Only ingest data fields your agent actually needs to complete its task. A support agent needs ticket content, customer name, and account status — not the customer's full payment history, social security number, or personal address.

Redact PII from logs and traces. When debugging agent runs, you need to see tool inputs/outputs — but customer email addresses, phone numbers, and financial data should be masked. Set retention limits on conversation history: 90 days for support agents, 30 days for transactional agents, configurable per customer for enterprise.

Data minimization is not just compliance — it reduces your liability surface and inference costs. An agent that processes 5 fields per task is cheaper and safer than one that ingests entire customer records 'just in case.'

Your agent product uses subprocessors — OpenAI, Anthropic, Google, AWS, Vercel, Supabase, and others. Enterprise customers will ask for your subprocessor list in the first security review. Have it ready.

Disclose all subprocessors in your privacy policy using a Data Processing Agreement (DPA) using standard templates (Common Paper, TermScribe, or your lawyer's template). For healthcare customers, you need a Business Associate Agreement (BAA) — confirm your LLM provider offers one (OpenAI and Anthropic do for enterprise plans).

Maintain a subprocessor change notification process: email enterprise customers 30 days before adding a new subprocessor. This is a GDPR requirement and an enterprise expectation. A simple subprocessor page on your website prevents weeks of back-and-forth during security reviews.

For regulated outputs — legal advice, medical recommendations, financial guidance — your agent assists licensed professionals; it does not replace them. Build this into your product UX and terms of service explicitly.

Require human review before any regulated output reaches an end customer. Display disclaimers: 'This draft was generated by AI and should be reviewed by a qualified [professional] before use.' Log who reviewed and approved each output for audit purposes.

This is not a limitation — it is a liability shield and a sales enabler. Legal firms buy 'AI research assistant' products. They do not buy 'AI lawyer' products. Medical practices buy 'AI documentation assistant.' They do not buy 'AI doctor' products. Position as assist, not replace, and regulated industries become addressable markets.