A midwest health insurance plan learned this the hard way last year. They’d embedded Microsoft’s AI into their utilization management workflow, only to discover that Texas HB 20’s algorithmic disclosure requirements couldn’t be met through Azure’s API.
The vendor’s roadmap offered no fix for 18 months. Their choice? Halt deployments in Texas or risk $10,000-per-day penalties. They chose a third option: rip-and-replace, for $3.2 million and six months of lost innovation velocity.
Their mistake wasn’t picking the wrong vendor, it was surrendering their sovereignty.
The world changed in November 2022. Generative AI isn’t another COTS system to purchase; it’s a generational paradigm shift that demands you think like a software company or lose your destiny to proprietary black boxes. But the real threat isn’t just vendor lock-in, it’s the emerging 50-state regulatory civil war.
The regulatory mirage: A "one rule" vacuum
At least 23 states and DC have enacted bills regulating AI in healthcare, and they aren't coordinating. Instead of a single standard, we are seeing conflicting archetypes emerge that make national compliance impossible for static systems:
- California (The safety mandate): CPRA amendments now regulate automated decision-making in healthcare claims, requiring "meaningful human review pathways" that many embedded AI models don't natively expose.
- Texas (The Black box ban): HB 20 mandates full algorithmic audit trails and neutrality, conflicting directly with the proprietary "black box" nature of most cognitive services.
- New York (The reporting burden): Proposed insurance law § 3209-b would require AI model cards to be filed with the Department of Financial Services - transparency that no legacy vendor currently provides.
- Colorado (The documentation trap): SB 4 requires developers of "high-risk" AI systems to provide extensive documentation, allowing for impact assessments, shifting the burden directly to the deployer.
The danger isn’t just the strictness of these laws - it’s their incompatibility. A model tuned to satisfy New York’s bias requirements might technically violate Texas’s neutrality statutes. If your strategy relies on a rigid, hard-coded vendor implementation, you are one legislative update away from non-compliance.
The federal gap: While the recent Executive Order ("EO 14365") aims to create a "minimally burdensome national standard" and tasks the DOJ with challenging state overreach, the reality on the ground is chaos. Until the courts resolve these challenges, state laws remain in effect. You cannot architect for a future federal preemption that may never come; you must architect for the messy reality of today. For a video from Eric on federal and state dynamics in AI, click here.
The TEFCA trap: Why integration-first compliance fails
When the Trusted Exchange Framework launched (see Marc’s previous post on the subject), healthcare organizations treated compliance as an integration project—custom connectors for each state’s HIE.
- Average cost: $1.8M per state.
- Timeline: 14 months.
- Failure rate: 68% abandoned due to maintenance overhead.
The same pattern is emerging in AI. Teams are building state-specific API wrappers around vendor AI, creating a brittle patchwork. When Virginia updates its privacy law, 23 application teams scramble to patch their wrappers. This is the "Integration Trap." You cannot API your way out of a regulatory flood; you need a dam.
Architecting for sovereignty: The platform defense
Platform engineering flips the model. Instead of asking developers to master 50 legal regimes (shifting left), we make the platform smarter (shifting down). By abstracting the "legal state" from the "application state," we decouple innovation from regulation.
Quantified Advantage: When a New Mexico law changed last quarter, one platform team updated a single OPA policy in 90 minutes. All 12 AI applications inherited the change automatically. Estimated savings: 400 engineering hours and $180K in avoided rework.
1. Policy-as-Code as legal firewall
We must move beyond static governance documents. Use Open Policy Agent (OPA) to encode jurisdictional rules as configuration, not code. California’s privacy mandates become a Rego module; Texas data restrictions are a separate policy bundle.
When a state updates its law, the platform team updates the policy once at the gateway level. All applications inherit the compliance change instantly, without a single line of application code being rewritten. This turns compliance from a bottleneck into a background service.
