them is not.

Anthropic's commitments

Anthropic processes the prompt to generate a response. Anthropic does not use customer prompts or completions to train its models. Anthropic's retention period for prompts and completions is governed by their Commercial Terms and the Data Processing Addendum (DPA) we sign with them; the contractual language is referenced from our legal page. Enterprise customers can opt into Zero Data Retention (ZDR) as an add-on — under ZDR, payloads are not retained beyond the request lifecycle. Contact sales for ZDR pricing and availability.

For Anthropic's own privacy commitments, see Anthropic's Privacy Policy and Trust Center.

Legal basis

We process your code on your behalf to deliver the review service you asked for — the legal basis is contract performance. ADO Pilot acts as a data processor; you remain the controller. The full processing terms, subprocessor list, and Standard Contractual Clauses for transfers out of the EEA, UK, and Switzerland are in the DPA, available from the legal page.

If you have a question about how a specific data class flows through ADO Pilot that this page doesn't answer, the tenant isolation page covers the storage and access boundaries on our side.