Executive summary

• Manual data reviews cannot keep up with AI demand.
• Slow access pushes staff into risky workarounds.
• HIPAA gives two ways to de-identify data: Safe Harbor and Expert Determination.
• De-identified data stops being PHI and can move freely.
• Open access to safe data and keep humans on the critical calls.

Most healthcare data governance was designed for a world that no longer exists: a few analysts requesting a few datasets, reviewed one at a time by a committee that met every other week. Data governance is the set of policies, roles and controls that decide who can use which data, for what purpose and under what safeguards. That review-queue model held up when demand was low. It does not survive contact with AI, where every clinician, every model and every new feature wants data continuously. When we built data handling for a clinical AI platform, the first lesson was that the queue itself becomes the risk.

‍

Why do traditional data governance models break down at AI scale?

Traditional governance rationed access because manual review was the only control available. A committee could approve a handful of requests a month. At AI scale demand arrives continuously from every team and every model. A review queue that clears a few requests a week becomes a permanent bottleneck. The bottleneck does not stop the work. It pushes the work underground.

‍

The tell is the workaround. When people cannot get governed access fast enough, they build shadow pipelines: a nightly export to a spreadsheet, a copy of a dataset on a laptop, a one-off extract that never gets deleted. Each workaround carries more risk than the access it replaces, because it lives outside the controls entirely.

‍

On Care Culture Talks, Dr. Gezzer Ortega, assistant professor of surgery at Mass General Brigham, made the same observation about language access: "There are a lot of workarounds that exist in this space. That tells you the problem is real and the system has not solved it." The logic transfers directly to data. Widespread workarounds are not a discipline problem. Medical staff do what they can with the resources they have at the point of care. The workarounds are evidence that the governance model has failed to keep up with real demand.

‍

Traditional governance breaks at AI scale in four predictable ways:

• Review queues that cleared quarterly reports cannot clear continuous model-training and feature requests
• Access is binary, approved or denied, when AI workloads need graded access by data sensitivity
• Controls live in review meetings and email threads, not in the systems where data actually moves
• The people who understand the data and the people who approve access sit in different queues, adding days to every request

‍

‍

How are data leaders embedding privacy, security and compliance into a foundational data layer?

Leading data teams are moving controls out of the review meeting and into the layer where data lives. A foundational data layer is the governed platform that stores, catalogs and serves data with privacy, security and compliance built in as properties of the system rather than gates applied after the fact. When the controls travel with the data, access can be broad without being ungoverned.

What belongs in the foundational layer and not the application layer?

The foundational layer owns identity, access policy, encryption, audit logging and data classification. Applications and AI features consume data through that layer. If a control has to be re-implemented correctly in every application to stay compliant, it belongs in the foundation instead.

‍

How does the HIPAA Security Rule map to a data layer?

The HIPAA Security Rule requires administrative, physical and technical safeguards to protect electronic protected health information (ePHI), covering its confidentiality, integrity and availability (HHS OCR, 45 CFR Part 164). In a foundational data layer those safeguards become concrete: technical safeguards as encryption and access controls, administrative safeguards as documented access policy and risk analysis, physical safeguards as controlled infrastructure. Building them into the layer means every application inherits them by default instead of implementing them one at a time and getting one of them wrong.

‍

‍

What are HIPAA's two de-identification pathways and how do Safe Harbor and Expert Determination differ?

HIPAA's Privacy Rule recognizes two methods for de-identifying protected health information so it is no longer PHI: Expert Determination and Safe Harbor (HHS OCR, 45 CFR 164.514). De-identification is the process of removing or obscuring identifiers so data can no longer be tied to an individual. Once data is de-identified under either method, HIPAA obligations no longer apply to it, which is what makes de-identification the pivot point for opening access.

What is the Safe Harbor method?

Safe Harbor removes 18 specified identifiers from the data, including names, most geographic detail smaller than a state, all date elements more specific than a year and any other unique identifying number or code. It also requires that the covered entity have no actual knowledge that the remaining data could identify a person. Safe Harbor is one-size-fits-all: less subjective, usable without an expert and fast to apply. The tradeoff is utility, because stripping dates and geography removes signals that many AI models need.

