Woodside Energy, the Perth-based oil and gas producer, now runs roughly 50 AI agents in live production across exploration, maintenance, and plant operations, according to vice president for digital Andrew Melouney, speaking on MIT Technology Review’s Business Lab podcast. The company’s flagship deployment, an agentic copilot called Startup Advisor, helps operators sequence the startup of liquefied natural gas plants, a process Melouney calls technically demanding and highly specialized. The tool does not run the plant. It sits beside the operator, replaying prior startups and surfacing recommendations while a human keeps the controls.
That distinction matters more than the framing usually applied to agentic AI. Woodside’s push into autonomous systems did not begin with chatbots or copilots. The company started building predictive analytics and machine learning tools across its asset base around 2015, Melouney says, years before “agentic AI” entered the industrial vocabulary. That earlier work on data governance, covering sensor feeds, maintenance records, and time-series data lakes, is what let the company layer autonomous agents on top of existing models instead of starting from zero.
The payoff shows up first in maintenance economics. Woodside’s Maintenance Intelligence tool correlates equipment performance data with maintenance logs pulled from SAP, and Melouney says it has cut maintenance hours by as much as 15 percent over five years on one asset where the company piloted it. Startup Advisor applies the same underlying pattern to a higher-stakes workflow, giving less experienced panel operators the kind of contextual guidance a veteran colleague would otherwise supply.
The industrial AI story unfolding in energy and heavy manufacturing reads differently from the consumer AI story dominating this year’s headlines. Chatbot vendors compete on autonomy and benchmark scores. Woodside is explicit that its agents remain accountable to a human operator, a choice that looks as much like liability management as it does philosophy. Melouney describes an ambition for what he calls an autonomous enterprise with “agents with agency,” yet every concrete example he gives keeps a person making the final call. That gap between the stated ambition and the deployed reality is worth watching as more industrial operators adopt similar language.
Governance appears to be doing real work here, not just serving as messaging. Every proposed use case goes through a structured assessment against privacy and cybersecurity controls before an internal AI council of senior leaders weighs in on prioritization and risk. Melouney says the company has not yet solved lifecycle management, meaning how it tracks model drift and decides when to retune an agent, at its current scale of roughly 50 agents. He calls managing 500 or 5,000 agents an open problem rather than a solved one, which is a more candid admission than most vendors selling agentic platforms tend to offer.
Woodside’s core systems run in partnership with Infosys as managed service provider, an arrangement Melouney frames as the operational reliability that earns his digital team room to experiment elsewhere in the business. Without that stability, he suggests, the organization would not extend the trust needed to deploy agents into safety-critical workflows.
For operators in heavy industry tracking the agentic AI wave, Woodside’s approach is a useful counterweight to autonomy-first framing coming out of Silicon Valley. The sequence that produced Startup Advisor, roughly a decade of data investment before any agent touched a live workflow, suggests that lifecycle governance, not model capability, is the binding constraint on scaling agentic systems in physical, safety-critical environments. Any energy or industrial operator evaluating agentic AI vendors this year should ask how the vendor handles agent retraining and drift at scale, since that is the problem Woodside itself has not yet finished solving.
Reported by MIT Technology Review on July 2, 2026.
