There is a fundamental misunderstanding in how the industry is thinking about AI and job replacement.
The assumption is that once AI is capable enough, roles will simply disappear. But that assumes jobs exist independently of the systems they operate within.
They don’t.
Modern roles are embedded in deeply interconnected IT landscapes, tied into workflows, data flows, approvals, controls and dependencies that have accumulated over decades. What a person does is only part of the role. The rest is how that role connects fragmented systems and ensures work completes end-to-end.
And those systems were never designed as a coherent whole.
This is key…the IT industry doesn’t usually replace things.
It adds to them, wholesale replacement is very rarely the default approach.
Each generation of technology is introduced alongside what already exists. Integration instead of removal. Mediation instead of resolution. Over time, the result is not a clean architecture, but a mesh of interdependent components where logic, data and behaviour are distributed across multiple systems.
In that mesh, nothing is isolated.
Remove a role, and you don’t just remove a function. You disrupt a network of dependencies that were never fully defined, but have become operationally critical. This is why replacement is so difficult. Not because organisations don’t want to do it, but because the system makes it too risky.
And even this view is incomplete.
Because not all of the enterprise exists within the system.
A significant portion of how work actually gets done lives outside it. Shadow IT, spreadsheets, email threads, ad-hoc processes, and the constant layer of human intervention that resolves what the formal system cannot. These “water butt” resolutions are not exceptions. They are structural.
So when we talk about AI replacing jobs, we are not just talking about automating tasks.
We are talking about removing nodes from a system that is both deeply entangled and only partially visible.
This is the dilemma AI agents now face.
They are being introduced into environments built through accumulation, not replacement. The instinct of the industry will be to do what it has always done. Add AI on top. Let agents navigate the complexity.
And for a time, that will work.
But it doesn’t remove the underlying structure. It extends it.
Agents inherit the fragmentation and the gaps. As their scope increases, so does the pressure on the system.
At that point, the industry faces a choice it has historically avoided.
Replace the underlying systems, or constrain the AI.
And if replacement is too disruptive, the system will revert. Agents will be limited. Human oversight will return. More layers will be added.
Not because AI failed.
But because the system could not absorb what AI requires.
You don’t replace a job.
You replace the system the job exists in.
And in most organisations, that system is far more entangled and incomplete than anyone is willing to admit.