Sangeet Paul Choudary has written an interesting piece on The problem with agentic AI in 2025. The article argues that while agentic AI (systems that act autonomously rather than simply assist) holds great promise, many organisations are ill-prepared for its demands: governance, data quality, and clarity of purpose. He says they risk doing more harm than good if they treat it as just another off-the-shelf tool.
His central point is that agentic AI should be approached like a railway – something that requires governance, coordination, and planning – rather than a slightly faster canal. And while I agree with him that governance will become essential, I think we’re missing a crucial step: you can’t govern a network that doesn’t yet exist.
Every era has its great leap forward. For the Victorians, it was the railway. For us, it’s artificial intelligence. And just like those early engineers, we’re caught up in a rush of progress, laying new lines at a dizzying pace. Every week brings another tool, another automation, another gleaming piece of track stretching off into the future.
But if you look closer, you’ll notice something about those first railways. Many of them didn’t connect. They were short, ambitious bursts of steel between nearby towns, each built by its own company, each proud of its speed and ingenuity, but each ultimately isolated. You could go fast, but only for a few miles.
AI feels very much like that today. We’ve built countless short pieces of track in tools that automate one step, models that improve one process, plugins that handle one task. Each is a marvel in its own right: a writer that drafts your email, a system that analyses customer feedback, a bot that sorts your calendar. But these are small lines between small stations. They start somewhere, they end somewhere, and then the train stops.
Sangeet writes that:
Agentic AI needs the same reframing today: moving beyond faster execution of isolated tasks to building the governance frameworks that allow many agents to act together coherently across a system.
It’s a fair point and he’s right that coherence will be critical. But before we can build the frameworks, we need to build the connections. We need those short tracks to link up, to form a working system where one process hands off naturally to the next. He hints at the importance of governance but doesn’t yet describe what that might look like in practice, and perhaps that’s because we haven’t yet seen the kind of joined-up network that would need it.
I believe the real transformation comes when the tracks connect – when data moves smoothly between systems, when an insight in one tool triggers an action in another, when the network itself starts to hum. That’s when you stop having a collection of clever gadgets and start having a functioning railway. Then, and only then, does governance truly matter.
It’s easy to be dazzled by the engines, by how fast a single tool can move. But speed on its own isn’t progress. The early railways only changed the world when they became networks and when it was possible to travel, uninterrupted, from the places people were to the places they wanted to be.
That’s the stage AI has to reach. Not just faster tools, but connected ones. Not just automation, but continuity. The future isn’t a patchwork of short lines running in parallel; it’s a joined-up system where every process leads naturally into the next, carrying value all the way along.
Governance and coordination will absolutely matter but a train that can only run a mile isn’t much use, no matter how fast it goes or how well governed it is.