When Christopher Nolan approached Hans Zimmer about composing for a new film, he did not begin with spaceships, black holes, or the scale of an epic science-fiction production. He gave Zimmer a short, personal story about parenthood and asked him to respond to its emotional premise. Zimmer wrote a fragile piece connected to his own son. Only afterwards did Nolan explain that the music was for Interstellar. Zimmer had written something intimate for a film of enormous scale, but Nolan recognized that the piece had found the heart of the story before the expected conventions of the genre could shape it.
That creative decision offers a useful principle for AI operating model redesign: defining the familiar form too early biases the solution toward that form.
If an organization begins with the idea of an “AI analyst,” it has already made several design choices. It has accepted the analyst role as the unit of work. It has inherited the analyst’s inputs, outputs, handoffs, approval points, and position in the hierarchy. The organization may automate the role brilliantly and still fail to reconsider whether the surrounding value chain should exist in its current form.
The same applies to an AI recruiter, project manager, editor, or customer-service representative. The label makes the capability easier to understand, budget, govern, and explain. It also narrows the imagination. It encourages leaders to ask how an agent can perform an existing job faster rather than what outcome the organization could create if cognition no longer had to follow the same human sequence.
This is the central tension in agentic transformation. Human-shaped roles are useful scaffolding. They should not automatically become the design units of the AI-native enterprise.
My own agent system contains the same compromise
My own experiment in setting up Open Source Consulting as an AI-native initiative ended up with me describing the named AI roles I use to operate it. Steve develops the strategy. Frank researches and writes. Richard evaluates career opportunities. Richi searches for them. Don challenges the commercial angle. Guddy thinks about the website as a product. The names make the system easier for me to operate because they create memorable boundaries around context, responsibility, and output.
The system works. It has helped me build an AI-assisted professional-services model while keeping a demanding day job. That was the original constraint: I could not scale myself, hire a conventional team, or give up the work that currently pays for the experiment. AI created the possibility of building something closer to a team around the capacity I actually had.
But the system is still recognizably human-shaped.
Steve briefs. Frank writes. I review. I approve. The article is published.
The sequence is faster than a traditional editorial workflow, and the agents carry more persistent context than a generic assistant. Yet the operating logic remains familiar:
brief → produce → review → approve → publish
That does not make the experiment invalid. It makes it an intermediate stage. Familiar roles give me legibility, control, and confidence while I learn what the system can do. They are a bridge between using AI as a tool and designing an operating model that does not reproduce human constraints.
The risk begins when organizations mistake that bridge for the destination.
Human operating models were built around human constraints
For most of history, the human mind and body were both the driver and the limiting factor of nearly every process chain. Hunting, farming, manufacturing, typing, printing, analysis, management, and decision-making all evolved around the fact that people have finite attention, imperfect memory, limited working hours, and a need to coordinate with other people.
Organizations learned to make those limits manageable. They divided work into roles. They created departments to concentrate expertise. They used meetings to synchronize knowledge. They produced documents so information could travel without the author. They introduced approval stages because decision-makers could not observe every activity directly. They scheduled reviews because senior attention was scarce. They built reporting cadences because analysis could not be continuous.
These structures were rational responses to human limitations. They became so familiar that we often experience them as natural properties of work rather than design choices.
That is why the idea of a digital employee feels intuitive. It preserves a model every leader already understands. There is a box on the org chart, a list of responsibilities, a manager, a queue of tasks, and an expected output. Replace or augment the person in the box with an agent, and the transformation appears concrete.
But AI does not share the same constraints. It has different ones.
An agentic system can potentially monitor evidence continuously, preserve explicit context, compare alternatives repeatedly, run several lines of analysis in parallel, regenerate outputs when conditions change, expose contradictions across functional boundaries, and trigger work when an event occurs rather than when a meeting appears on the calendar. It can act through tools, maintain state, and iterate over its own intermediate results. Anthropic distinguishes predefined workflows from agents that dynamically direct their own processes and tool use; the architectural distinction matters because different forms of orchestration create different possibilities for how work is sequenced and controlled.
None of this means AI possesses judgment, legitimacy, purpose, or accountability. It does not. It means the operating model should reflect a different distribution of capability instead of assuming that AI is simply a faster person. This is also why workflow accountability cannot be delegated merely because execution becomes agentic.
The cognitive difficulty is similar to trying to understand time dilation. We know intellectually that two observers can experience the passage of time differently because of their relative velocity, but the idea conflicts with a lifetime of shared human experience. In the same way, it is difficult to imagine a process that does not revolve around sequential human contribution because every organization we have known was built that way.
Breaking the model requires more than adopting a new tool. It requires stepping outside a comfort zone shared by almost every human being who has ever organized work.
