Sociotechnical Systems Before Generative AI

Long before generative AI, organizational researchers had already encountered a recurring transformation failure: a technically superior production system could produce worse organizational outcomes when the surrounding work system was designed badly. Sociotechnical systems research gave that problem a durable form. The technical system and the social system are interdependent. Changing one changes the conditions of the other.

That principle is unusually useful in 2026. Foundation models, copilots and agents can alter who performs a task, how quickly information moves, which decisions can be delegated, where exceptions appear and what expertise remains visible. Treating those capabilities as software features misses the larger design problem. The relevant unit is the Human–AI work system.

1. The coal mine was an organization-design experiment

E. L. Trist and K. W. Bamforth’s 1951 study of British coal mining examined the consequences of changing from an older form of group-based mining to the longwall method. The new production technology reorganized the work itself. Tasks became more specialized and spatially separated; coordination patterns changed; workers became more dependent on conditions outside their immediate control.

The importance of the study was not nostalgia for an older method. It showed that technical efficiency could not be evaluated independently of the social organization required to make the system function. A production technology carries assumptions about roles, coordination, autonomy, information and control. When those assumptions fit poorly with the operating environment, the technical improvement can create new losses elsewhere.

2. Joint design is different from technology plus change management

A common transformation sequence selects technology first and treats organization as a downstream adoption problem: configure the platform, train users, communicate the change, monitor uptake. Sociotechnical reasoning reverses part of that logic. Work design and technical design are coupled choices.

For enterprise AI, that distinction is consequential. Consider a system that drafts a maintenance recommendation, prepares a credit decision, triages customer cases or assembles management analysis. The model’s output quality is only one variable. The intervention also changes who sees the evidence, who checks exceptions, how much context remains available, whether expertise accumulates or atrophies, what happens when the system is uncertain, and who owns the resulting action.

Those are not post-deployment concerns. They determine the architecture of the intervention.

3. AI makes task boundaries fluid

Traditional automation often replaced or accelerated relatively specified activities. General-purpose AI makes the boundary more fluid because the same capability can summarize, classify, draft, retrieve, reason over context and interact with tools. NIST’s 2024 AI Use Taxonomy reflects this shift by describing Human–AI work in terms of outcome-oriented activities rather than tying use to a particular AI technique or domain.

This matters organizationally because a role is not a bag of independent tasks. Tasks share context, tacit knowledge, relationships, sequencing and accountability. Removing one activity can change the information available for another. Accelerating analysis can move the bottleneck into review. Automating routine cases can leave people handling only difficult exceptions, changing the expertise and cognitive load required of the role.

The design question is therefore larger than “what percentage of this job can AI do?” It is what configuration of tasks, context, authority and capability produces the strongest work system after AI changes the feasible allocation?

4. Exposure is not the same as replacement

The International Labour Organization’s 2025 global analysis reinforces the task-system view. Its refined index assesses nearly 30,000 occupational tasks and finds that one in four workers globally are in occupations with some degree of generative-AI exposure. Yet the report concludes that transformation of work is more likely than wholesale replacement because most occupations retain tasks requiring human input.

That finding shifts attention from job counts to work design. If occupations are transformed unevenly at task level, organizations need to decide how the remaining human work changes: whether judgement becomes more concentrated, whether review burdens rise, whether workers retain enough primary experience to detect machine error, and whether productivity gains translate into better work or simply higher work intensity.

The ILO’s earlier work also highlighted autonomy and work intensity as potential consequences. Sociotechnical design makes those variables part of performance rather than treating them as externalities.

5. Human oversight has to be designed as work

“Human in the loop” sounds reassuring but says little about the actual operating system. A person may formally approve an AI recommendation while lacking the time, evidence, expertise or authority to challenge it. Conversely, excessive review can recreate the full manual process and eliminate the value of automation.

NIST’s AI Risk Management Framework connects technical design to organizational roles, responsibilities, competencies and lifecycle processes. Its actor model explicitly includes domain experts, human-factors specialists, governance roles, data specialists and affected communities alongside technical developers. That is a contemporary expression of a sociotechnical reality: reliable AI operation depends on a configuration of actors, not a model in isolation.

Useful oversight therefore requires concrete design choices: what evidence accompanies the recommendation; which uncertainty is visible; what triggers escalation; what the reviewer can change; which actions can proceed automatically; how overrides are captured; and how the organization learns from disagreement between people and the system.

