Assumption: CCPA is a database and cookies problem
Reality: CCPA follows the data, not the storage layer. The moment your application feeds customer data into an AI model, an LLM API, or an automated decision system, a new set of obligations kicks in — separate from anything your database or cookie banner already covers. Working with mid-market companies across Southern California, we see this gap show up over and over. It isn't malicious. It's structural: engineering teams scope CCPA to the systems they think of as "storage," and the AI pipeline never makes the list.
Assumption: Calling a third-party AI API is just processing
Reality: it's a data transfer. When your application sends customer data to OpenAI, Anthropic, or any third-party API, CCPA requires you to disclose the transfer in your privacy policy, ensure the third party has adequate data protection agreements, and honor deletion requests across all systems — including any data retained by the AI provider.
Most companies we audit have AI features that send customer names, emails, and behavioral data to LLM APIs without updating their privacy policy or data processing agreements. This is low-hanging fruit for regulators.
Assumption: AI outputs are our data, not the customer's personal information
Reality: if your application uses AI to create customer segments, risk scores, or behavioral predictions, those outputs are "inferences" under CCPA — and inferences are personal information. Customers have the right to know what inferences you've drawn about them and the right to request deletion.
A SaaS client in Newport Beach was using AI to generate "customer health scores" for their B2B platform. Those scores were derived from usage patterns, support ticket sentiment, and payment history. Under CCPA, every one of those scores is personal information subject to access and deletion requests.
Assumption: Deleting an account deletes the customer's data
Reality: if you fine-tune models or build embeddings using customer data, that data doesn't disappear when the customer deletes their account. It lives inside your model weights or vector database. CCPA's deletion requirements mean you need a strategy for handling this — whether that's retraining models, maintaining deletion logs, or designing your pipeline to avoid persistent customer data in training sets.
The same mismatch applies to your deletion process more broadly. Teams assume it covers everything, but the primary database is only one destination. A compliant deletion has to reach vector stores, cached embeddings, conversation logs, and any derived data.
Assumption: Automated decisions are just a feature
Reality: if AI is making decisions that affect customers — credit approvals, pricing, service levels — CCPA's opt-out provisions for automated decision-making apply. You need a human review pathway for consequential decisions, not just a model output that ships straight to the customer.
What CCPA actually requires you to do
The fixes follow directly from the mismatches above:
Audit your data flow. Map every place customer data touches an AI system — LLM APIs, embedding databases, fine-tuning pipelines, automated decision systems. Most companies discover three to five undocumented data flows.
Update your privacy policy. Disclose AI-related data processing specifically. Generic language about "service providers" doesn't cover LLM API usage adequately.
Implement AI-specific deletion. When a customer requests deletion, your process needs to cover not just your primary database but vector stores, cached embeddings, conversation logs, and any derived data.
Add human review for consequential decisions. Where automated decisions affect customers, build the review pathway CCPA's opt-out provisions require.
Log everything. Maintain records of what data was sent to which AI systems, when, and for what purpose. When a regulator asks, you need to answer quickly and accurately.
Assumption: Enforcement is theoretical
Reality: California's AG and the new Privacy Protection Agency are actively enforcing CCPA, with a focus on AI and automated decision-making. Fines of $2,500 per violation add up fast when every customer record counts as a separate violation.
For a mid-market company with 50,000 customer records and a non-compliant AI feature, the theoretical exposure is $125 million. Regulators rarely pursue the maximum, but even a fraction of that is existential for a mid-market company.
The fix isn't expensive. The audit typically takes two weeks. The engineering changes take four to six weeks. Compared to the alternative, it's the best investment you'll make this year.