Sociotechnical systems in software and AI: Who actually captures the productivity gains?

AI promises big productivity wins, but sociotechnical realities mean the gains often land unevenly. This post untangles the 'last mile' problem, who benefits today, and pragmatic steps organisations can take to close the value gap.

Introduction: the promise versus the landing zone

It is tempting to imagine AI as a magical productivity multiplier that simply plugs into work and everyone gets richer — faster code, sharper analysis, fewer meetings. The reality, as multiple investigations and practitioner reports show, is more sociotechnical than technomagical. AI changes the software and work systems it touches, and whether value arrives at the boardroom, the bench, or the individual engineer depends on how those systems are reorganised.

The "last mile" problem: AI does 80%, humans do the rest

One clear finding from MIT Sloan's research is a practical one: AI often gets teams about 80% of the way to a finished output, but the final 20% needs human judgement, context and domain expertise. Tasks that map cleanly from input to output — a meeting transcript to a summary, for example — are high-fit and show quick wins. Tasks that require external knowledge, strategic framing or nuanced evaluation — such as competitive analysis — are low-fit and rely heavily on human validation.

That 80/20 split is crucial because it reframes popular claims about productivity. AI improves task-level performance substantially, but the sociotechnical work of integrating those outputs into decisions, products and workflows is what actually realises value.

Who is capturing the gains today?

Evidence suggests the winners are not just those who buy models, but those who rebuild systems around them. BCG's research identifies a set of "future-built" companies that capture far more value: roughly five times the revenue impact and three times the cost reduction from AI compared with laggards. These firms invest in capabilities, reinvest returns into people and architecture, and even allocate meaningful budget to agentic AI — BCG notes roughly 15% of AI budgets going to agents in their examples.

Put simply: the gains concentrate where organisations have aligned incentives, modern infrastructure, and processes that make human-and-AI collaboration routine rather than exceptional.

Why the value gets stuck — the sociotechnical bottlenecks

There are several recurring bottlenecks where AI's promise stalls:

Task fit and the last mile — as MIT Sloan shows, many outputs need human synthesis. If organisations don't redesign roles and handoffs, AI assists but does not displace labour or create measurable output changes.
Execution and modular architecture — Bain's discussions with financial services leaders emphasise that strategy alone is insufficient. Execution depends on modular, cloud-native architectures, reusable interfaces for agents, and governance that lets new systems coexist with legacy code.
Organisational incentives and measurement — pilots can boost individual tasks but fail to translate to KPIs. Forbes and other analyses point out a productivity paradox: large task-level gains (14–55% in some studies) have not yet moved macroeconomic indicators, and many pilots (the studies suggest a high failure rate) do not scale.
Cultural and skills gaps — AWS's practical guidance highlights the need to change processes, roles and incentives across several dimensions (strategy, processes, culture, infrastructure, skills and governance) to close the gap between AI experiments and sustained value.

Where the value lands in practice

When organisations do capture value, it tends to land in a few places:

Product teams that redesign workflows — embedding agents in existing flows (customer service, developer tooling) converts task output into user-visible improvements.
Operational centres that automate repeatable work — finance, compliance and routine data ops where AI reduces manual toil and human reviewers handle exceptions.
Startups rethinking sociotechnical paradigms — as Eli Dukes argues, new businesses that stitch LLMs, vector databases (the long-term memory for models) and novel interfaces can create entirely new work systems that capture disproportionate value.

Practical guidance: a sociotechnical checklist to capture AI value

If you are a leader, product manager or engineer trying to turn AI promise into predictable outcomes, here are concrete moves rooted in the evidence:

Map task fit, not just features — build a task-level integration map (MIT Sloan). Classify tasks as high-fit (low human judgement) or low-fit (high domain oversight) and design collaboration patterns accordingly.
Design for the last mile — accept that humans will often finish the job. Allocate time, craft review checklists and instrument handoffs so AI outputs are easy to vet and compose into final artifacts.
Invest in modular, cloud-native stacks — follow Bain's playbook: use modular services and APIs so models and agents can be iterated independently of legacy systems, and so integrations scale across teams.
Align incentives and KPIs — measure outcomes you care about (revenue, cycle time, error rates), not just model metrics. Reinvest early returns into people and platform, as BCG reports future-built companies do.
Close the skills and culture gap — run focused change programmes across the seven dimensions AWS recommends: strategy, process redesign, culture, infrastructure, skills, governance and measurement.
Pilot with exit criteria and scaling paths — design pilots that test both technical feasibility and process fit; require explicit plans for operationalisation to avoid the common pilot graveyard.
Consider new sociotechnical platforms — for transformational opportunities, evaluate whether a new product that reconfigures data, memory (vector DBs) and agents can create superior workflows rather than retrofitting old ones.

Measuring success realistically

Expect a phased curve. Forbes highlights that some sectors experience initial productivity dips as systems and labour reconfigure — manufacturing studies found substantial early declines before recovery over several years. At the macro level, early adopters may show task-level gains of 14–55%, but aggregate measures lag while adoption, process change and reinvestment unfold.

That means you should track short-term leading indicators (time saved on tasks, error reduction, user satisfaction) and medium-term business metrics (throughput, revenue per head). Don't conflate a tidy model metric with realised organisational value.

Conclusion: capture is a sociotechnical problem

AI is not magic; it is a powerful component in sociotechnical systems. The productivity gains it offers are real at the task level, but whether those gains benefit individuals, teams or firms depends on architecture, incentives, skills and how the last mile is organised. Future-built companies that explicitly redesign systems and reinvest returns are already pulling ahead. For everyone else, the practical work of mapping tasks, reworking workflows and aligning incentives is where the race will be won — or lost.