# Governance that scales: running AI oversight across multiple plants
A predictive-maintenance model trained on vibration sensors at a Michigan engine plant gets copied to a sister plant in Monterrey. Nobody updates the risk log. Six months later, the Monterrey line has different machines, different failure modes, and a model quietly generating false negatives. No one owns the fix, because no one owns the model. This is not a hypothetical: it is the default failure mode when manufacturers scale AI from one pilot line to a global footprint without scaling governance alongside it.
Governance that works for one plant, one model, one champion, breaks the moment you have twenty plants, two hundred models, and a rotating cast of plant managers. This lesson covers how to build oversight that scales with deployment, not behind it.
A pilot usually lives under one engineer's desk. Approval is informal: the plant manager signs off, IT connects the data feed, and everyone moves fast.
That model collapses under three multi-site pressures:
Regulators and standards bodies increasingly expect manufacturers to treat AI risk the way they already treat quality risk under systems like ISO 9001: documented, owned, auditable, and tiered by consequence.
Scaled governance rests on three building blocks.
1. A model inventory. A single, mandatory register of every AI system in production or pilot, across every plant. Minimum fields: model name, purpose, plant(s) deployed, data sources, owner, risk tier, last validation date, next review date. If a model is not in the inventory, it cannot go live. This mirrors the "AI system inventory" expectation in the EU AI Act's obligations for providers of high-risk systems, and is good practice regardless of jurisdiction.
2. Risk tiering. Not every model deserves the same scrutiny. A practical three-tier scheme used by several industrial firms:
| Tier | Example | Oversight |
|---|---|---|
| Tier 1: Safety/compliance-critical | Weld defect detection feeding into pass/fail for structural parts | Full validation, human sign-off, quarterly audit |
| Tier 2: Operational | Predictive maintenance, demand forecasting | Documented monitoring, semi-annual review |
| Tier 3: Low-stakes | Internal chatbot for shift-scheduling FAQs | Lightweight logging, annual review |
Tiering determines audit frequency, documentation depth, and who must sign off, so effort concentrates where consequences are highest.
3. Named accountability owners. Two distinct roles, often conflated to the point of failure:
Both names go in the inventory. If a torque-check model misfires, everyone knows within minutes who investigates the model and who halts the line.
A workable structure has three layers, avoiding both a single overloaded committee and a lawless free-for-all:
This structure echoes how manufacturers already run NIST's AI Risk Management Framework functions (govern, mapmapUsing software to automate repetitive marketing tasks and campaigns, enabling personalisation at scale across channels like email, web, and social.View full definition →, measure, manage) but mapped onto a physical multi-site footprint rather than a single business unit.
Audits should not be a compliance afterthought. A minimal, practical cadence:
A simple drift check a quality engineer can run without a data science degree:
# Compare recent sensor readings to training baseline
import pandas as pd
baseline = pd.read_csv("training_data_baseline.csv")
recent = pd.read_csv("last_30_days_sensor_data.csv")
# Flag if the mean has shifted more than 2 standard deviations
threshold = baseline["vibration_reading"].std() * 2
drift = abs(recent["vibration_reading"].mean() - baseline["vibration_reading"].mean())
if drift > threshold:
print("ALERT: possible data drift, escalate to model owner for review")
else:
print("Within normal range")This kind of check does not replace a data science review, but it gives a plant-level owner an early warning signal between formal audits.
Knowledge check
1. In the Michigan-to-Monterrey example, what was the root cause of the model quietly generating false negatives?
2. Why does governance designed for a single pilot plant tend to break down at multi-site scale?
3. Why does the lesson argue that a vision-inspection model failing on a packaging line and a similar failure on a safety-critical aerospace torque-check should NOT be tracked identically?
4. Select ALL correct answers describing the multi-site pressures that cause single-plant AI governance to break down.
Select all the correct answers.
5. Select ALL correct answers about what regulators and standards bodies increasingly expect for AI risk management in manufacturing.
Select all the correct answers.
Manufacturers operating in the EU face the EU AI Act, which entered into force in 2024 with phased obligations through 2026 and beyond. Safety-related AI used in machinery (think AI-based safety controllers on a production line) can fall under "high-risk" classification, triggering requirements for risk management systems, technical documentation, human oversight, and post-market monitoring, obligations that mapmapUsing software to automate repetitive marketing tasks and campaigns, enabling personalisation at scale across channels like email, web, and social.View full definition → closely to the tiering and audit structure above.
In the US, there is no single federal AI law equivalent to the EU AI Act as of early 2026. Instead, manufacturers navigate a patchwork: OSHA (Occupational Safety and Health Administration) rules on machine safety still apply regardless of whether AI is involved, the FTC (Federal Trade Commission) has signaled scrutiny of deceptive or unsafe AI claims, and NIST's framework remains voluntary but is increasingly referenced in customer contracts and insurance underwriting. Multi-site US manufacturers exporting to the EU should assume EU AI Act obligations apply to any EU-deployed systems regardless of where governance is headquartered.
The practical implication: build your inventory and tiering system to the stricter standard (generally the EU AI Act's documentation depth) and it will satisfy lighter-touch US expectations by default, rather than running two parallel governance systems.
🎬 [VIDEO: "How Siemens Approaches AI Governance in Manufacturing" - youtube.com - search for recent Siemens or manufacturing-industry conference talks on scaling AI governance across plants, illustrating inventory and audit practices in a real industrial setting]
1. Freeze new AI pilots for two weeks and inventory everything already running, including unofficial spreadsheet models nobody flagged.
2. Tier every existing model using the three-tier framework.
3. Assign named model owner and business owner to each; escalate gaps to the corporate council immediately.
4. Set audit cadence per tier and put dates in a shared calendar, not an email thread.
5. Before any model is copied to a second site, require cross-plant review board sign-off confirming local revalidation.