Building a Safer Path to Autonomous Industrial AI
As AI takes on more autonomous roles in industrial settings, organizations need new approaches to data, governance, and human oversight, says Arti Garg, chief technologist at AVEVA.
By MIT Technology Review Insights | October 8, 2026
Industrial AI is entering a new phase. Rather than simply assisting human operators with recommendations and analytics, AI systems are increasingly being trusted to act—adjusting processes, optimizing energy use, and making real-time decisions across complex industrial environments.
That shift toward autonomy promises significant gains in efficiency, safety, and sustainability. But it also raises urgent questions about how organizations manage data, govern decision-making, and keep humans meaningfully in the loop.
In a conversation with MIT Technology Review Insights, Arti Garg, chief technologist at AVEVA, outlined what it will take to build a safer path to autonomous industrial AI—and why the foundations laid today will determine whether that future is resilient or risky.
From Assistance to Autonomy
For years, industrial AI has mostly played a supporting role: surfacing anomalies, forecasting maintenance needs, and helping engineers interpret vast streams of sensor data. In 2026, that relationship is changing. AI agents embedded in industrial control and operations systems are increasingly authorized to take action on their own, within defined boundaries.
"We're moving from AI that tells you what to do to AI that does it," Garg says. "That's a fundamentally different risk profile."
In high-stakes environments such as refineries, power grids, and manufacturing lines, an errant decision can cascade quickly. The margin for error is narrow, and the consequences—safety incidents, environmental damage, costly downtime—are severe. Autonomy, therefore, cannot be treated as a simple upgrade. It requires a deliberate rethinking of how AI systems are designed, deployed, and monitored.
Data Is the First Line of Defense
Much of the conversation around industrial AI focuses on models. Garg argues that the real foundation is data.
"If your data is fragmented, stale, or inconsistent, no amount of model sophistication will save you," she says. "Autonomy amplifies whatever quality you already have—good or bad."
That means industrial organizations need unified data infrastructures that span operational technology (OT) and information technology (IT) environments. Real-time sensor readings, maintenance logs, engineering diagrams, and process histories must be integrated into a coherent, trustworthy foundation that AI systems can rely on.
Context matters just as much as cleanliness. An AI model operating a production line needs to understand not only what the data says but also what it means—the physical constraints, safety limits, and historical precedents that define acceptable behavior. Without that context, even well-trained models can make decisions that are technically valid but operationally dangerous.
Governance Must Scale With Autonomy
As AI takes on more responsibility, governance frameworks need to evolve at the same pace. Garg points to several principles that should guide that evolution:
- Defined boundaries. Autonomous systems should operate within clearly specified limits, with escalation paths when those limits are approached.
- Traceability. Every AI-driven decision should be logged, explainable, and auditable—especially in regulated industries.
- Continuous evaluation. Models should be monitored in production, not just validated before deployment, since industrial conditions shift over time.
- Accountability. Clear ownership must exist for every autonomous system, even when the decision-making is distributed.
"Governance isn't a brake on autonomy," Garg says. "It's what makes autonomy sustainable. Without it, you get fragility disguised as efficiency."
Keeping Humans in the Loop—Meaningfully
One of the most misunderstood aspects of autonomous industrial AI is the role of human oversight. Too often, "human in the loop" becomes a checkbox—a nominal approval step that offers little real protection.
Garg argues for a more substantive approach. Human operators should remain engaged with the systems they oversee, with the training, tools, and authority to intervene when needed. That means designing interfaces that surface the right information at the right time, and building organizational cultures where raising concerns is expected rather than discouraged.
"The goal isn't to remove humans from the process," she says. "It's to place them where their judgment matters most."
A Roadmap, Not a Race
The path to autonomous industrial AI is not a sprint. Organizations that treat it as a race to deploy the most advanced models risk building systems that are powerful but brittle. Those that invest in data foundations, governance, and human-centered design will be better positioned to scale autonomy safely.
"The question isn't whether industrial AI will become more autonomous," Garg says. "It's whether we'll build the guardrails before we need them—or after."
For an industry defined by physical risk and long asset lifespans, the answer will shape not only competitiveness but also safety, trust, and the long-term viability of AI in the world's most critical infrastructure.
This article was sponsored by AVEVA and produced by MIT Technology Review Insights.