Ph.D., MHA @PQE Group
Knowing this for seventy years has not stopped organizations from doing it anyway. Part 3 of this series described the pattern directly: AI system design typically involves technologists, vendors, and project managers, while the employees who will use the system are consulted last, or not all. McKinsey’s most recent research confirms that the pattern persists at scale. Most organizations, its researchers find, are still applying AI to individual tasks within workflows that were built for a pre-AI world, rather than redesigning the workflow itself, and that gap likely explains why so few see the technology reach the bottom line [1].
Sequential design is not a failure of intent. It is what happens by default when a technical team owns technical decisions and expects organizational development to arrive after the architecture is fixed. AI makes this default costlier than it was with earlier generations of enterprise software, because AI systems do not behave like static tools. They are adaptive, they generate outputs that require human judgment to interpret, and their value depends on how well they are woven into decisions employees are already making. A sociotechnical framework specific to AI integration captures this: AI adoption succeeds or fails less on technical accuracy than on whether the organization builds the social processes—socialization, trust, and shared understanding of roles —needed to integrate it into real work [4]. Design the technology first and treat those processes as an afterthought, and the “afterthought” becomes the whole implementation problem this series has spent four parts diagnosing.
Classical sociotechnical theory treated technology as a fixed object that people had to design around. AI complicates that premise, because AI systems are not fixed; they adapt; they are appropriated in ways their designers never intended (Part 4’s central argument), and they increasingly participate in decisions rather than simply executing them. Recent work extending sociotechnical theory into the AI era captures this shift directly. Researchers Wei Xu and Zaifeng Gao propose what they call an intelligent sociotechnical systems framework, arguing that joint optimization for AI must operate simultaneously across the individual, the organization, the broader ecosystem, and society [5; Figure 1]. In their framing, an AI system is not a tool a social system optimizes around; it is closer to a collaborator inside a human-AI system, one whose design should treat the redistribution of judgment, not just the automation of tasks, as the thing being optimized.
Fig. 1 Intelligent Sociotechnical Systems (iSTS) (adapted from [5])
That reframing matters practically. It means joint optimization is not a single design gate passed through once before launching. It operates at nested levels that shift at different speeds: an individual employee’s changing relationship to their own judgment, a team’s evolving division of labor, an organization’s governance of AI use, and a broader environment of norms, trust, and disclosure the organization does not fully control (the “bring your own AI” dynamic from Part 4 is exactly this last level asserting itself). Design decisions made at one level ripple into the others. A technically excellent tool that redraws decision rights without anyone deciding to redraw them is not a design success; it is an unplanned experiment in organizational structure.
Read against joint optimization, the theories introduced across this series stop being four separate diagnostic lenses and resolve into a single discipline. Organizational Support Theory explains why the social half of the design cannot be an afterthought: employees calibrate their commitment to whether the organization’s actions, not its messaging, signal that their contribution and well-being are valued [6]. Adaptive Structuration Theory explains what happens when the technical and social halves are not designed together: users appropriate the resulting system on their own terms, faithfully when the design fits real work, unfaithfully when it does not [7]. Actor-Network Theory explains why the work is never finished: once an AI system is embedded in daily practice, it becomes an active participant in the organizational network, continuously reshaping the very relationships the original design assumed would hold steady [8].
None of these theories is an optional context for the others. OST tells you whether people will accept the terms of a jointly optimized design. AST tells you, after the fact, whether the design was actually joint or only technical with a social layer added later. ANT tells you that even a well-optimized system will not stay optimized, because the system itself keeps changing what it touches. Joint optimization is not the fifth theory in the series; it is the design discipline the other four exist to protect.
In practice, joint optimization changes where certain conversations happen, not just how many of them happen. Decisions about who retains judgment, which workflows change shape, and how success will be measured belong in the design room alongside the technical architecture decisions, not in a change management plan that follows six weeks later. Frontline employees belong in that room as co-designers, not as a validation panel consulted after the technical direction is set, for the same reason Part 3 argued it: they see the judgment-intensive edge cases and informal dependencies that technical teams, working from process documentation rather than lived experience, routinely miss.
