A No-Copy PHI Cloud Environment That Cleared AI Governance and Funded Research Through Year 3

Challenge

A national hospital system with a longstanding academic research partnership wanted to accelerate clinical research using modern AI, but the path ran straight through a walled, security-heavy environment: strict PHI controls, an internal data-science organization, AI-governance committees, and analytics workflows bound to legacy SAS. The decision that had to hold up was not a model choice; it was an operating pattern that let AI touch real clinical data without PHI ever leaving the client’s network, and that committees could approve on their own terms. Any approach that relied on copying data out, or on a black-box model making decisions, would not survive review.

Approach

We stood up a HIPAA-eligible cloud analytics environment (GCP Vertex AI and BigQuery, with AWS Bedrock and PHI guardrails) built on the client’s own standardized consumer-project pattern, then proved an AI-augmented workflow on an internal test case before extending access to the academic partner. The AI was confined to reading and language (extracting structured signal from unstructured note text and drafting); the analytic decisions ran as deterministic, versioned code against governed tables, so results were reproducible and auditable.

• Used authorized views with policy-tag, column-level PHI masking instead of data copies, so PHI stayed inside the client’s network and never moved to a separate analytics store.
• Confined the LLM to extraction and language at build time (pulling structured fields from unstructured clinical notes), while the cohort logic, mortality and payor analyses ran as deterministic, versioned code that any reviewer could re-run.
• Validated the pattern end to end on an internal test case: reanalysis of published studies plus a sepsis and payor-mortality analysis with unstructured-note text extraction, so the method was proven before the academic partner was given access.
• Authored a combined infrastructure-plus-project SOW that became a de facto AI playbook for the program, including a tiered cost model so spend stayed predictable as usage grew.
• Built a governance and language framework written to clear the client’s AI review committees, framing what the AI does and does not decide in terms the committees could approve.
• Phased the program deliberately (prove internally, then extend), so each access expansion rested on an already-validated pattern rather than a new exception.

Impact

• Master agreements executed, establishing the contractual base for a multi-year program rather than a one-off pilot.
• The HIPAA-eligible cloud environment was cost-bounded to roughly $1,000 to $2,000 per month, low enough to run without finance-approval friction (figure reflects the scoped consumption envelope, not a fixed ceiling).
• The AI-augmented workflow pattern was validated end to end on the internal test case, including unstructured-note extraction feeding deterministic analysis.
• A multi-year research collaboration was funded into Year 3.
• An expansion pipeline opened with the CMO office and corporate analytics, moving the relationship beyond the original research partnership.
• In parallel, we scoped enterprise AI-enablement for business analysts oriented around reading, QA-ing, and guiding AI output rather than writing Python, a path that fits a historically SAS-bound analytics organization.

Capabilities demonstrated

• No-copy PHI architecture: authorized views with policy-tag, column-level masking so data stays in the client’s network.
• Deterministic-by-design AI: LLM scoped to extraction and language, decisions run as versioned, auditable code.
• AI governance enablement: a governance and language framework written to clear enterprise AI review committees.
• Combined infrastructure-plus-project SOW that doubles as a reusable AI playbook, with a tiered, predictable cost model.
• Phased delivery that proves a pattern internally before extending access to outside partners.
• Practical analyst enablement for a legacy-SAS organization moving toward AI-assisted work.

Anonymized by design: client names stay off the narrative per our reference policy. Figures that are modeled, small-sample, or targets are identified as such above.