Abstract
The wrong way to think about AI is as a better worker that should be inserted into existing organizations. The better way is to treat AI as a force that changes how work is organized. That shift does not start with the org chart. It starts with the workflow: the sequence of steps, handoffs, decisions, exceptions, and responsibilities through which work becomes real.
Reworking Order starts from a simple operating principle: meaningful change should remain human-centred at the level of intention, authorization, and accountability. AI should expand the option space, compress execution, and reduce friction. Humans should still define the objective, frame the problem, choose the tradeoff, and own the decision to proceed. That is not a sentimental preference. It is the most coherent design principle for an environment where AI is powerful, but work still requires legitimacy, judgment, and responsibility.
This is not a call to use AI lightly. It is a call to use it intensely, but in the right order. Today, most important changes should be human-initiated and human-authorized. Over time, that balance may move along a continuum as system quality, trust, and calibration improve. But for now, the best model is not agent-first. It is human-framed, AI-expanded, and human-authorized.
The Real Design Problem
Most organizations are still asking the wrong question. They ask where agents should sit, which jobs they should replace, or how to redraw reporting lines around them. That is the old logic with a new object attached. It is useful only if the workflow itself stays intact.
The deeper change is that AI alters the task chain. Work is not a single unit. It is a sequence of steps, decisions, handoffs, and exceptions. When AI enters that sequence, the economics of each step change. Some work can be automated. Some can be augmented. Some still needs a human because the work is not mainly execution. It is judgment, legitimacy, coordination, or accountability.
That is why the org chart is not the point. The workflow is.
The research is already pointing in this direction. A formal NBER paper on AI adoption argues that organizations are composed of interacting tasks and that AI adoption often requires system-wide change, not just task-level substitution. Another paper on the productivity J-curve shows why general-purpose technologies like AI need complementary investment in processes, human capital, and new business practices before the gains show up. The point is not that AI has no value at the task level. The point is that the value only compounds when the surrounding system is redesigned too.
That is the first lesson for Reworking Order: the unit of redesign is not the agent. It is the work system.
Human-Framed Does Not Mean AI-Light
The phrase human-centred can sound cautious if it is not defined properly. In this practice, it does not mean slow, manual, or anti-automation. It means that the locus of initiation and authorization remains human while AI expands what can be explored, tested, drafted, and executed.
That distinction matters.
In software, for example, AI can propose a feature, draft the spec, write the code, generate tests, and surface edge cases. But a human should still define the feature, explain why it matters, choose among the tradeoffs, authorize release, and remain responsible for what follows. The same logic applies beyond software. AI can draft a policy, outline a workflow, or simulate an implementation path. Human judgment still has to decide whether the change is worth making and what risks are acceptable.
This is not only a moral preference. It is also a practical one. Research on decision authority shows that when humans must still learn the payoffs of choices, principals may rationally keep authority with the human agent rather than simply hand it to AI. Research on human-AI collaboration also shows that combinations of humans and AI are not automatically better than either alone. In a large meta-analysis, human-AI systems were often worse than the best single option, especially in decision tasks, even though they did better in content creation. That means the default assumption should not be that more AI involvement always improves outcomes. The design question is how to preserve human judgment where it matters and let AI help where it is genuinely additive.
That is why the center of the practice is not "AI everywhere." It is "AI in service of human-directed change."
A Continuum Of Authorization
The strongest version of this thesis is not that humans must always be the only source of ideas. That would be too rigid and not supported by the evidence. The better position is that human initiation and human authorization should remain the norm for consequential change, but that the system can move along a continuum over time. That continuum is an operating inference from the research, not a settled law of nature.
At the start, most important changes should be:
- human-initiated
- human-framed
- human-authorized
- AI-assisted in exploration and execution
That is the safest and most coherent model for a mature organization or a serious product team. But as systems improve, teams can move along a gradient:
- from human-only ideation to AI-assisted ideation
- from human-approved prototypes to AI-prepared prototypes
- from human-defined choices to AI-suggested options within a human-defined policy boundary
- from manual review to selective human oversight
This is not a binary. It is a continuum of confidence, trust, and delegation.
That continuum is important because it keeps the thesis practical. It avoids two bad extremes. The first is agent-worship, where the tool is treated as the origin of strategy. The second is human-only romanticism, where AI is used so timidly that it never changes the economics of work. The right position is in between. Human direction stays central, while machine capability is progressively embedded into the chain of work.
There is also a psychological reason to preserve this order. Recent evidence suggests that passive reliance on AI can reduce self-efficacy, ownership, and meaning, while active collaboration preserves those qualities much better. In other words, how AI is introduced into the workflow changes how people relate to their work. A practice that wants durable change should not just optimize throughput. It should preserve human agency while improving execution.
What This Means In Practice
If this thesis is right, then the practical test for any new change is simple: who framed it, who authorized it, and who is accountable for it?
That gives the practice a clear rule of thumb.
If the change is strategic, consequential, or irreversible, then the human should initiate it or at minimum ratify it after AI-assisted exploration. AI should help widen the search and compress the work, but the change should enter the world through human judgment. If the change is routine, reversible, or bounded by a clear policy, then the delegation can move further along the continuum. The system can be more autonomous there, because the cost of error is lower and the value of speed is higher.
That is how the continuum should be read. Not as a claim that AI should never initiate anything, but as a design discipline for deciding when initiation can safely shift. For now, most important work still belongs in the human-initiated, human-authorized zone. Over time, some categories may move outward as calibration improves, but the burden of proof should stay high.
The same logic applies to the Build vs Deliver Gradient. Early in a project, people need to decide how much structure is worth creating before the deliverable is finished. Too little tooling creates permanent inefficiency. Too much tooling creates drag. The right answer depends on repeatability, payoff horizon, and the amount of future work the structure will save. AI changes that calculation, but does not remove it. In some cases it lowers the cost of exploration, which means teams can test more before committing. In other cases it raises the value of having a human keep the final authorizing role, because the decision now moves faster and can spread further.
This is where Reworking Order differs from generic AI optimism. The practice does not assume that the presence of AI itself creates better organization. It assumes the opposite: the organization has to be redesigned so the AI can actually do useful work without eroding judgment, ownership, or coordination.
That is why the language matters. The right framing is not "agents are the center." It is "humans remain the source of intention, while AI becomes a powerful amplifier of option space and execution." That is both more defensible and more useful to leaders who are trying to make real decisions now, not ten years from now.
Diagram
Why This Is A Good Business Thesis
The value of this thesis is that it gives the practice a stable point of view without making it inflexible.
It is stable because it says something clear:
- AI is powerful, but work is still a social and organizational problem
- workflow is the real unit of redesign
- human initiation and authorization matter
- complementarity has to be designed, not assumed
It is flexible because it allows for change over time:
- some work can become more autonomous
- some decisions can become more AI-assisted
- confidence can rise as systems prove themselves
- the boundary between human and AI participation can move
That is a serious operating model, not a slogan.
Sources
AI adoption is shaped by interacting tasks, which means organizational redesign matters.
General-purpose technologies require complementary investments in processes and human capital before the gains show up.
Decision authority is a central design variable, not an afterthought.
Human-AI combinations are not automatically superior; task type and design matter.
How AI is integrated affects agency, ownership, and meaning.
AI changes the bundling of steps, tasks, and jobs under new coordination constraints.
Early gains tend to appear first in individually adjustable work, not coordination-heavy work.
AI adoption is associated with workforce reorganization and flatter hierarchies.