It also changes what gets measured. A technical rollout is complete when the system performs specifications. A jointly optimized system is never finished in that sense, because Actor-Network Theory’s warning holds: the human-AI division of labor keeps shifting after launch. Organizations that take joint optimization seriously treat post-launch monitoring —the stakeholder mapping described in Part 3, the appropriation tracking described in Part 4—as a continuation of design, not a separate governance obligation bolted on afterward. The distinction is not cosmetic. A governance function that treats itself as separate from design will always be reacting to a system that has already drifted.
This is the fifth and final article in this series. The argument has moved in one direction across all five parts: from naming why AI implementation fails as an organizational problem, to locating when organizational support must be built, to mapping where interventions belong across the implementation lifecycle, to confronting what happens once employees, not the organization, control how a system actually gets used. Joint optimization is where those threads meet. It is not a new intervention to add to the first four. It is the standard the first four were always implicitly asking organizations to meet: design the technical and human systems as one system, because they were never really two. The organizations most likely to close the value gap McKinsey is now measuring will not be the ones with the most capable AI. They will be the ones willing to treat the redesign of work, roles, judgment, and decision rights as part of the AI implementation itself, not as a downstream people problem to be managed after the technology ships. That is organizational development work, and it is exactly the discipline this series has argued AI implementation has been missing. For organizations working through what that redesign looks like in their on operating environment, PQE Group’s advisory practice partners with clients on exactly this kind of sociotechnical implementation work, connecting organizational development expertise to the technical and regulatory realities of the industries we serve.
The following steps operationalize the principles above. They are intended for planning and future application:
This is article number 5 of our series on driving successful AI adoption through organizational change. Explore the previous pieces below:
1. Why 93% of Companies Are Implementing AI but Only 33% of Employees Know About It
2. Before You Flip the Switch: Building Organizational Support Before Introducing AI Systems
References
[1] McKinsey & Company. (2026). A new year's resolution for leaders: Redesign work for people and AI. https://www.mckinsey.com/mgi/media-center/a-new-years-resolution-for-leaders-redesign-work-for-people-and-ai
[2] Trist, E. L., & Bamforth, K. W. (1951). Some social and psychological consequences of the longwall method of coal-getting. Human Relations, 4(1), 3–38. https://doi.org/10.1177/001872675100400101
[3] Walker, G. H., Stanton, N. A., Salmon, P. M., & Jenkins, D. P. (2008). A review of sociotechnical systems theory: A classic concept for new command and control paradigms. Theoretical Issues in Ergonomics Science, 9(6), 479–499. https://doi.org/10.1080/14639220701635470
[4] Makarius, E. E., Mukherjee, D., Fox, J. D., & Fox, A. K. (2020). Rising with the machines: A sociotechnical framework for bringing artificial intelligence into the organization. Journal of Business Research, 120, 262–273. https://doi.org/10.1016/j.jbusres.2020.07.045
[5] Xu, W., & Gao, Z. (2025). An intelligent sociotechnical systems (iSTS) framework: Enabling a hierarchical human-centered AI (hHCAI) approach. IEEE Transactions on Technology and Society, 6(1), 31–46. https://ieeexplore.ieee.org/document/10744034
[6] Eisenberger, R., & Stinglhamber, F. (2011). Perceived organizational support: Fostering enthusiastic and productive employees. American Psychological Association. https://doi.org/10.1037/12318-000
[7] DeSanctis, G., & Poole, M. S. (1994). Capturing the complexity in advanced technology use: Adaptive structuration theory. Organization Science, 5(2), 121–147. https://doi.org/10.1287/orsc.5.2.121
[8] Walsham, G. (1997). Actor-network theory and IS research: Current status and future prospects. In A. S. Lee, J. Liebenau, & J. I. DeGross (Eds.), Information systems and qualitative research (pp. 466–480). Springer. https://doi.org/10.1007/978-0-387-35309-8_23
