From Scientific Management to Human–AI Work Systems

The current AI transformation debate can look unprecedented: general-purpose models, copilots, agents and machine reasoning are spreading through knowledge work at extraordinary speed. But the organizational problem underneath them has a much longer history. For more than a century, managers and researchers have been asking how work should be decomposed, how decisions should be made under uncertainty, how people and technology should be jointly designed, how information should flow, and when a process should be redesigned rather than automated.

That history matters because it changes the question. Instead of asking, “How do we add AI to the organization?” a more useful question is: what configuration of work, decision rights, information processing, technology and human capability is now justified because AI changes what the enterprise can do?

1. 1911: work becomes an explicit object of design

Frederick Winslow Taylor’s The Principles of Scientific Management, published in 1911, is an obvious early milestone. Taylor sought to replace rule-of-thumb work methods with systematic observation, task analysis, training and management coordination. The legacy is controversial, and rightly so: extreme decomposition can strip work of context, autonomy and tacit knowledge. But Taylor helped institutionalize an idea that still sits beneath modern transformation programmes — work can be studied, decomposed and deliberately redesigned rather than accepted as fixed.

The AI-era parallel is immediate. Many organizations are now decomposing knowledge work into tasks to identify what a model, copilot or agent might perform. The danger is also familiar: decomposition can improve clarity while simultaneously destroying the context that made good judgement possible. The practical lesson is not “Taylor was right” or “Taylor was wrong.” It is that task decomposition is a design instrument, not a complete theory of work.

2. 1951: technology and the social system cannot be designed separately

E. L. Trist and K. W. Bamforth’s 1951 study of British coal mining examined what happened when a new production technology altered not only the technical workflow but also the social structure and experience of the work group. This line of research became foundational to sociotechnical systems thinking: organizational performance cannot be understood by optimizing the technical system in isolation from people, roles, relationships and local coordination.

That principle is highly relevant to generative and agentic AI. An agent may make a workflow technically faster while weakening accountability, creating hidden review work, confusing escalation paths or eroding the information people need to exercise judgement. A technically capable intervention can therefore be an organizationally poor design.

3. Bounded rationality: organizations exist partly to cope with cognitive limits

Herbert Simon’s work on bounded rationality shifted attention from idealized decision makers to real people making decisions with limited time, information and computational capacity. In his 1978 Nobel lecture, Simon described modern organizations as mechanisms for coping with human limits in the face of complexity and uncertainty.

AI changes part of that constraint. Models can search, summarize, classify, generate alternatives and perform forms of analysis at scales that were previously impractical. But that does not eliminate the organizational problem. Objectives can conflict. Authority still matters. Incentives still matter. Source quality, uncertainty, accountability and the consequences of error still matter.

The relevant transformation question is therefore not whether AI is “smarter” than a person at an isolated task. It is which cognitive constraints have moved, which have not, and how the decision system should change as a result.

4. 1974: organization design as information processing

Jay Galbraith’s 1974 information-processing view linked task uncertainty to organizational form. His framework explicitly connected cognitive limits to design choices involving information systems, lateral relationships and group problem solving. The core logic is powerful for AI transformation: as uncertainty and information-processing demands change, organizational arrangements that were previously adequate may become insufficient — or unnecessarily heavy.

Foundation models alter information-processing capacity dramatically. But the appropriate response is not always “more AI.” Sometimes the better design is to simplify a workflow, clarify decision rights, standardize data, remove an approval layer, or keep a stable human process because its current configuration is already sufficient.

5. 1990: do not automate a process that should be redesigned

Michael Hammer’s 1990 Harvard Business Review article on reengineering argued that organizations should not simply use information technology to accelerate obsolete processes. The article became associated with a broader reengineering movement whose implementation history includes both genuine redesign and excessive, often destructive restructuring. The enduring lesson is narrower and still useful: automation and redesign are different decisions.

That distinction has returned with force. A copilot can assist the current workflow. An agent can execute parts of it. But neither answers whether the workflow itself, the role boundaries or the management routine should remain unchanged. The ability to automate a sequence does not prove that the sequence is still the right one.

6. 2021–2022: general-purpose AI changes the economic boundary of experimentation

Stanford’s 2021 report on foundation models described a shift toward models trained broadly at scale and adaptable across many downstream tasks. This mattered organizationally because AI capability stopped arriving only as narrow systems built for narrowly specified use cases. A common capability layer could now be adapted across writing, analysis, coding, search, support, decision preparation and many other forms of knowledge work.

OpenAI’s public research preview of ChatGPT on 30 November 2022 then lowered the practical barrier to experimentation for millions of people. The enterprise consequence was not simply “better software.” It was a sharp reduction in the cost of trying AI inside everyday knowledge work. Individual experimentation could now spread faster than enterprise architecture, governance, operating-model redesign and value measurement could keep up.

7. 2025–2026: adoption broadens faster than transformation matures

The more recent evidence suggests a widening distinction between AI use and enterprise transformation. McKinsey’s 2025 global survey reported regular AI use in at least one business function at 88 percent of respondent organizations, while only about one-third reported that their companies had begun to scale AI programmes. Its March 2025 research also found workflow redesign to be the factor with the strongest relationship to EBIT impact among the organizational attributes it tested.

PwC’s 2026 Global CEO Survey provides a complementary signal: only 12 percent of CEOs reported both revenue and cost benefits from AI, while 56 percent reported no significant financial benefit to date. These are survey findings, not universal laws, but together they reinforce an important distinction: access, experimentation and even broad usage do not automatically produce enterprise-level value.

The strongest current pattern is therefore not “AI adoption.” It is the emergence of organizations trying to redesign workflows, roles, operating mechanisms and technology foundations around AI while retaining explicit accountability.

8. What changed — and what did not

AI does not make a century of organization theory obsolete. It changes the parameters those theories were trying to manage.

What changed: the breadth, accessibility and rate of improvement of machine capability; the speed at which individuals can experiment; the range of cognitive tasks that can be partially delegated; and the possibility of software that can plan and act across multi-step workflows.

What did not: the need to define value; the limits of data and context; the importance of task interdependence; the reality of conflicting objectives; the need for decision rights; the social consequences of work design; the requirement to coordinate across functions; and the responsibility to learn from evidence rather than momentum.

9. The transformation implication

Enterprise AI Transformation is best understood as the contemporary form of a longstanding organizational design problem under a rapidly moving capability frontier.

The central management task is not to maximize AI. It is to determine, repeatedly and with evidence, what should change, what should remain stable, how work and decisions should be reconfigured, what technology is sufficient, and what evidence justifies the next commitment.

A credible AI transformation system therefore needs both implementation capability and continuously refreshed research intelligence. The technical frontier moves too quickly for a fixed target operating model; organizational history is too rich to justify pretending that every problem began with generative AI.

What remains uncertain

Several questions remain genuinely open: how far agentic systems will reliably operate across high-consequence workflows; which forms of human oversight create real control rather than ceremonial approval; how quickly enterprise architectures will consolidate around agent platforms; how work roles will change once agents become persistent rather than episodic; and how organizations should measure value when AI changes capacity, quality and decision speed simultaneously.