2. Golden paths with escape hatches
Platform engineering reduces cognitive load by providing standardized, compliant workflows - "Golden paths." For example, a "HIPAA-compliant chatbot" template that comes pre-configured with PII redaction and audit logging.
But these must be paths, not cages. Expose parameters for edge cases: when a specific regulatory nuance requires bypassing the standard model, developers can toggle an “audit-intensive mode” that triggers legal review workflows automatically. This keeps the 90% use case fast while managing the 10% risk.
Predicting the puck: Even if state configuration laws get struck down, states will retain authority over the "scope of practice" - how AI provides clinical care versus administrative tasks. By building models with fundamental traceability and disclosures now, you insulate yourself against future liability shifts regarding who owns the harm - the developer, the physician, or the AI.
3. The Internal Developer Portal as sovereignty interface
Tools like Backstage become the single pane of glass for discovering sovereign-compliant AI resources. This isn’t just for humans - it’s the “internal developer portal for agents.” It acts as the chain of custody, logging exactly which model version was used, what training data it accessed, and which state policies were active at the time of inference. This traceability is your primary defense in a courtroom.
AI-specific technical risks you’re ignoring
Sovereignty isn’t just about where the data lives. It’s about technical control over the model's behavior:
- Model governance: How does your OPA policy enforce constraints on model cards or RLHF alignment decisions?
- AI security: Standard security tools miss prompt injection, model inversion, and data poisoning—attacks that could force regulatory disclosure under state breach laws.
- The Open-Source spectrum: If you reject OpenAI to avoid lock-in, what is your stance on Llama 3.1 or Mistral? Sovereignty exists on a spectrum; "pure build" isn't always optimal, but "pure buy" is always risky.
The sovereignty maturity model
Where does your platform stand today?
- Level 1 (Vendor dependent): Using embedded AI; regulatory changes require a vendor roadmap. Cost: $2-5M per unexpected regulatory shift.
- Level 2 (Reactive): Custom wrappers per application; fire-drill updates.
- Level 3 (Platform sovereign): Policy-as-code, golden paths; compliance updated in one place. Developers move fast because guardrails are baked in.
The forcing function: Cyber insurance & quantified risk
Executives don't care about OPA policies; they care about liability. To get buy-in, you need to speak the language of "Regulatory Lag."
The quantified risk:
- California: Non-compliance costs up to $7,988 per intentional violation.
- Texas: Breaches cost up to $50,000 plus $100 per individual per day.
If a proprietary vendor takes 90 days to update their model for a new Texas transparency law, your organization is exposed to those fines for 90 days. Now multiply that risk by 50 states.
- The cost of sovereignty: A one-time investment in a platform you control.
- The cost of dependency: A compounding liability that grows with every new state bill.
Sovereignty isn't a tech refresh; it is an insurance policy. Insurers are already asking about third-party tooling, incident response plans, and patch discipline. If you can't demonstrate rapid kill-switch capability, insurers will treat your GenAI initiatives as uninsurable amplifiers of risk.
The human element: Shadow AI as regulatory exposure
The genie is out of the bottle. Treating developers as customers isn’t just about productivity - it’s about preventing sovereignty leaks. When your legal team says “pause Texas deployments,” but the Epic AI button is still right there in the clinician’s workflow, they’ll use it.
Your first step this week: The sovereignty audit
Don’t build a platform yet. Don’t write code. Start with an audit that maps the real world, not the policy world.
- Inventory: Identify every AI entry point (EHR buttons, browser copilots, IDE assistants).
- Verify: For each tool, map what data sends (PHI/PII), where it’s processed, and what controls exist.
- Detect Leaks: Look for unmanaged accounts, personal API keys, and client-side SDKs that bypass your gateway.
- Close the Loop: Check for evidentiary artifacts like centralized logs. If you can't produce an audit trail, you are relying on hope.
(Contact Marc Mangus for a full list of sovereignty audit questions: https://www.linkedin.com/in/marcmangus/)
