Farooque Munshi, Partner at EY. Leads Data/AI for Advanced Manufacturing across the Americas. Focused on turning AI ambition into outcomes.
Ask a chief data officer (CDO) to describe the company’s AI estate, and you’ll likely get the usual inventory. A predictive maintenance pilot, vision inspection on two lines and a copilot rollout in engineering, each with an owner, a business case and a slide in the steering deck.
However, if you walk the floor, you might discover a different answer. The machining center that arrived last quarter self-adjusts its tool offsets with an embedded vision model. The compressors run a vendor’s anomaly detection. The new spectrometer classifies defects using a model trained on other customers’ data, quite possibly including your competitors.
None of it appeared in the CDO’s inventory or went through model validation. It all came in the way a pallet jack does: through procurement, evaluated on price, delivery and warranty.
Most manufacturers now run two AI estates. One is deliberate, governed and presented to the board. The other is bigger, growing faster and invisible, assembled by category managers who have no idea they’re making architecture decisions, because nothing in the purchasing process tells them they are.
The model inside your grinding machine has a warranty card.
It’s getting harder to buy dumb equipment, and every embedded model has the properties that make AI a governance problem in the first place. It’s trained on data you can’t inspect, making decisions that affect your output, capable of drifting quietly. Your enterprise models have owners, monitoring and rollback plans. But for anything you bought this year, try answering the following:
• What was it trained on?
• When does the vendor retrain it, and do they tell you?
• What data does it phone home?
• If it scraps a shift’s output, can you audit why?
For your own models, the governance framework answers all of this. For the ones bolted inside equipment, the honest answer is usually, “We don’t know, and we signed nothing that lets us find out.”
Nobody screwed up, and that’s the problem.
Procurement is doing exactly what it was built to do. The request for proposal (RFP) template predates embedded AI. The contract playbook covers IP, liability and spare parts, not retraining cadence or data provenance. The governance function, meanwhile, was designed to catch models built in-house and software bought as software. Capital equipment lives in a different budget and a different approval chain. Both systems are working as designed, and the gap is between them.
This is why “procurement should loop in the AI team” isn’t enough. Just as developers weren’t going to consistently involve security teams on their own, procurement teams working against deadlines won’t reliably escalate AI-related questions voluntarily. The answer is to build the check into the purchasing process itself, so identifying embedded AI becomes a standard part of evaluating equipment rather than an extra step someone has to remember to take.
Here’s what it’s costing you.
Today, it is silent drift. Your material supplier changes or the vendor pushes an update overnight, and an embedded model starts making bad calls with the full authority of automation. There’s no baseline or audit trail. Operators get burned once and route around the intelligence permanently.
Then there’s data leakage. Vendors pool operational data across their installed base to improve their models and sell the improved model to everyone, including the plant across town. Years of hard-won process optimization is quietly averaged into the industry mean.
Regulation adds another consideration. While the EU AI Act’s high-risk obligations effective date has been postponed to December 2027, manufacturers should not assume they can wait until the rules apply to start identifying where AI is embedded in their operations. Some responsibilities fall on deployers, not just vendors, including requirements around human oversight, monitoring and record-keeping.
The new Machinery Regulation also brings certain AI-enabled safety functions into conformity assessment from 2027. “We didn’t know the machine had a model in it” isn’t a defense. It’s the kind of governance gap these requirements are designed to expose.
There are four ways you can fix this.
1. Count everything.
Walk the asset register and flag anything with adaptive control, vision, anomaly detection or optimization. Then ask vendors directly what AI capabilities are embedded in those systems. If a vendor is evasive, that is useful information in itself. Build a simple registry that captures the asset, what the model does, what decisions it can make, what data leaves the organization and how the model is updated or retrained.
2. Fix the contracts.
For anything with embedded intelligence, require the following:
• Disclosure of AI functions
• Notice before remote model or software updates, with the right to defer them
• Documented operating conditions and failure modes
• Explicit terms governing how operational data can be collected, used and shared
• Audit rights
Vendor resistance is information, too. A supplier that won’t tell you when it retrains the model controlling part of your process is quoting a price that doesn’t include the risk.
3. Add a tripwire, not a committee.
Put a simple routing rule into the capital-approval process: purchases flagged as AI-enabled generate a short disclosure in the registry, with deeper review reserved for systems that give a model meaningful authority over quality, safety or production decisions. The point is to embed the check in a process people already use, rather than create another approval layer that everyone learns to work around.
4. Treat them like assets.
Baseline embedded models when equipment is commissioned and monitor their behavior alongside the physical asset. When an AI-assisted system produces an unexpected result, make understanding why part of the standard troubleshooting process.
The goal isn’t to turn every machine operator into an AI engineer. It’s to make sure the organization can recognize when the intelligence inside an asset has changed and has a way to investigate it.
Conclusion
The governed estate with the framework and the steering committee may now be the minority of the intelligence running your operations. The majority came in through the loading dock. The good news is that this AI problem can be fixed with an inventory, five contract clauses and a routing rule.
Do this, and you’ll know what intelligence runs your plants, on whose data and under whose control. Don’t, and you’ll keep presenting your AI strategy to the board while the real one accumulates, unread, in a filing cabinet full of purchase orders.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
