You're considering an AI assistant to draft emails, summarize support tickets, or generate sales insights. It promises to save your team hours. Then the question lands: where does our customer data actually go? Most founders gloss over this. They sign the terms and assume their data stays private. Sometimes it does. Often, it doesn't—not because the vendor is malicious, but because nobody set clear rules at the start. The problem isn't AI itself. The problem is letting AI access customer data without boundaries. An AI tool trained on your customer conversations, invoice history, or contact details can become a liability faster than you can update your privacy policy. Here are five guardrails you need in place before any AI tool touches a single customer record. 1. No training on your data—ever This is the most critical rule, and it's often buried in the fine print. When you feed customer data into a generic AI service—ChatGPT, Claude, Gemini—that data goes into the model's training pipeline by default. Your customer's invoice details, support history, or contract terms become part of a broader dataset. The model learns from it. Other users might eventually see patterns from your data. Some vendors claim they don't do this. Verify it. Specifically ask: "Does this tool ever use our data to train or improve the underlying model?" Get it in writing. Better: use AI tools that run on your own infrastructure or offer a strict "no training" contract clause. Enterprise-grade AI services (and many newer platforms built for business) offer this. Cheap consumer tools often don't. If the AI vendor profits from training on your data, your data isn't secure—it's an asset they're mining. 2. Automatic data deletion after task completion Even if the vendor promises not to train on your data, they'll still hold it. For how long? Set a hard rule: any customer data sent to the AI tool must be deleted immediately after the task completes . Not kept for 30 days "in case you need it." Not archived for analytics. Gone. This applies to: Conversation summaries passed to AI for draft responses Customer names, emails, or phone numbers used in prompts Invoice or contract details fed to AI for analysis Support tickets uploaded for sentiment analysis or categorization If the vendor won't commit to auto-deletion, you need to manually purge it from their system on a schedule. Document when you do this. Make it someone's responsibility. The same rule applies to your own infrastructure. If you build an internal AI workflow using Orin's built-in AI or a self-hosted model, configure it to process data in-memory and discard it after the operation completes. Never let customer data sit idle in logs or temporary storage. 3. Encryption in transit and at rest Data sitting on a vendor's server unencrypted is data a hacker can read in seconds. Require end-to-end encryption: your data should be encrypted on your device before it leaves, and decrypted only when needed by the AI tool. At rest, it should be encrypted with keys only your organization controls (or keys the vendor can't access). Ask these questions: Is data encrypted in transit (HTTPS/TLS)? (Most tools do this. Non-negotiable.) Is data encrypted at rest? (This is rarer. Many don't.) Who holds the encryption keys? (If the vendor holds them, they can read your data.) Does the tool support zero-knowledge architecture? (Only you hold the keys.) If a vendor can't explain their encryption model in clear terms, that's a red flag. 4. Zero access to PII or sensitive identifiers Some AI tasks don't need customer names, email addresses, or phone numbers to work. Strip them out before passing data to the AI. Example: if you're using AI to summarize a support ticket, remove the customer's email address from the prompt. The AI can still understand the issue without it. If you're using AI to categorize leads, pass job title and industry—not the customer's full name or LinkedIn profile. This principle is called data minimization : only send the AI the specific information it needs to complete the task, nothing more. Build this into your workflows: Create a data mapping rule: what fields are safe to send to the AI, and what fields must stay local? If you use Orin automations with AI, configure them to exclude PII by default. For every AI integration, audit the prompt. Remove unnecessary customer identifiers. This also protects you if the AI vendor has a data breach. If you never sent the customer's SSN or credit card number, it can't be leaked. 5. Clear audit trail and user consent When an AI tool processes customer data, you need to know: Who initiated the AI action? (Which team member, or which automation?) When did it happen? What data did it access? Why was it needed? (The specific business task.) Where is the output stored? This audit trail becomes critical if a customer asks how their data was used, or if you face a compliance review. You need to be able to say: "On January 15th, our support team used