Building a safer path to autonomous industrial AI

In an MIT Technology Review Insights podcast produced in partnership with AVEVA, the company's chief technologist, Arti Garg, argues that industrial AI is entering a riskier phase. Foundation models, physical AI and agents can now automate more complex work in plants, power systems and mines, where an unexpected decision has physical consequences.

Her central point is one we agree with: newer AI systems are harder to predict and explain, and their behavior can change as they are tuned, which is a very different risk profile when the output is a set point on industrial equipment rather than text on a screen.

Let's be clear about what this is. The piece comes from Technology Review's custom content arm, not its newsroom, and it was made with AVEVA. That does not make it wrong. It does mean the success stories deserve outside confirmation before anyone leans on them.

The most useful part is how Garg describes moving from recommendation to automation. AVEVA has piloted systems that adjust equipment set points on their own instead of only suggesting them, with guardrails like keeping the automation inside a fixed operating band or limiting it to certain areas. She is candid that the industry is still figuring out the human supervisor role, and that guardrails for AI will not look like the ones we give people. We think that honesty is the right posture for anything touching critical infrastructure.

The headline results need more scrutiny. Garg cites SCG Chemicals, a Thai petrochemical company, targeting around 99% plant reliability and an early-pilot return of almost nine times on its AVEVA platform. She also mentions a study suggesting industrial AI adoption rose nearly 78% over two years, without naming it. These are claims from a vendor about its own tools, and we would want to see independent data.

The workforce angle is where this leads if it scales. Garg says close to half of the industrial workforce is set to retire within five years, and frames AI as a way to capture that expertise. That is a real opportunity. It also risks turning decades of judgment into a model that newer workers trust without the experience to question it, which is exactly the failure she warns about with her own toy robot giving bad troubleshooting advice.

Our read: the governance philosophy is sound and worth copying. The proof points are marketing until someone else checks them.

Garg says she chairs the IEEE P7100 working group developing a standard method to measure AI's environmental impact across energy, resources, water and carbon. She also describes AVEVA work with Idaho National Laboratory on AI for grid resilience as renewable generation grows.