What would an AI-native organization look like if we founded it today, with one hundred people, capable AI systems and no inherited departments? Not an existing business with a chatbot added to every role. An organization designed from the outset for a world in which access to cognitive capability is becoming less constrained.
This is a thought experiment, not a forecast. I am not suggesting that legal obligations, professional expertise or human relationships disappear. I am asking something more uncomfortable: which parts of our organizations exist because they create value, and which survive because we have forgotten the constraints that originally made them sensible?
A week to understand last week’s work
Consider a ritual I recognize from professional services. Employees have to record their working hours precisely. Their project supervisor may spend part of the weekend approving those entries. A disciplinary supervisor then has until Tuesday to approve them again. Controllers assemble the numbers, and by Thursday managers can finally read a report on productivity and project margin.
This is not a claim that every firm uses precisely those deadlines. It is a recognizable pattern in which each step has a plausible explanation. Timesheets support billing and cost allocation. Project leaders need to understand delivery. Line managers have people responsibilities. Finance needs reliable information. Yet the whole sequence can turn last week’s work into this week’s administrative procession, only to arrive at information that is already ageing.
Now imagine designing the system from the outcome backward. The desired outcome is trustworthy, timely understanding of project economics and delivery risk—not the successful completion of four approvals. Reliable source data could be reconciled as events occur; anomalies could be highlighted immediately; the appropriate human could investigate the exceptions. Controls and audit trails would remain, but the routine sequence would no longer determine when the business is allowed to learn.
This distinction reveals why optimizing a workflow is not necessarily redesigning it. An AI assistant might make timesheet approval faster. Another might produce Thursday’s report in seconds. Neither necessarily challenges the premise that multiple people must routinely bless the numbers before anyone may use them. We have accelerated an inherited critical path instead of asking whether the critical path is still valid.
We have accelerated an inherited critical path instead of asking whether the critical path is still valid.
Approval for the sake of approval
Some organizational steps are more stubborn than their technology suggests. Approval is a good example because it can satisfy a human need for importance. Being the person who authorizes a client invoice or signs off a routine decision can feel like evidence of status. The approver becomes the gate through which others must pass. That feeling does not vanish when software is introduced.
The point is not that approvals are bad. Financial controls, segregation of duties, security review and professional accountability often require genuine independent judgment. The question is whether an approval exists to mitigate a named risk or to maintain a place in the hierarchy. An automated three-stage approval remains a three-stage approval. If the underlying purpose is ceremonial, faster routing simply makes the ceremony more efficient.
A better redesign begins with four questions. What decision is being made? What risk requires protection? What evidence would establish that the risk is controlled? And which human must remain accountable if the decision proves wrong? Sometimes the correct answer still includes two independent approvals. Sometimes it includes an automated check and an exception owner. What matters is that authority follows purpose, not nostalgia.
An AI-native organization starts with outcomes
For our imaginary one-hundred-person firm, I would not begin by assigning people to departments. I would begin with the outcomes customers are trying to achieve, the obligations the business must honor and the distinctive human contributions customers are prepared to value. This does not mean an organization without standards. An enterprise cannot lead effectively if its capabilities change chaotically every week, and it cannot sustainably accommodate every client’s wishes without a repeatable delivery model.
The difference is that standards can govern outcomes and assurance without fixing every intermediate output. Customers may need the same confidence in quality, safety and value while benefiting from quite different paths to the result. One engagement might need rapid experimentation before any large business case. Another might require formal evidence and controlled authorization before a small trial. The route becomes adaptable; the standard of accountability remains.
Three layers of an outcome-first organization
- Intent and outcomes: define customer value, success criteria and accountable human judgment.
- Dynamic capabilities: combine repeatable expertise and AI around flexible paths to the outcome.
- Knowledge, execution and learning: test assumptions continuously and change the route as evidence develops.
These are proposed design principles for a thought experiment, not empirical proof of one superior organizational structure.
That distinction between output and outcome is foundational. A slide deck, steering report, campaign plan or requirements document may be useful. None is valuable simply because it exists. Its value depends on whether it changes understanding, judgment or the result. The established OSC discussion of output versus outcome applies directly: when a proxy becomes the job, an organization can produce more while achieving less.
If we kept the functions, what would their purpose become?
Marketing becomes continuous market understanding
Marketing offers the most accessible example. Much of its familiar cadence reflects the time once required to gather evidence, interpret market signals, coordinate creative work, obtain approval and wait for a sufficient volume of results. There are good reasons for planned campaigns, but information maturity need not always arrive according to a calendar determined by human production effort.
