The most interesting AI work I have seen in the last two years was not a chatbot or an image model. It was getting a useful answer out of a utility billing system that was designed before I was born. That is the unglamorous frontier, and it is where I think the real money and the real difficulty in enterprise AI actually live.
I came up through enterprise ERP, so I have a specific bias here. When most people picture AI changing an industry, they picture greenfield: a clean new product, a clean new model, no legacy. But the industries that run the physical world, energy, water, the grid, do not get to start clean. They run on systems that have been accreting since the 1970s, and those systems are not going anywhere, because they work and the cost of being wrong is measured in people without power.
The rip-and-replace fantasy
The pitch a utility usually hears is that AI requires them to replace their core systems. This is mostly a fantasy, and an expensive one. You do not rip out the system of record for a million customers because someone demoed a clever assistant. The institutional knowledge baked into those platforms, the decades of edge cases and regulatory rules, is the actual asset. Throwing it away to chase a cleaner architecture is how transformation projects become cautionary tales.
So the real work is the opposite of replacement. It is layering. You leave the system of record exactly where it is, and you put AI on top of it: assistants, analytics, and APIs that read from the old world and present a humane interface to the new one. Less heroic, far more useful.
What that looks like in practice
Take a utility call center. A customer calls because their bill tripled. Behind that simple question sits a tangle of meter reads, tariff rules, weather, and a fifty-year-old data model that was never designed to explain itself. The old way, an agent spends four minutes navigating screens. The useful AI way, a layer on top reads all of that and hands the agent, or the customer directly, a plain answer: here is why, here is what changed, here is what you can do.
None of that requires replacing the billing engine. It requires understanding it well enough to translate it. Billing transparency, lower call volume, self-service that actually serves. The wins are concrete and boring, which is exactly why they are real.
The hardest part of enterprise AI is not the model. It is having the patience to understand a system everyone else wants to throw away.
Why this changed how I think about AI everywhere
Working on this rewired my sense of which AI actually matters. The flashy demos assume a blank slate. The valuable work assumes the opposite: that the world is full of systems that are load-bearing, half-understood, and impossible to switch off, and that the job is to make them legible rather than to replace them.
That is true far beyond utilities. Most of the economy runs on software somebody is afraid to touch. The teams that win with AI will not be the ones with the best models. They will be the ones willing to do the patient, unfashionable work of understanding the old system before they wrap it. We run our own stack for exactly this reason: it is the only way to learn that the boring part is the whole job.
I find this genuinely exciting, which I realize is a strange thing to say about utility billing. But there is something honest about an AI problem where you cannot hide behind a nice demo, because a real person is waiting to find out why their bill went up. That constraint makes the work better. It usually does.