Start with the problem, not the AI role
During my years doing Design Thinking work with clients, the opening question was often some version of: “How can we make a better product?”
The question sounded reasonable. It was frequently wrong.
In several engagements, broader discovery showed that the end customer did not need a better product. They needed a better service, a different commercial model, clearer guidance, or a more coherent experience around the product. Starting from “better product” had already narrowed the solution space before the real problem was understood.
The Design Council’s Double Diamond remains a useful representation of this challenge. Its first movement is divergent: understand the problem rather than assume it. Then define the challenge more precisely. The second diamond broadens the possible answers before converging on what should be tested and delivered. The point is not the diagram itself. The point is the discipline of knowing when to widen the frame and when to narrow it.
AI operating-model design needs the same discipline.
Organizations often start at the wrong end of the sequence. They ask whether a use case is technically feasible before validating whether the problem is worth solving. They see a new model capability and search for a business process to attach it to. The technology gives life to ideas that may solve no meaningful problem.
A stronger sequence is:
- Does a valid problem exist?
- What outcome should improve?
- Is AI the right instrument for that outcome?
- Does the current state of the technology support the required behaviour reliably enough?
- Is the result viable in financial terms and in the human effort required to operate, govern, and trust it?
This is not anti-innovation. It is how innovation avoids becoming theatre.
The equivalent of Nolan withholding the phrase “epic science-fiction film” is refusing to begin with “AI analyst.” The organization should first understand the purpose, the decision, the evidence, and the value. Only then should it decide which combination of agents, workflows, people, controls, and interfaces can produce the outcome.
The opportunity is not only faster steps, but different relationships between steps
The easiest first improvement to understand is not the elimination of every stage. It is the ability to make previously sequential activities simultaneous.
Consider a familiar corporate sequence:
request → analysis → documentation → meeting → approval
Traditionally, each stage depends on the completion of the previous one. A person receives a request, performs the analysis, creates a document, schedules a meeting, collects feedback, revises the work, and returns for approval. The process contains days or weeks of latency because human attention, production, and senior availability are all constrained.
AI can change the relationship between those activities.
The request itself can be assisted and clarified as it is made. Analysis and documentation can happen together. During the meeting, the system can update the recommendation, test alternatives, expose missing evidence, and regenerate the artifact while the discussion is still taking place. The meeting can end not with a list of actions for the next iteration, but with an agreed and documented outcome.
For decades, it was considered bad practice to “work on the thing” during the meeting. The meeting was for discussion; the work happened before and after. That separation made sense when production required sustained individual concentration and iteration was expensive. With AI, co-development can become the meeting’s most valuable activity.
This does not necessarily remove the human decision. It can strengthen it by giving leaders the ability to see their assumptions translated, challenged, and revised in real time.
The relevant question is no longer only, “How do we shorten every step?” It is, “Which steps can now occur together, and which delays no longer create value?”
That is already a meaningful operating-model change. It also points toward a more radical one.
A two-week assessment became an hour, but the closed loop mattered more
I saw this directly in an enterprise experiment designed to generate a strategy for recurring revenue.
The inherited sequence required one person to produce at least three iterations of a draft. A group of senior leaders then had to approve the result before the work could move into ideation. The process was built to protect leadership intent and maintain quality, but it was limited by scheduling, asynchronous handoffs, and the availability of people whose attention was already scarce.
The work could take roughly two weeks.
Using AI, we brought the relevant leaders together for a concentrated session. The system could interpret input, generate options, capture changes, and iterate quickly enough for the group to work through the substance together. Instead of sending SLA-like requests from one senior person to another and waiting for the next response, we needed a few hours of shared attention.
The experiment succeeded until we left the meeting.
At that point, I still needed the leaders to provide a formatted result that I could present to the country president. We had transformed the thinking process but allowed the old document-production handoff to survive at the end.
That was the more valuable lesson.
AI enables a closed loop. The meeting should end only when the outcome, reasoning, evidence, and presentation artifact are complete enough for the next consequential decision. The participants may spend less total time than they would across multiple asynchronous iterations, but the more important improvement is that alignment, analysis, documentation, and approval can become one integrated act.
A superficial reading says AI reduced a two-week workflow to one hour. A better reading says the original workflow existed because cognition, drafting, coordination, and senior attention had to be sequenced. Once that constraint changed, the operating model should have changed with it.
Automating A → B → C → D is not the same as redesigning the value chain
Now consider a strategic investment decision.
A conventional process might look like this:
- An executive frames the question.
- An analyst gathers evidence.
- A team develops a recommendation.
- A committee reviews the proposal.
- A leader approves or rejects it.
An organization can assign an agent to each stage. One agent structures the request. Another researches the market. Another prepares the business case. Another checks risks. The process becomes dramatically faster.