6. Local optimization can damage the enterprise system

AI makes local optimization cheap. A team can automate drafting, analysis or case handling without waiting for enterprise redesign. That speed is valuable, but it can also produce fragmented work systems: duplicated agents, conflicting data access, invisible handoffs, inconsistent decision rights and new manual reconciliation between locally optimized processes.

Sociotechnical thinking suggests a practical test. Evaluate the intervention at more than one boundary. Does it improve the task? Does it improve the workflow? Does it improve the role? Does it improve cross-functional coordination? Does it improve the enterprise outcome? A gain at one level can become a cost at another.

7. Autonomy is an allocation problem

The Tavistock Institute’s current work extends sociotechnical thinking into Human–AI systems through the idea of adaptive autonomy: autonomy is distributed and can change as tasks, context and system capability change. That is a better description of enterprise AI than a fixed binary between manual and autonomous work.

An AI system may prepare evidence in one workflow, recommend an action in another, execute bounded actions in a third and operate with broader autonomy in a low-consequence technical domain. The allocation should depend on consequence, reversibility, observability, reliability and the quality of available context.

This makes autonomy an operating-model variable. As model capability improves, the enterprise can change the allocation — but only if measurement and feedback reveal whether the previous boundary worked.

8. The practical architecture of a sociotechnical AI intervention

A useful intervention can be designed through five linked questions:

1. Outcome: what observable enterprise result should improve?
2. Work: which tasks, decisions, handoffs and exceptions produce that result today?
3. Allocation: what should people, models, rules and software each do in the new configuration?
4. Operating conditions: what authority, information, controls, skills and coordination does the configuration require?
5. Learning: what evidence will show whether the configuration should expand, change or stop?

This sequence prevents the model from becoming the center of the architecture. The model is a capability inside a work system whose purpose is an organizational outcome.

9. Why the old theory becomes more useful as AI improves

Better AI does not make organizational design less important. It increases the number of feasible configurations. When technology can perform only one narrow task, the design space is constrained. When models and agents can participate across many cognitive activities, enterprises have more choices about delegation, coordination, role boundaries and process structure.

More possibility creates a larger architecture problem. The advantage shifts toward organizations that can redesign work deliberately, instrument the new configuration and learn faster than the capability frontier moves.

Sociotechnical systems thinking therefore survives generative AI for a simple reason: technology changes what is possible; organization determines what that possibility becomes in practice.

10. From joint optimization to co-evolution

Recent research pushes the sociotechnical argument beyond a one-time design exercise. Parker and colleagues describe a co-evolving sociotechnical systems perspective for AI-enabled work: technical capability, work practices and organizational arrangements continue changing after deployment. This matters because generative AI is unusually mutable. Models change, interfaces change, workers develop new practices, managers alter controls and organizations discover uses that were not anticipated at implementation.

The implication is operational. A Human–AI configuration should be treated as a versioned operating design rather than a finished rollout. Teams need to observe where judgement is improving, where review is becoming ceremonial, where cognitive load is moving, where capability is eroding and where new coordination costs are appearing.

11. AI adds new design variables, not a new organizational universe

Research published in 2025 on AI-enhanced human–robot teams adapts established sociotechnical principles to newer conditions. It adds explicit attention to adaptive autonomy, cognitive workload, collaborative sensemaking, transparency, trustworthiness and unpredictable behavior. Work-system research on generative AI similarly argues that AI can no longer always be treated as a passive technical component; in some settings it behaves more like an intelligent actor participating in the work.

These extensions matter, but they reinforce rather than erase the older lesson. Human–AI work still has to be designed around interdependence. The new variables are the speed with which allocation can change, the breadth of cognitive tasks available to machines, the opacity of some model behavior and the possibility that software can now coordinate or act through tools.

12. A compact design test for enterprise AI

Before scaling an AI intervention, test the whole work system across six dimensions:

1. Outcome: is the enterprise result improving, or only local task speed?
2. Allocation: are human and machine contributions deliberately assigned?
3. Authority: is the right to recommend, approve and act explicit?
4. Capability: are people gaining the skills needed for the redesigned work, including effective challenge?
5. Coordination: did the intervention remove handoffs and friction, or create new reconciliation work elsewhere?
6. Learning: can evidence from use change the workflow, controls and autonomy boundary?

If these questions cannot be answered, the organization has probably deployed an AI capability without yet designing the Human–AI work system around it.

Open questions

Several questions remain empirical: which Human–AI allocations preserve enough direct experience for people to remain capable reviewers; when agent autonomy improves system resilience versus creating opaque dependencies; how productivity gains affect work intensity and learning over time; and which organizational structures can revise Human–AI boundaries quickly without producing constant operating instability.