78% of manufacturers are not confident their AI models could pass an independent audit, according to Grant Thornton's 2026 survey.
Most have no way to tell whether the predictive maintenance model running on their floor is still right, or has been wrong since the last supplier change.
Before we get into how to catch it, quick question:
How does your plant monitor AI model performance after go-live?
👋🏻 I'm Leonardo Ubbiali. This week we're looking at the AI failure mode that does not crash, and what GlobalFoundries built to catch it.
GlobalFoundries' Fab 7 in Singapore was named a WEF Lighthouse Factory on September 16, 2025, after deploying more than 60 4IR use cases.
Labour productivity improved 40 percent and new product introduction prototyping time improved 30 percent.
Yew Kong Tan, Senior Vice President and General Manager of APAC Manufacturing and Singapore Site at GlobalFoundries, said the milestone was about "advancing the adoption of Industry 5.0 from 4.0" at GF's AI Centre of Excellence in Singapore, reshaping the workforce with digital solutions and building a robust digital ecosystem through strategic partnerships.
What most plants are not copying is the structure underneath those 60 use cases.
What's The Structure Like
Each model has a named owner and a drift threshold set at go-live. Population Stability Index above 0.25 is the working default across most ML teams.
Retraining is tied to recipe changes and material changes rather than a calendar cycle.
Incident response plans are written and tested before the model scales.
The peer-reviewed University of Sheffield AMRC paper published January 30 makes the case for why this matters.
If the model is a black box then the drift is invisible.
And if the model can show which sensor readings it relied on, a shift in those features is the early warning.
SHAP and Grad-CAM are among the techniques the paper applies across casting defects, metal surface defects, and acoustic anomaly detection.
The defense in every case is human review tied to feature-attribution signals.
I have not seen many plants outside the Lighthouse cohort that have this structure in place.
Grant Thornton surveyed 950 senior business leaders for its 2026 AI Impact Survey.
7 percent of manufacturers have a tested AI incident response plan.
78 percent lack full confidence they could pass an independent AI governance audit within 90 days.
50 percent of operations leaders say they need a formal AI governance plan in place within the next six months.
Tom Puthiyamadam, Managing Partner of Advisory Services at Grant Thornton, says AI deployment is simply outpacing the infrastructure that supports it.
That is the gap GlobalFoundries closed and most plants have not started on.
Anyone scaling AI past the second use case this year should expect drift in the first one.
Five Things You Can Do This Quarter
The problem: You need to know whether the AI models running on your floor are decaying.
What you need: A list of every AI model in production at your plant, when each was last validated or retrained, what data sources feed it, and what events could have shifted those inputs.
The Prompt (copy this):
I'm a [YOUR ROLE] at a [FACILITY TYPE] manufacturer. The following AI models are running in production at our plant: [LIST EACH MODEL, ITS PURPOSE, WHEN IT WAS LAST VALIDATED, AND ITS DATA SOURCES]. Recent changes to inputs or upstream conditions include [LIST RECIPE CHANGES, SUPPLIER CHANGES, SENSOR RECALIBRATIONS, OR PROCESS CHANGES IN THE LAST 6 MONTHS].
Tell me: Which of these models is most likely to be drifting based on the input changes I described? What drift-monitoring metric should I implement for each one? What retraining cadence should I tie to which event? How should I structure an incident response plan for an AI model that is silently producing wrong outputs?
What you'll see:
A drift-risk ranking of your production models, a recommended monitoring metric for each, a retraining schedule tied to your upstream events, and an incident response template.
NIST AI Risk Management Framework Playbook
The Playbook breaks the four AI RMF functions (Govern, Map, Measure, Manage) into specific actions a plant can run against. The Measure and Manage sections cover the drift monitoring and incident response work the body argues for.
Time to value: 30 minutes
GlobalFoundries built the structure over 60 use cases. When did your team last check whether the first one is still right?
Hit reply. I read every email.
Leo