It may also remain fundamentally unchanged.
The deeper questions are different:
- Why is evidence gathered only after an executive asks?
- Why are alternatives evaluated sequentially?
- Why is the recommendation a static document?
- Why is review tied to a committee meeting rather than a material change in evidence?
- Why are assumptions inspected only at defined checkpoints?
- Why does each function hand off its contribution instead of interacting with a shared, continuously updated decision model?
An AI-native decision system could monitor relevant evidence continuously, maintain competing value hypotheses, update confidence as conditions change, simulate alternatives in parallel, and escalate when ambiguity, consequence, or divergence crosses a defined threshold. Humans would not disappear. They would enter where their contribution is premium: reframing the problem, judging weak signals, resolving value conflicts, building trust, and accepting responsibility for the consequential decision.
That is not an automated analyst.
It is a redesigned decision system.
The distinction matters because research on earlier general-purpose technologies repeatedly found that technology investment and organizational innovation are complements. New technology creates more value when firms also change processes, workplace structures, skills, and business models. The lesson is not that AI will repeat the past exactly. It is that installing a powerful technology inside an unchanged organization has never been the same as capturing its full economic potential. The same logic underpins the broader OSC argument that AI is not merely a productivity layer, but a value-chain question.
Letting go of human quality as the only acceptance criterion
One of the habits that will slow this transition is the idea that AI becomes valuable only when it approaches a human response.
That criterion sounds demanding, but it can be conservative. It defines progress relative to the existing human process rather than the desired business outcome.
I learned a related lesson while taking over a backlog containing approximately 1.700 items. The backlog represented a sincere attempt not to neglect any request, idea, feature, or activity that had ever been mentioned. In theory, it preserved completeness. In practice, it prevented reliable execution.
I reduced the backlog to around 70 items. There were fewer complaints than many people expected. More importantly, the 70 items were completed. The backlog then scaled into a healthy continuous system of roughly 60 to 80 activities rather than returning to uncontrolled accumulation.
The paradigm shift was not a better method for prioritizing 1.700 items. It was accepting that most of the 1.700 should not survive.
“Done is better than perfect” can be a dangerous slogan when it excuses poor judgment. It can also be a necessary challenge to systems that confuse completeness with quality. In AI transformation, organizations may need a similar willingness to accept a strong 70% result, learn from its use, and improve the system rather than demanding human-like perfection before allowing any operating-model change.
This does not mean lowering standards for consequential decisions. It means choosing standards that relate to the outcome. An agent does not need to imitate every human behaviour if the system can produce a more useful, explainable, and governable result through a different combination of machine and human contribution.
The Human Premium Enterprise is designed around what humans make valuable
The Human Premium Enterprise is not simply an organization in which humans approve whatever AI produces. That would preserve the old hierarchy while increasing the volume arriving at the approval gate.
It is an organization that redesigns work so human attention is concentrated where it creates disproportionate value.
Three contributions matter especially.
First, judgment under ambiguity. AI can synthesize and compare at extraordinary speed, but organizations still need people who can recognize when the evidence is incomplete, when values conflict, when a plausible answer is politically or ethically untenable, and when the problem itself has been framed incorrectly.
Second, trust building. AI-native employees will remain a minority for some time. Organizations will need people who can create collective trust systems, involve experts in defining better guardrails, explain why a new process is safe enough to use, and help colleagues distinguish reasonable caution from institutional avoidance. Trust is not a communication layer applied after implementation. It is part of the operating design.
Third, imagination. AI is rapidly improving execution, synthesis, and analysis. Imagination remains a distinctly human privilege, and many organizations have neglected it because their people were consumed by coordination, reporting, formatting, and production. Capacity returned by AI should not merely be reinvested in producing more artifacts. It should create room to imagine outcomes that the current value chain was never designed to deliver.
There is also a long list of valuable work that organizations say they lack time or resources to pursue: building better relationships, involving overlooked experts, testing neglected hypotheses, improving institutional memory, and exploring opportunities that do not fit the current planning cycle. An AI-native operating model can make room for that work, but only if leaders resist filling every recovered hour with more of the same output.
Named agents are scaffolding, not the destination
This brings the argument back to Steve, Frank, Richi, Richard, and the rest of my own experiment.
Giving agents human-shaped roles has been useful because I began with a practical constraint. I wanted Open Source Consulting to become an almost self-running system, but I could not scale myself. The role metaphor helped me define responsibilities, provide context, set guardrails, and keep myself as the final validation gate. The names also made the experiment enjoyable, which matters more than enterprise design language sometimes admits. Most importantly, it allowed me to invest my Human Premium in finding things that excite me, which I believe can excite others, without questioning if executing on them would be worth the effort.
The next iteration should stop being so sequential.
