Energy giant Woodside now runs around 50 AI agents in production, and the story of how it got there is one of the clearest playbooks yet for scaling AI in a heavy-industry setting. In a conversation published by MIT Tech Review, Woodside’s Andrew laid out the philosophy behind the rollout: “think big, prototype small, and scale fast.” What stands out is that this isn’t a tech company talking about chatbots. It’s an operator running liquefied natural gas plants, some of the most technically demanding facilities on the planet.
The headline shift Andrew describes matters for every company past the pilot stage. Woodside started broad, handing generative AI to employees for personal productivity and staying open about which business problems to solve. That built trust and familiarity. But it didn’t scale. So they pivoted to a tighter focus on high-value solutions, according to MIT Tech Review. This is the move most enterprises are wrestling with right now: the honeymoon phase of “let everyone play with AI” eventually hits a wall, and you have to decide where the real money is.
What agentic AI looks like on a real plant
The flagship example is Startup Advisor, an agentic AI tool that sits beside the operators who fire up LNG plants from a control panel. Starting an LNG facility is brutally complex and demands specialist skill. Startup Advisor acts like a copilot. It lets operators replay previous startups, track how the current one is progressing, and get insights on how to optimize the sequence.
The practical payoff is knowledge transfer. Andrew’s framing is sharp: give a junior panel operator “a copilot that can help them almost like an experienced panel operator sitting next to them.” That’s the quiet promise of agentic AI in industrial settings. Not replacing the expert, but cloning their judgment and putting it next to the person who needs it at 3 a.m.
The three lessons worth stealing
Andrew boiled Woodside’s learning down to three pillars, and they translate to almost any organization trying to scale AI:
- Move from isolated solutions to enterprise-wide capability. Point solutions solve small problems in one corner. Agents with agency work across whole workflows.
- Standardize the platform and reuse patterns. “We don’t want to build 50 solutions in 50 different ways,” Andrew said. Repeatable patterns let teams ship fast and safely.
- Build governance that can actually keep up. Traditional ways of governing software rollouts won’t scale to the breadth AI demands. Strong governance isn’t a brake here. Woodside treats it as what lets them go fast.
That third point is the one most companies get backwards. They see governance and speed as a tradeoff. Woodside argues they’re the same thing, because you can’t scale 50 agents into live operations without a framework people trust.
Why the Infosys piece matters
Woodside runs this on top of a managed service partnership with Infosys, which handles core operations. Andrew’s line here is worth repeating: “our license to innovate is based on our license to operate.” His team only gets to reimagine how work gets done because the underlying platforms and applications run reliably every single day. Innovation sits on a foundation of boring, consistent operations. Skip the foundation and the fancy agent layer collapses.
What to take away
For practitioners and businesses watching this, a few things are worth acting on now:
- If your AI effort is still “everyone experiment freely,” plan the pivot to a focused, high-value shortlist before budgets get questioned.
- Pick a platform and enforce repeatable patterns early. Fifty bespoke builds is a maintenance nightmare waiting to happen.
- Treat governance as an enabler you design up front, not a cleanup job for later.
- Make sure your operational base is rock solid before you layer agents on top.
Woodside isn’t a Silicon Valley lab, and that’s exactly why this case is useful. It shows agentic AI moving out of demos and into live, high-stakes environments where mistakes carry real consequences. The companies that win the next phase won’t be the ones with the flashiest models. They’ll be the ones with the discipline to scale them safely. Full details are in MIT Tech Review’s original conversation.