If analysis and experimentation become less expensive, Marketing could increasingly function as a continuous market-understanding capability. It would observe changing needs, test ideas, connect qualitative customer conversations with quantitative evidence and support quicker commercial decisions. That does not mean more automated promotional noise. It could mean fewer irrelevant outputs because the organization learns what matters before multiplying them.
People would still decide the brand’s promise, the relationships worth building and the trade-offs that no optimization algorithm should make alone. Greater analytical frequency makes human judgment more important, not dispensable.
Human Resources becomes Human Potential
The HR example is more provocative—and, to me, more strategically interesting. When routine documentation, administration and coordination can increasingly be supported by intelligent systems, what becomes the most valuable purpose of a people function? I would explore the possibility of Human Potential as an organizational capability.
The Human Premium is not a sentimental label for anything AI cannot yet accomplish. It concerns the value of trusted relationships, empathy, contextual judgment, responsibility, finesse and the ability to discuss another person’s actual pain points. A Human Potential function would make it its business to foster these qualities, alongside learning, health, resilience and conditions for meaningful performance.
That would move some concerns often relegated to employee benefits into the heart of operating-model design. What creates avoidable cognitive burden? When does coordination consume the attention required for substantive judgment? How do we develop deeper expertise without exhausting the people who hold it? How might an organization use regained capacity to take mental health seriously, rather than treating every minute saved as an invitation to increase pressure?
There is no guarantee that AI adoption improves well-being. Without governance, an organization may use the same tools to intensify surveillance or expectations. Human Potential therefore cannot be a branding exercise; it would need authority over the choices that shape the experience of work.
The other functions are also open to reinterpretation
Product could evolve toward continuous value discovery, testing the customer’s real problem before locking an expensive delivery plan. Finance could emphasize dynamic capital allocation, shifting investment as validated evidence changes while retaining fiduciary controls. Operations could become outcome orchestration, coordinating changing capabilities without imposing every inherited handoff. IT could operate cognitive infrastructure, making identity, knowledge access, reliability, interoperability, evaluation and security dependable enough for the enterprise to work differently.
Leadership, meanwhile, would become less about supervising established sequences and more about designing the architecture of judgment: defining intent, decision rights, constraints, escalation and learning. Renaming departments, however, changes nothing unless incentives, authority and measures change with them. An old department with a new title is still an old department.
What research supports—and what remains a hypothesis
Evidence boundary: faster tasks are not a redesigned enterprise
We have reason to take expanded cognitive capability seriously without claiming that the organizational thought experiment has been proven. In a field study of customer-service work, Brynjolfsson, Li and Raymond’s research on generative AI at work reported an average productivity improvement of roughly fifteen percent with AI assistance, alongside marked differences across worker groups. That result concerns a particular setting, measured task and deployment. It does not demonstrate that eliminating departmental boundaries necessarily produces better firms.
Research into workplace coordination likewise highlights the cost of interruptions and fragmented attention. Microsoft’s 2025 analysis of the ‘infinite workday’ describes a pattern of disruptive messaging and meeting activity in the user populations examined. Its telemetry does not represent every global worker and excludes EU tenants; it should not be presented as a universal measurement of office life.
The important extrapolation is ours to test. AI may help individuals complete tasks more quickly, yet the full organization may still wait on the same committees, decision rights and information bottlenecks. Faster local work can even crowd the next queue. The proper experiment therefore measures the whole route from customer intent to validated result, including quality, cost, trust and human burden.
The Capacity Dividend is an investment choice
The Capacity Dividend has always carried a consequential question: what happens to the capacity that becomes available? The answer is not a fairy tale in which every efficiency improvement automatically produces more revenue and happier employees. The dividend is an opportunity whose value depends on decisions.
A leader might reinvest it in a broader portfolio of initiatives, more rigorous discovery, better customer outcomes, higher-quality assurance, renewed capabilities or sustained growth. In some circumstances it might improve margins or reduce cost. The distribution is a human choice, subject to commercial and ethical trade-offs. The important possibility is that previously unfostered potential can finally receive attention: experiments too expensive to try, innovations postponed by reporting, relationships neglected because routine coordination consumed the week.
A redesigned critical path could also lower pressure on individuals. When cognitive effort is no longer spent on unnecessary waiting, repetitive preparation and defending status, some of it can be redirected toward understanding, creativity and thoughtful work. Better outcomes and greater useful output need not require a proportionate increase in mental load. Yet leaders can erase that advantage immediately by increasing demands whenever tools improve.