‍

What is the Expert Determination method?

Expert Determination has a person with appropriate statistical and scientific expertise assess the data and certify that the risk of re-identifying an individual is very small, then document the methods used. The HHS guidance does not put a number on "very small," which leaves the judgment with the expert. This method fits the de-identification to the dataset, so it can preserve more of the detail AI workloads depend on. The cost is that it requires expertise and periodic reassessment, since re-identification risk changes as outside data becomes available.

‍

Which pathway fits AI workloads?

The choice is a tradeoff between speed and utility. Safe Harbor is the faster path for lower-risk sharing and for teams without a de-identification expert on hand. Expert Determination is worth the effort when a model needs the dates, ages or geography that Safe Harbor strips and when the volume of downstream use justifies the upfront work. Many mature programs run both: Safe Harbor as the default, Expert Determination for the specific datasets where utility loss would break the use case.

‍

How does de-identification apply to medical interpreting?

Language access data is a sharp example of why the pathway choice matters. No Barrier for instance, produces an interpreted clinical encounter which is free-text transcripts of what a patient said about their health, in two languages, which is some of the most sensitive data a system holds. Safe Harbor is blunt against free text: spoken names, dates and places have to be scrubbed wherever they surface in the conversation. 

‍

Expert Determination usually fits this data better, because a qualified person can certify that re-identification risk is very small while preserving the linguistic signal a model depends on. The speed-versus-utility tradeoff is the same. It just lands harder when the patient data is a conversation rather than a table.

Expert Determination usually fits this data better, because the expert certifies the method rather than reading every record. Under HIPAA the determination can cover "the method that generated such a data set" (HHS OCR), so a qualified person validates an automated de-identification pipeline, measures the residual risk statistically and documents that it is very small across the dataset. That preserves more of the linguistic signal a model depends on.

‍

‍

How do you democratize data access for AI without rationing it or losing control?

You classify data and tier access, rather than choosing between locking everything down and opening everything up. De-identified data flows freely to most users and most models. Identified PHI stays behind least-privilege access with full audit logging. Rationing, defaulting to deny, starves AI programs of the data they need. Ungoverned openness creates breach and compliance exposure. Tiering is the path between them.

‍

Centralized governance tools make tiering enforceable at scale: a data catalog so people can find what exists, policy-as-code so rules apply automatically instead of in review meetings, tiered access tied to data classification and audit logging that records every access for later review. The point of centralizing the tools is not to centralize the decisions. It is to make safe access the default so you do not have to ration it.

‍

Democratized access does not mean automated everything. Some outputs still need a human in the loop. The regulation is explicit about it. Under Section 1557 of the Affordable Care Act, when a covered entity uses machine translation for content that is critical to a patient's rights or care, that translation must be reviewed by a qualified human translator (HHS OCR, 2024). Human oversight is a governance control like any other. It works best when it is designed into the workflow up front.

‍

Ortega framed the same principle for policy. Healthcare tends to react to technology after the fact, when the better path is to get the technology, the health system and the clinicians in the room together from the start. Governance built proactively into the data layer is the same move. You decide where human oversight lives before the incident, not after it.

‍

This is not an abstract problem for us. Medical interpreting generates some of the most sensitive data in a health system: utterance-level records of clinical conversations, plus the audit logs that prove what was said. Serving interpreting for the 29.6 million people in the US with limited English proficiency (GSA, December 2025) means living inside this governance question every day. No Barrier is built on it: end-to-end encryption, SOC 2 Type II certification, a HIPAA posture with a standard BAA, PHI deletion after seven days and human oversight designed into the workflow rather than added after a problem. More detail on that posture is on our FAQ. 

‍

The bottom line

The health systems getting this right stopped treating governance as a gate and started treating it as a foundation. De-identification is the mechanism that lets most data move freely, tiering is how you avoid rationing and human oversight is where you keep control of the decisions that matter. Build those into the layer where data lives and access stops being a queue.

‍