The human reorg
You don’t need another model, you need new roles
Most enterprises still treat AI like a software rollout: pick a vendor, run a pilot, publish a success story, move on. But systems that learn, adapt, and collaborate don’t fit neatly into that pattern. They don’t stabilize after “go-live.” They keep changing, and so do the ways people use (or avoid) them. That reality forces a quieter but more radical shift: a human reorg.
Traditional IT, data, and operations roles were built for static systems. You defined requirements, shipped a release, and then measured uptime and tickets. With live and learning systems, the job looks different. Someone has to decide which behaviors are acceptable, which failures are tolerable, and how models should adapt to new data or new regulations. That’s not “maintenance”; that’s continuous product management.
A new class of hybrid operators is starting to emerge: AI systems architects who understand both infrastructure and workflows; feedback engineers who design how signals from users, logs, and outcomes flow back into training; human-in-the-loop trainers who handle edge cases and escalate genuinely ambiguous decisions. These are not pure ML roles. They sit at the intersection of domain expertise, UX, and risk.
The real bottleneck, though, isn’t technical—it’s human.
“People aren't going to go to another portal and create another login to use AI in their workflow. They want something that feels seamless and feels like it's not there. Jan from accounting isn't going to read the academic leaderboards... she wants a report card, she wants a nutrition label, in language that she can understand.” – Lydia Andresen
Most organizations still roll out AI the way they roll out dev tools: they train the enthusiasts and hope everyone else catches up. Engineering-style prompt training—syntax tips, clever hacks, token talk—dazzles tech teams and terrifies typical users, who just want to know: Does this help me do my job, or is it here to replace me? If the answer isn’t unambiguously about augmenting their work, adoption stalls or goes performative: people nod in workshops, then quietly revert to old spreadsheets and email chains.
In practice, success depends less on how “advanced” the model is and more on who owns it. If AI is a board mandate delegated to a technology function, it will be treated like compliance: something to be reported, not something to be used. You get dashboards, steering committees, and no real behavior change.
“I think you're going to see a jobs explosion… the enterprise sector is going to need to hire literate people in AI and data in a way and in places that they've never thought that they needed to before.” – Jordan Cealey
The organizations that actually get leverage from AI do something different: they push ownership into the business. A sales leader owns the AI copilot in their pipeline reviews. An operations leader owns the automation that reallocates work between humans and agents. Their KPIs, incentives, and headcount planning all assume that AI is part of how the function works—not a side project.
That shift forces uncomfortable questions:
There isn’t a neat org chart pattern that solves this. But there is a clear failure mode: treating AI as “something IT does” instead of as a change in how the business operates.
The next wave of value won’t come from another model upgrade. It will come from companies that are willing to reorganize people, power, and responsibility around systems that learn—rather than pretending those systems are just another tool in the stack.