This is why the investment policy matters more than a productivity headline. The Human Premium Enterprise does not equate abundant cognition with limitless human endurance. It asks how organizations turn cognitive leverage into trusted, accountable capability while protecting the conditions under which human judgment remains valuable.
What we must not design away
A serious AI-native organization must preserve the relationships between people. Context can be gathered more completely and a response can adapt to volatile conditions faster, but empathy, finesse, trust and painful conversations still require someone to take responsibility for another person’s experience. An AI-supported diagnosis of a client’s problem is not the same thing as a trusted conversation about what that problem means.
Nor can an organization abandon legal duties, security constraints, financial stewardship, customer confidentiality, professional expertise or the need for escalation. Some decisions deserve slower independent consideration. Others require continuity that dynamic teams alone cannot provide. The ambition is not to remove all friction; it is to make the remaining friction intelligible and justified.
The NIST AI Risk Management Framework provides an authoritative framework for thinking about AI risk. The system is socio-technical, and its safeguards must be too. Permissions and monitoring matter, but so do incentives, human behaviour, ownership, professional accountability and the environment in which decisions are made. We should be able to explain both why a safeguard exists and how it contributes to a better outcome.
Protect the human premium in the experiment
A test of a different sequence needs more than a time-saving target. Establish the current baseline before introducing AI: how long a customer waits, how many times information is re-entered, which exceptions reach the right person, how often decisions need correction and how people experience the work. Then change the route for a bounded class of cases, with a named human owner and a way to restore the existing control if the new arrangement fails. The question is not whether an AI-assisted team looks impressive in a demonstration. It is whether the enterprise becomes more reliable under ordinary conditions.
Equally important, examine who benefits. If the firm produces a proposal in hours rather than days, has the client received a better answer or merely a faster document? If an employee spends less time assembling reports, have they gained space for difficult conversations and professional development, or have their targets risen immediately? If approval becomes exception-based, are managers exercising more meaningful judgment or simply receiving fewer signals about emerging trouble? A credible experiment should welcome answers that complicate the business case. Otherwise the organization has merely replaced its old approval theatre with new AI theatre.
The most uncomfortable redesign concerns executive leadership
The final question of the thought experiment may unsettle its sponsors. Executives readily ask employees to become more productive, learn new skills and adapt to the Cognitive Revolution. But a transformation that reduces information scarcity and coordination costs also challenges the economic rationale for management layers built around controlling access to information and approval.
What happens when an executive’s highest paycheck becomes harder to justify than the salaries of the people delivering value? That is not a prediction that leadership will disappear or a claim that employees will automatically earn more. It is a challenge to the explanation for managerial compensation. Being the person everyone must ask is not the same as being the person whose judgment makes the enterprise more valuable.
The Human Premium of leadership would need to show itself in selecting the right problems, allocating capacity responsibly, accepting accountability, protecting trust, enabling other people and making difficult choices under uncertainty. These contributions can be extraordinarily valuable. They must not be confused with the number of approvals a title commands.
There is a symmetrical implication for employee compensation. If people, assisted by AI, can make more valuable contributions, it becomes harder to defend pressure on their pay solely by arguing that the underlying tasks have become cheaper. Who receives the gains will depend on bargaining power, markets, institutions and choices—not on a technological law. Executives must be prepared to have that conversation with the same rigor they demand from everyone else.
The inherited-sequence decision test
- Outcome: What customer result is the process meant to create or protect?
- Necessity: Which step actually controls risk or contributes independent evidence?
- Alternative: What changes if continuous information and exception handling replace routine waiting?
- Validation: Are customer value, decision quality, trust and employee burden better?
Test one inherited sequence before redrawing the chart
An enterprise does not need to abolish departments to learn from this thought experiment. Start with one customer outcome. Trace the path from an expressed need to a verified result. Identify each approval, report, handoff and waiting period. Ask what risk each step controls, what information it provides and whether its purpose can be achieved differently. Preserve essential human decisions, then test one alternative route with measurable safeguards.
Measure elapsed time, certainly—but also customer value, decision quality, financial exposure, learning speed and cognitive burden. Compare what changed, not merely how many actions the AI completed. This is consistent with OSC’s work on organizational judgment and objective architecture: making a process faster is not success if the process is pointed at the wrong result.
The organization we would never have designed before AI may initially look quite familiar. Its difference may not be the chart on the wall. It may be the willingness to stop mistaking inherited sequences for necessities—and to apply the freed capacity to better outcomes, healthier work and a more demanding standard of human leadership.
