Data Privacy: When Local Models Are Non-negotiable

When you send a prompt to a frontier model, the prompt travels over the internet, hits a data center, and is processed by hardware you don't control. The provider's privacy policy says what they do with that data. Most major providers say they don't train on your data (especially on paid tiers and via the API), and they typically retain data for 30 days for abuse monitoring before deletion. Read the specific policy of the specific service you're using; it's not all identical.

For most personal and small-business work, this is fine.

For some work, it isn't.

DATA THAT SHOULD NEVER GO TO A FRONTIER MODEL VIA UNCONTROLLED PROMPTS:

• Personal Health Information (PHI). HIPAA-regulated data. Patient names, diagnoses, medications, treatment plans.
• Personally Identifiable Information (PII) at scale. Lists of customers' Social Security numbers, addresses, dates of birth.
• Financial data covered by regulation. PCI (credit card numbers in plaintext), bank account numbers, SOX-regulated financial statements pre-publication.
• Trade secrets. Source code or formulas you're contractually or legally obligated to protect.
• Customer data covered by NDAs or data-processing agreements. Many B2B contracts prohibit sending the customer's data to third-party AI services.
• Government data with classification or controlled status. Anything CUI, FOUO, or higher.

For these categories, you have three options:

OPTION 1: Use a frontier model with the right contracts in place. Anthropic, OpenAI, and others offer enterprise tiers with HIPAA BAAs, SOC 2 attestations, zero-data-retention modes, and so on. If you have the volume and the budget, this is the cleanest path. Not all data categories are eligible. Government classified data, for instance, generally isn't.

OPTION 2: Use a local model. The model lives on your hardware. The data never leaves the room. Capability is lower than frontier (see section 4.1), but for many routine tasks it's enough.

OPTION 3: Don't use AI for this work. Some data shouldn't go through any model, local or frontier, because the human review process and audit trail matter as much as the answer.

The decision tree:

1. Is this data covered by a regulation, NDA, or contract? YES → Local or enterprise tier. Don't use consumer chat interfaces.
2. Is this customer data, even if not regulated? YES → Default to local, unless the customer has agreed to AI-assisted processing.
3. Is this internal-only business data? OK to use frontier on most paid tiers. Read the privacy policy to confirm.
4. Is this public information or your own work? Frontier is fine.

For small businesses, the practical answer is often: a Mac mini in the office running a local model, used for any agent task that touches customer data. Frontier models for everything else. You get both worlds without breaking your contracts.

A separate frame on the same problem: most data covered by these rules is also AFD (see §1.4). PHI, PCI, customer data under NDA are not just regulated; they're categories where being wrong hurts people who aren't you. The decision tree above is the operational answer. The AFD frame is the ethical answer. They mostly point the same direction.