I can imagine a system that operates around maturity gates rather than named-agent handoffs. Evidence could accumulate continuously. A gatekeeping mechanism could determine when enough support exists to advance an argument, regardless of which agent happens to be active. Competing theses could be tested in parallel. Weak claims could be rejected before a draft exists. My personal and professional experience could be requested only when the system reaches a point where original judgment is required.
I would become a supporting actor at defined moments of premium human contribution rather than the person pushing every artifact through a synthetic production line.
The technical configuration will be challenging. The larger challenge will be my own skepticism. At some point, I will have to make what I think of as the final payment of the Trust Tax: allowing the system to operate with enough freedom to prove that it can deliver value without reproducing my own working habits at every stage.
That is why a human-shaped role can be an effective interface for an agent without becoming the design unit of the enterprise.
The transformation cannot bankrupt the business it is trying to reinvent
The strongest objection is also the most reasonable one:
We have a business to run. I can give you one part-time employee to explore AI, but I cannot pause the organization for a blank-sheet redesign.
That constraint is real. Companies need roles, budgets, controls, accountable owners, and revenue while they transform. No serious recommendation should pretend that leaders can discard the operating model and wait for a better one to emerge.
The practical risk is that AI may be unusually unforgiving of half-built transformation. A few copilots and isolated agents can create visible activity without creating meaningful return. They can accelerate local tasks while increasing coordination costs, review traffic, and uncertainty elsewhere. The organization may become busier and faster without becoming better.
The answer is not a revolutionary cutover. It is to protect space for purposeful redesign inside the running business.
Begin with one consequential outcome. Map the current value chain honestly. Identify which steps exist because human cognition is scarce, sequential, or distributed. Test which activities can become simultaneous, continuous, or event-driven. Preserve human control where consequence and accountability require it. Measure whether the new system improves the outcome, not merely whether one task finishes faster.
Then use the capacity returned by the experiment to finance the next redesign.
This leads to an emerging question for the Human Premium Enterprise:
Where do we declare ourselves satisfied with the capacity AI has returned, and reinvest that capacity in the full reinvention of the company?
If every recovered hour is absorbed by more work inside the old model, the transformation never begins.
An executive diagnostic for AI operating model redesign
Before creating another digital role, leadership teams should ask:
- Does a valid problem exist? What evidence shows that the current outcome is insufficient?
- What outcome should improve? Define the decision, customer value, risk, revenue, learning, or service result before defining the AI form.
- Which parts of the current process exist because human cognition is scarce, slow, sequential, or distributed?
- Which activities can now happen simultaneously? Can analysis, documentation, challenge, and iteration occur in one shared loop?
- Which activities could become continuous rather than episodic? What should the system observe without waiting for a request?
- Which handoffs could disappear? Where does information move only because the people involved cannot share context directly?
- Where does human contribution remain premium? Identify the moments requiring judgment, trust, imagination, negotiation, or accountability.
- Is AI the right instrument? Do not use an agent where a simpler workflow, rule, product change, or service intervention would solve the problem better.
- Can the technology support the required behaviour reliably enough today? Separate ambition from current capability.
- Is the model viable? Include financial return, operating effort, governance burden, adoption, and the cost of maintaining trust.
- What existing meeting, report, role, or approval should change if the redesign succeeds? If nothing changes, the organization may only be automating the old sequence.
- How will returned capacity fund further reinvention? Do not allow every gain to disappear into more output.
These questions are not a universal recipe. They are a way to widen the first diamond before the organization converges too early on a familiar agent-shaped answer.
The road was not redesigned for a faster horse
The familiar “faster horse” analogy is useful even though its common attribution to Henry Ford is disputed. The important idea is that a new capability should not be constrained by the form it replaces.
Roads, vehicles, supply systems, cities, and patterns of movement were not transformed because the horse became incrementally faster. They changed because a different technological system made different arrangements possible.
AI operating model redesign requires the same conceptual break.
The equivalent of building a car rather than a faster horse is not using AI to do what humans already do, only faster. It is using AI to rethink the outcomes an organization wants to achieve without being limited by its current value chain.
That does not mean rejecting every role, workflow, or control inherited from the human enterprise. It means forcing each one to justify its continued existence against the outcome rather than preserving it because it is familiar.
Named agents may help us begin. Digital colleagues may make new capabilities understandable. Automated roles may generate early savings and build confidence. But the Human Premium Enterprise will not emerge by recreating yesterday’s org chart in software.
It will emerge when organizations start with purpose, broaden the solution space, redesign the relationship between activities, and bring human beings into the system where their judgment, trust, imagination, and accountability create the greatest value.
The biggest mistake in agentic transformation may not be asking AI to do too much.
It may be defining what AI is allowed to become by giving it yesterday’s job description.
