A proposed enterprise function for the responsible adoption, lifecycle, and oversight of artificial intelligence across PACS — built on the knowledge that already powers our care.
Before we propose what to do about it, a snapshot of what is happening in our environment today — across Copilot, Studio agents, vendor AI, and the ways AI is being used by staff that no one has officially sanctioned.
Before asking PACS to audit anyone else's AI use, an audit of my own — and a demonstration of the workflow this function would apply at scale.
Three patterns I have used to make my own work more effective, each of which would fail the governance review I am proposing — and each of which I have already remediated, or am in the process of moving, into PACS-sanctioned tooling.
This is a working example. It demonstrates how the function operates: identify, classify, remediate, document. The pattern is the same whether the subject is the governance lead or anyone else.
Before any AI tool touches PACS data, the BAA question has to be answered. This is the source-of-truth map. Covered means the Microsoft Online Services DPA / Business Associate Amendment reaches the surface; not-covered means no PHI, period.
Every entry below reflects PACS' current licensing posture and Microsoft's published BAA-applicable services list. Status indicators are deliberately coarse — covered, not covered, or conditional. Tier and SKU notes are surfaced inline because the same product name often carries different protection depending on the SKU (consumer vs. enterprise vs. free).
Every Copilot agent we build, every Studio workflow we deploy, every clinical or operational LLM we eventually permit — all of them are downstream of one thing: the quality, structure, and governance of the knowledge they read.
PACS's knowledge base, SharePoint architecture, ServiceNow configuration, and document metadata are the substrate on which our AI will operate. If those sources are inconsistent, unindexed, or stale, every AI tool built on top inherits and amplifies their defects — at the speed and scale of automation.
This is why AI governance and knowledge governance cannot be separated. They are the same function, viewed from two angles. The proposal that follows treats them as one.
AI governance is not a future regulatory possibility for PACS. Multiple federal and state regimes already apply to AI as used in healthcare today. A summary of the most material — and the specific PACS control gap each one creates.
Four public AI failures from the past 24 months — each preventable with a governance function in place. None of the affected organizations had one.
Three existing functions touch AI at PACS today, each with a specific and necessary scope — and none of them owns the operational lifecycle of AI being built inside our walls.
Scope: Does PACS admit this AI vendor?
Scope: Is PACS legally compliant?
Scope: Identity, network, endpoint, threat.
PACS will not invent its own framework. The proposed function aligns to the same standards Workday, JPMorgan, IBM, and peer regulated enterprises align to — published, voluntary, increasingly cited by HHS, and certifiable.
The proposed AI & Knowledge Governance function operates across four operational pillars — bounded, named, and individually measurable.
A multiplier, not a bottleneck. Governance scales with risk — self-service on low-risk productivity, gated review on PHI- or SOX-touching automation.
The seven-stage operational SOP every new agent passes through. Same workflow whether the requester is a field engineer asking for a help-desk assistant or the CFO asking for a forecast bot.
The single most useful artifact for day-to-day decisioning. Read a row left-to-right to know which AI platforms are admissible for which class of PACS data. Framework-driven, not preference-driven.
Four layers, adapted from Microsoft's Cloud Adoption Framework for AI. Each layer answers a different question. Layers 1–3 govern every AI surface at PACS — Microsoft, third-party, or custom. Layer 4 is the build surface for agents we own.
Most of what we need is already licensed. The unfilled boxes are deliberate — they identify the explicit spend decisions ahead, not gaps left by omission. PointClickCare is the named architectural gate because it sits inside Layer 1 (data governance) but is not natively addressable by Purview today.
Pause · Questions · Stretch
We are halfway. A good moment for a refill, a stretch, or any questions on the framework, the regulatory floor, or the coverage gap before we move into structure and people.
The AI keeps going. So does David. You don't have to.
The engineering restructure is independently valuable — and necessary regardless of the AI governance proposal.
Not a department — at least not yet. One full-time lead, four task-force members carrying primary roles, and a cross-functional review board that meets monthly. Designed as Stage One. Y2 re-evaluation against KPIs and workload determines whether the function graduates to a formal department with named FTEs.
Operational ownership of all four pillars. 40+ hrs/week dedicated. Backfilled in engineering by restructure.
Carry existing positions; contribute to standards, agent reviews, and pillar-specific work without leaving current scope. Hold Entra roles: AI Administrator & Agent ID Administrator (scoped, audit-trailed, least-privilege).
Convenes for medium-/high-risk reviews and quarterly governance posture. Not standing committee meetings.
The reporting line for the proposed Director of AI & Knowledge Governance role is genuinely a leadership decision. Four credible options; tradeoffs surfaced; one primary recommendation grounded in the Workday case study and SOX SoD principles.
Governance capability builds in two parallel tracks: people (credentials, training, hires) and tools (tenant settings, policies, agent registries, environments). Click any phase to see which milestones land where — and what each unlocks for the other track.
Four time horizons. The first 30 days are already underway — NIST AI RMF training scheduled, agent inventory audit in progress, this proposal itself an artifact.
Six Y1 KPIs the function commits to. Public dashboard, quarterly review.
The body of this proposal addresses the governance question itself. Three adjacent considerations — credentialing, organizational placement, and timing — sit just outside that scope and are worth holding in mind. Tap each card to see the context.
This is true across PACS today — no current employee holds an AI governance credential. The role and the credential are emerging together at the industry level.
The proposal pairs internal practitioner fluency with a published credentialing path: NIST AI RMF training May 26 · IAPP AIGP within 90 days · ISO 42001 Lead Implementer by Y2. External hires arrive credentialed but require 12–18 months to absorb PACS-specific architecture (Copilot Studio agents, parent/child structures, evaluation sets, Graph indexing).
Both paths are defensible. The choice is between fluency first, credentials on a timeline and credentials first, fluency on a timeline.
SOX 404 ITGC contemplates segregation of duties between developers and governance. The same principle that separates check-writing from check-approval applies here.
The Workday case study published by NIST explicitly separates developer reporting from governance reporting for this reason. A governance function inside the team it governs is structurally constrained from saying "slow down" when deployment falls short of standards.
This is the structural argument behind Option A in Section 17 — peer-level placement rather than embedded placement. The two functions are complementary, not interchangeable.
The engineering restructure and the AI governance proposal are independent. Either can be approved, deferred, or declined without affecting the other.
The regulatory calendar is on its own clock: Section 1557 AI provisions have been enforceable since May 1, 2025; the HIPAA Security NPRM contemplates AI inventory; OCR enforcement activity is ongoing. The substantive question is sequencing and pace — not whether.
Starting organized while the volume is still manageable is materially different from starting after an incident. That is the trade-off worth holding.
Each decision is independent, reversible, and defensible. Approve, discuss, or defer any item. Approvals fill three outcome buckets — Governance Framework, Knowledge Base, AI Governance Role. These three asks mirror the live deck and the decision memo.