Sociotechnical systems in software and AI: Who really gets the productivity gains?

AI promises big productivity boosts, but sociotechnical realities shape where the value actually lands. This practical post explains the gaps between promise and payoff, who benefits today, and how teams can steer gains towards people and organisations rather than vendors or dashboards.

Introduction: a familiar mismatch

When a new wave of tools arrives—think electricity in factories, or the first enterprise software—people picture instant productivity miracles. With AI, the headlines promise similar leaps. Yet the reality is messier. The benefits of AI are real at task level in many studies, but where the value actually lands depends on a sociotechnical ecosystem: people, practices, platforms, incentives and infrastructure all need to adapt. Treating AI as a plug‑and‑play productivity hack is a good way to be disappointed.

What the evidence actually shows

Micro studies are generally encouraging. Recent surveys of the literature find mixed but mostly positive productivity effects at the task level, with some experiments reporting substantial improvements—sometimes 50% or more—for particular activities (for example, coding or drafting) while other tasks show modest gains (Alex Imas summarises this research well).

At the same time, macroeconomic and firm‑level measures are more conservative. Some analysts expect AI could raise long‑run growth materially (estimates of 1–1.5 percentage points annual uplift appear in several reports), while other economists argue the aggregate effect may be much smaller—on the order of 0.1% per year—if adoption and integration are slow or uneven (as discussed in CEPR summaries).

There is also persuasive evidence that new general‑purpose technologies produce a productivity J‑curve: investments and organisational change precede measurable gains by years. In other words, money and effort must be spent first on infrastructure, skills and process redesign before output metrics catch up (a point highlighted in analyses that draw on Brynjolfsson and Syverson’s work).

Why the value often doesn't land where expected

Several sociotechnical frictions explain the gap between AI's promise and realised value:

Infrastructure and capital: AI adoption often requires new datacentres, hardware and platform investments. Citigroup's analysis emphasises that the era of AI is still very much an adoption phase with capital being built out—so the firms that supply AI infrastructure and tools often capture early returns.
Organisational adaptation: Tools change workflows. Without redesigning processes, incentives and governance, gains at the task level don't aggregate to firm‑level productivity. That J‑curve is basically bureaucracy having a nap while the machines learn their lines.
Uneven adoption: Representative data suggests relatively low formal adoption in many contexts—one estimate puts AI use under 5% of firms in the US—so aggregate effects are muted simply because not enough organisations have reconfigured around AI yet.
Skills and roles: Where AI is adopted, the benefits often cluster with teams that have both the technical skills and the mandate to change processes. Surveys show many companies are investing in training (nearly half reporting training programmes), but this is not uniform across sectors or firm sizes.
Vendor versus user capture: Early in the cycle, enabler companies—those making AI models, chips and cloud stacks—tend to capture a disproportionate share of value. Only later, with deeper integration and business model adjustments, can adopters capture more of the downstream productivity gains.

Who benefits now (and who might later)?

Immediate winners tend to be:

AI enablers and platform providers who sell compute, models and tools.
Organisations that already had strong data infrastructure, skilled staff, and capital to invest—these groups can pilot, iterate and scale faster.
Individual workers whose tasks are highly automatable and who get effective tooling and training; some studies suggest that AI can equalise outcomes for less experienced workers in certain contexts.

Potential later winners include broader groups of adopters once the organisational changes—process redesign, governance, measurement and training—have been implemented. Surveys of firms that have deployed AI report widespread perceptions of improved productivity, and many enterprises are already offering training and creating roles centred on AI. That pattern fits the classic diffusion story: early adopters and suppliers first, mainstream benefits later.

Practical guidance: how to make the gains land where you want them

If you're leading a team, or building software that leverages AI, the sociotechnical perspective helps guide practical choices. Here are steps that improve the chances that productivity gains actually accrue to your organisation and people.

Start from the task, not the hype. Break down workflows into discrete tasks, measure baseline performance, and pilot AI on the highest‑value, well‑measured activities. Micro studies show the clearest benefits at the task level; use that evidence to choose where to experiment.
Design the sociotechnical system. Treat AI as part of a system that includes process, incentives, training and governance. Bring product managers, engineers, operators and end users into the same room and make shared success metrics explicit. This reduces the risk that improved tool outputs get bottlenecked by unchanged processes.
Invest in measurement and feedback loops. Create short learning cycles to capture real productivity changes and side effects. The J‑curve is less scary if you can see incremental wins and course‑correct quickly.
Align incentives. If vendors are paid per‑seat or per‑API call, their incentives differ from yours. Consider contracting, pricing and internal reward structures that share gains with teams and individuals who adopt and improve performance.
Prioritise uplift in skills. Practical training—embedded in workflows—helps more workers benefit. Data from corporate surveys shows many organisations are already offering training; make yours targeted and hands‑on.
Think long term about infrastructure. Some gains require reliable data pipelines and compute. Balance short pilots with a roadmap for the necessary engineering investments so quick wins can scale.
Consider equity and diffusion. If policy or HR is within your remit, design pathways so productivity gains don't accrue only to a small subset of staff. Structured upskilling, job redesign and transparent promotion paths help spread benefits.

Final thought: a pragmatic lens

AI is neither a magic wand nor merely a vendor story. It is a sociotechnical force that delivers value when the technical layer is stitched into the social fabric of organisations. The literature and industry reports show task‑level gains, a likely lag before broad macro benefits, and uneven distribution of early returns. Leaders who plan for the sociotechnical work—measurement, process redesign, investment and fair incentives—stand the best chance of turning promising models into realised productivity improvements that benefit more than a few.