Is AI (finally) making us more productive? With John Burn-Murdoch and Sarah O’Connor

Episode Summary

Executive Summary: The episode examines whether AI is boosting productivity, concluding that macro data still shows little clear evidence, but micro-level signs—especially in software—are increasingly suggestive. The hosts argue AI's impact is uneven, often hidden by bottlenecks, quality-control costs, regulation, and measurement issues, while agentic tools may be starting to create real output gains in certain sectors.

Main Topics: Macro productivity data still lags (Priority: 5/5): The hosts argue that GDP and labor-productivity statistics are backward-looking and do not yet show a broad AI-led productivity boom, though some analysts believe the U.S. may be nearing an inflection point. Why micro evidence can conflict with macro stats (Priority: 5/5): Individual users report big time savings, but organizations are systems with bottlenecks, review processes, and quality controls that can erase or obscure gains when measured at scale. Industry-level studies on AI adoption and productivity (Priority: 4/5): They discuss analyses comparing AI adoption rates with productivity growth across sectors, noting weak or mixed correlations in the UK and more promising but still limited evidence in the U.S. Software as the leading indicator (Priority: 5/5): Software development is presented as the clearest area where AI output may already be rising, especially as agentic tools can generate, test, and iterate code in ways that fit deterministic workflows. Productivity gains vs job displacement (Priority: 4/5): The conversation distinguishes higher productivity from fewer jobs, invoking Jevons paradox and stressing that cheaper, faster production can expand demand rather than reduce headcount in some fields. Which professions are most exposed (Priority: 4/5): The hosts judge regulated fields like medicine, radiology, and law as more protected because humans must sign off, while creative and freelance work appears more vulnerable because cheaper AI output can replace human labor. Personal reactions and job redesign (Priority: 3/5): Both hosts describe feeling both excitement and anxiety about AI, with concern shifting from total automation to job redesign, loss of control, and the possibility that work will be reorganized around AI.

Key Arguments: Current macroeconomic data does not yet show a decisive AI-driven productivity jump, but this may reflect data lags and measurement limitations rather than absence of impact. Claims of AI productivity gains are strongest at the task level, yet organizations operate through interdependent systems, so one faster task can create bottlenecks elsewhere. Quality matters: if AI increases output while lowering accuracy or requiring extra checking, measured productivity gains may be overstated or reversed. Industry-level evidence is mixed: UK adoption-productivity correlations look absent or negative, while U.S. evidence is beginning to look more positive but remains statistically thin. A broader, longer-run definition of AI and a focus on software suggests real productivity gains may already be embedded in recent U.S. tech-sector growth. Agentic tools appear to be changing software development more visibly than earlier LLMs because code is deterministic and can be checked, fixed, and iterated automatically. Regulated sectors are likely to keep human oversight because insurers, regulators, and liability concerns require a responsible sign-off, limiting full automation. Creative industries and freelance work may be more exposed because AI can produce cheaper, adequate substitutes that firms may prefer over higher-quality human work. Job losses are not the same as productivity gains; in some areas lower costs could expand demand and preserve employment, while in others they may reduce labor needs. Much of the apparent slowdown in graduate hiring may be due to macroeconomic conditions and interest-rate changes rather than AI alone.

Data Points: Bullishness on AI productivity now (Sarah): 3/10 - Sarah's rating of AI's current productivity impact Bullishness on AI productivity in five years (Sarah): 8/10 - Sarah expects stronger productivity effects later Bullishness on AI productivity now (John): 4/10 - John's rating for more nitty-gritty productivity effects Bullishness on AI productivity in five years (John): 7/10 - John's forward-looking rating Goldman Sachs corporate anecdote average: 32% productivity increase - Referenced as a compilation of company stories about AI savings Meter study sample size: 16 software engineers - Small but carefully run study on perceived vs actual productivity with AI Meter study perceived productivity before work: 20-30% more productive - Engineers' expectation before using AI tools Meter study perceived productivity during/after work: 20% more productive - Engineers' self-assessment after using AI Meter study measured productivity: 20% less productive - Actual measured outcome in the study Software share of non-farm U.S. economy: 6% - Used in the longer-run AI/productivity paper discussed in the episode Share of productivity-growth acceleration attributed to software: about one-third - Same paper's finding over a multi-year period Time window for software output inflection: late 2025 into this year - Hosts say output data began rising after agentic tools came online Time lag in macro data: about 6 months - Government statistics are described as lagged and rearview-mirror-like

Pivotal Quotes: "I don't think we see signs yet of a massive liftoff in that microdata." — John Burr Murdoch: On the lack of clear macro evidence that AI has already produced a broad productivity boom "Jobs are not tasks, and organizations are certainly not tasks." — Sarah O'Connor: Explaining why task-level AI speedups do not automatically translate into firm-wide productivity gains "I think it's pretty much impossible now to look at the data and argue that we're not seeing an increase in output." — John Burr Murdoch: On recent software-output evidence after agentic tools became available

Implications: AI’s productivity impact is likely real but uneven, appearing first in software and other deterministic work. For most sectors, the bigger story may be job redesign, quality control, and bottlenecks rather than immediate mass automation.