The Compliance Gap Every Manufacturer Running AI Needs to Know About

Most manufacturers running AI systems assume their existing compliance frameworks cover them. SOC 2, ISO 9001, IATF 16949 — pick your audit. The assumption is that if the compliance stack is clean, the AI is governed.

It isn't.

SOC 2 audits your infrastructure security. Your quality system audits process consistency. Neither one was built to answer the question that matters most when you're running AI in production: is your system doing what you think it's doing, and what happens when it's not?

This week, we're breaking down exactly where the gap is and what to do about it.

---

Industry Spotlight: The Automotive Supply Chain

The automotive supply chain has become a live case study in AI governance failure modes.

Tier 1 and Tier 2 suppliers are running computer vision systems on stamping, welding, and assembly lines. Most of these systems were implemented during the last three years, trained on historical data, and validated against controlled test sets. They passed validation. They went live.

Here's what the validation process typically didn't account for: material specification changes from suppliers, tooling wear patterns that gradually shift part geometry, lighting degradation in the inspection environment, and the long tail of edge cases that only appear at production volume.

The result is model drift that no existing compliance framework flags. The quality management system audit checks whether the inspection process exists and is documented. It doesn't check whether the AI component of that process is still calibrated to current production conditions.

I've seen this play out in a specific and consistent way: the model continues to report high confidence scores while passing parts that would have failed incoming inspection under the previous regime. The number looks fine. The outcome isn't.

The manufacturers catching this early share one characteristic. They measure AI performance against operational outcomes — not just model-reported accuracy — and they've assigned someone with real accountability for monitoring that gap. That's NIST AI RMF MEASURE 2.5 and GOVERN 1.1 in practice. Not compliance theater. Operational discipline.

---

Framework Deep Dive: NIST AI RMF for Manufacturers

NIST AI RMF has four functions. Here's what each one actually means for a manufacturer with AI systems in production.

GOVERN

GOVERN is where most manufacturers are most behind. It's not about policies for their own sake. It's about ensuring that before an AI system influences a production decision, someone with authority has answered the hard questions: what failure modes are acceptable, which decisions require human sign-off, and who is accountable when performance degrades.

GOVERN 6.1 specifically addresses organizational roles and responsibilities. In manufacturing, this means explicit documentation of who owns each AI system's performance — not the vendor, not IT by default, not the engineer who championed the implementation. A named person with the operational context and authority to act.

MAP

MAP is your AI inventory plus context. MAP 1.1 requires understanding intended use and deployment context. MAP 2.2 requires categorizing systems by risk. The categorization that works in manufacturing is consequence-based:

• Tier 1 — Safety implications and autonomous control
• Tier 2 — Quality outcomes and regulatory exposure
• Tier 3 — Internal efficiency with low-consequence errors

The same rigor doesn't apply uniformly. A scheduling optimization tool and a robotic safety interlock system are not the same governance problem.

MEASURE

MEASURE is where SOC 2 comparisons break down completely. SOC 2 measures controls. MEASURE asks whether the AI system is performing as intended and detects when it isn't.

MEASURE 2.5 pushes for operational performance metrics, not just vendor accuracy benchmarks. MEASURE 4.1 establishes ongoing monitoring cadence calibrated to risk tier. For a Tier 1 system, that means real-time monitoring with automated alerts. For a Tier 3 system, monthly review is defensible.

MANAGE

MANAGE closes the loop. MANAGE 1.3 requires pre-defined risk response plans for AI failures. MANAGE 2.2 covers AI incident response specifically.

The critical implementation point: AI incidents are not IT incidents. When a quality inspection model starts passing defects, the root cause investigation requires AI/ML competency, not just infrastructure diagnosis. MANAGE 3.1 requires feedback loops from the people using AI systems. Your operators know when outputs are wrong. Build a formal channel for that knowledge to reach whoever owns the system.

The practical sequence that works for mid-size manufacturers: inventory first, categorize second, build measurement third, operationalize response last. Don't try to implement all four functions simultaneously. Start with your one highest-consequence system and build the pattern there.

---

This Week in AI

EU AI Act conformity assessments are increasingly relevant for manufacturers exporting to European markets. "High-risk" under that framework includes AI used in safety components of products and AI systems that are safety components themselves — language that maps directly onto predictive maintenance and process control applications common in manufacturing.

If you're running NIST AI RMF implementation, the governance structures you're building aren't just domestically relevant. The documentation and accountability requirements align closely with what EU AI Act conformity assessments will require. One framework, two compliance use cases.

On the vendor side, several major MES and ERP platforms have added AI-driven features in recent update cycles without corresponding updates to their documentation about how those features work, what data they use, or how performance should be monitored. If you've updated platform software in the last twelve months, it's worth reviewing release notes specifically for AI or ML language. Features you didn't consciously adopt may already be influencing production decisions.

---

Action Item This Week

One concrete thing to do: pull a rough AI system inventory.

Don't aim for perfect. Aim for complete enough. List every system, tool, or platform that uses ML or AI components — including SaaS tools, ERP modules, and anything a business unit may have procured independently of IT.

For each system on the list, write down one thing: what decision does this system influence, and what's the consequence if that decision is wrong?

That exercise alone will surface your governance priorities. The systems where you can't quickly answer the consequence question are the ones that need immediate attention. Most manufacturers find two or three systems they weren't thinking about as AI governance priorities that turn out to be their highest-risk exposure.

Keep the list. It's the foundation everything else in NIST AI RMF MAP builds on.

---

Chase Sutphin is the founder of AI Governance Solutions and an enterprise security engineer at Fortinet. He helps organizations implement the NIST AI Risk Management Framework through expert-led consulting and a purpose-built GRC platform.