Data Privacy Considerations When Building AI Features for Regulated Industries
The first architecture we proposed for a healthcare client's AI intake assistant sent patient-described symptoms directly to a third-party LLM API, and it didn't survive the client's compliance review — not because the vendor's security was inadequate, but because their BAA didn't cover the specific data flow we'd designed, and the review process alone added six weeks we hadn't budgeted for.
We rebuilt around a de-identification step that stripped protected health information before anything reached the model, replacing names, dates, and identifiers with placeholder tokens that got mapped back after the response returned. It added a processing step and a small amount of latency, but it moved the entire LLM interaction outside the scope of several compliance requirements that would otherwise have applied.
Data residency came up in ways we hadn't fully anticipated for a financial services client operating across multiple jurisdictions — some customer data legally couldn't leave a specific region, which ruled out several API providers outright regardless of their security posture, and shaped our vector database and inference provider selection around region-locked deployment options from the very start of the project.
Logging is where privacy work quietly gets undone if nobody's watching. We've caught more than one project logging full prompts and completions for debugging purposes — including the sensitive data they were supposed to be protecting — in a system with far weaker access controls than the production database it was meant to help debug. Log redaction is now a standard line item in every regulated-industry project, not an afterthought.
The broader lesson: for regulated clients, the compliance review is part of the engineering timeline, not a gate at the end of it. Looping in the client's compliance or legal team during architecture design, rather than after a prototype is built, has consistently saved weeks compared to designing first and asking permission later.
