The org chart is a cached process
Most AI rollouts fail for the same reason most software rollouts in the 90s failed. Nobody redesigns the process. They take the process as it already exists, with all its handoffs, approvals, and waiting, and they point a model at one step of it. The step gets faster. The process does not.
An org chart is not neutral. It is a process, frozen at the point someone last decided who does what. Every box and every line between boxes encodes an assumption about how work has to move. Add AI without touching that structure and you have not automated the work, you have automated the org chart, waiting periods and all.
I think this is the real reason "apply AI to the business" so often produces nothing. The task got faster. The queue it sits in did not move.
What actually has to change
I ran into this directly at Sourceful, on a five-person team building a product across firmware, backend, mobile, and design. If we had kept our roles as they were on paper, mine was product and design, one teammate was iOS, another was backend, another was firmware, and our CTO ran the platform, we would have spent half our time in handoffs a team that size cannot afford. So the roles blurred. Everyone became full stack in practice. I wrote code. Engineers weighed in on UX. Nobody waited for a handoff that used to exist because a job title said it had to.
What did not blur was judgement. Each of us stayed the person who decided what "correct" looked like in our domain. When we needed an eval for a piece of AI-generated firmware logic, our firmware engineer wrote it, because he was the one who actually knew what a bad answer looked like. When a generated interface needed a bar to clear, that was mine to set. The expert did not disappear when the task got shared. The expert became the person who defined the standard everyone else, including the agents, had to meet.
The role stopped being a description of which tasks you did. It became a description of which judgement calls you owned.
That distinction is the entire redesign, in one sentence. Task ownership can spread across a team, across agents, across whoever is available. Judgement ownership cannot. Somebody still has to be accountable for knowing what good looks like, and that person's job gets more important as the volume of task-doing around them gets automated, not less.
Why the same move gets harder, not easier, at scale
Five people can renegotiate who does what over a Slack thread and a bit of friction. A five-thousand-person company cannot. The instinct to keep roles exactly as they are is stronger the bigger the org gets, because more people's jobs, titles, and reporting lines are built on top of the current shape. Nobody in accounts payable is empowered to tell finance leadership their fourteen-step process should be five. So it stays fourteen steps, and whatever AI gets bolted on just makes those fourteen steps faster.
The scale problem is not technical. It is that redesigning roles alongside a process requires someone with the authority to say a role should change, and in most companies that authority sits several layers above the person who actually understands the process well enough to redesign it. The people who can see the fourteen unnecessary steps are rarely the people who can remove them.
This is why the boring-sounding part of AI adoption, deciding who owns which judgement calls once the routine work is gone, matters more than the model you pick or the framework you build on. Get that part wrong and you end up with the enterprise version of what almost happened to our five-person team: everyone doing a bit of everything, nobody clearly accountable for whether any of it is actually good.
What this looks like done properly
Done properly, the shape is consistent whether it is five people or five thousand:
- Decide what the routine work is, and let it move to whoever, or whatever, can do it fastest. Task assignment stops being a status symbol.
- Keep domain experts as the owners of judgement, not the owners of tasks. Their job becomes defining what correct looks like, and building the evals or review gates that hold agents to that bar.
- Push the authority to redesign down to where the knowledge is. The person who understands why a process has fourteen steps is usually the person who can tell you which nine are load-bearing.
- Expect the org chart to move. If roles have not changed shape in six months, the process underneath them probably has not either, and the AI on top of it is just making the old shape faster.
None of this is really about AI. It is about whether a company is willing to change how work is organised, not just what tool does it. AI is the reason this conversation is happening again, but the underlying problem, mistaking the org chart for the process, has been sitting there the whole time.
Redesign the process. Then build the agent that runs it. But redesign the roles too, or the agent just inherits the same handoffs the humans were stuck in.