A few years ago, I sat in a meeting where a team proudly demoed a generative AI tool that could summarize contracts in seconds. Everyone was impressed. Six months later, almost no one was using it. The model hadn't gotten worse. Nothing about the technology had changed. What had happened was more ordinary and more familiar: the tool never got embedded into how people actually worked, nobody had resolved who was accountable when it made a mistake, and the people expected to use it had never been brought into designing how it would fit into their day.
I've now seen this pattern enough times, across enough organizations, to be confident it isn't a one-off. It's the default outcome when AI is treated as a technology rollout instead of what it actually is: a change management problem wearing a technology costume.
The part everyone gets right
Picking a capable model is the easy part now. Between the major frontier labs and the tooling built on top of them, most organizations have more raw AI capability available to them than they know what to do with. That was not true even three years ago, and it's easy to still be operating as if it were — spending months evaluating models when the model was never going to be the constraint.
The part almost everyone gets wrong
The constraint is almost always somewhere else: a business problem that was never clearly framed, data that isn't ready for the use case that got picked, a workflow that was never redesigned to actually incorporate the AI's output, or a group of users who were told to adopt a tool rather than invited to help shape it.
In my own work, the AI initiatives that actually stuck shared a few things in common. We picked use cases that were high-volume and well-defined, where a mistake was cheap to catch — not the most impressive demo, but the one most likely to survive contact with real, messy data. We built human review into the workflow at the exact points where errors would matter, instead of treating oversight as a compliance checkbox added at the end. And we spent real time with the people who'd actually be using the tool — in one case, coaching tax professionals on how to work with AI-assisted output — because their trust in the tool mattered as much as its accuracy.
What "operationalizing AI" actually means
I don't position myself as someone who builds AI models, and I think that distinction matters more than it gets credit for. My role in this work has been business problem framing, use-case selection, workflow redesign, governance, and adoption — the unglamorous work that determines whether a promising pilot becomes a permanent part of how a team operates, or a slide in a deck nobody opens again.
That's a different skill set than model development, and honestly, most organizations need far more of it than they currently have. There's no shortage of AI capability sitting on the shelf right now. There's a serious shortage of people who know how to turn that capability into something a business actually trusts enough to run on.
Before greenlighting your next AI initiative, ask three questions first: what specific decision or task does this change, who is accountable when it's wrong, and have the people expected to use it helped design how it fits into their workflow? If you can't answer all three, you're not ready to scale it — no matter how good the demo looked.
None of this makes AI less exciting to me — if anything, it's made me more convinced the value is real. But the value shows up on the operational side of the work, not the model side, and that's exactly the layer I've spent my career building.