AI and productivity: the ladder of leakage

Why AI-driven productivity gains leak out at every stage of aggregation

I expect AI to be transformational for productivity—but only in the medium- to long-term. Right now, productivity gains are real but dispersed and uneven. At the macro level in the UK, they are undetectable. Encouragingly, firm-level analysis is starting to find benefits: a study by CEBR finds that AI adoption increased European firms’ labour productivity levels by 4% (with no evidence of reduced employment).

However, because AI adoption remains patchy, it will take some time for productivity gains to come through. More importantly, even when adoption happens, it is often ineffective: improved output per hour for specific tasks doesn’t always add up to improved productivity at the individual, team, function, firm, sector, or economy level. I call this the “ladder of leakage”. As the chart below shows, adoption remains limited and, crucially, most firms have not yet moved beyond pilots into scaled deployment.

Productivity gains from AI are not automatic

It is clear that AI can improve the quality and quantity of many labour outputs. But realising those gains requires deliberate effort. Without it, they dissipate at every step:

• Task → individual: time saved is not redeployed to higher-value activities, and may be absorbed as slack

• Individual → team: working patterns and roles remain unchanged, creating inefficiencies rather than higher output

• Team → function: gains are too small or fragmented to change headcount, hours, or scope; caution keeps humans in the loop even for low-risk tasks

• Function → firm: budgets and resources are not adjusted to capture efficiencies, nor redirected toward higher-value activities

• Firm → market: competitive advantages translate only slowly into market share, due to switching costs, risk aversion, and inertia

• Market → economy: gains in faster-improving sectors are offset by shifts toward slower-productivity-growth activities (Baumol’s disease)

At each level, frictions prevent improvements from adding up to system-wide gains.

Reallocation is the key bottleneck

Much of the guidance on capturing productivity gains from AI emphasises redesigning end-to-end workflows and the roles of the workers involved. This can, indeed, unlock significant value.

But at a more fundamental level, realising those gains—at each rung of the “ladder of leakage”—depends on reallocation. Time saved must be actively redirected to higher-value activities. This is particularly important at the firm level. Past analysis suggests organisations tend to remain anchored to existing budgets rather than reallocating resources toward emerging opportunities. In one survey, a third of companies said they reallocated just 1% of their capital annually. The same likely applies to headcount.

From a public policy perspective, reallocation between firms is equally important. Research shows that in the US between 2011 and 2019, the scaling of more productive firms contributed 0.6 percentage points to annual productivity growth, while the exit of less productive firms added a further 0.5 percentage points. Overall, reallocation accounted for nearly half of total productivity growth. Over the same period, UK productivity growth among comparable firms was effectively zero. Weak business dynamism is increasingly recognised as a key constraint—and may be what limits the economy-wide payoff from AI in the UK.

Leadership and management make a big difference

I will return to the centrality of leadership and management in identifying and capturing opportunities from AI in future posts. Clearly, senior decision makers are pivotal in driving the kind of reallocation mentioned above. But their influence and importance goes well beyond this. Firms with better management practices are much more likely to adopt AI. So are businesses where AI adoption is actively encouraged. Businesses with leaders who demonstrate strong ownership and commitment to their AI initiatives are much more likely to be AI high performers.

The bottom line is clear. AI is already improving individual tasks. But whether those gains translate into higher productivity depends on whether organisations are willing—and able—to reallocate time, resources, and effort. That is a much harder problem than deploying the technology.