AI and productivity: what the statistics (won't) tell us
The UK lacks timely and granular data on a critical economic variable
The key economic impact of AI can be framed in two contrasting ways. A negative framing focuses on whether it will displace workers. A positive framing asks how much it will boost productivity. In a simple mental model, these are two sides of the same coin: if machines can produce more output, fewer workers should be needed.
There are several problems with this false dichotomy. I have previously discussed the many scenarios in which higher productivity coincides with more, not fewer, jobs. But regardless of which scenario eventually emerges, one thing is clear: we need timely and reliable data to understand what is actually happening to both employment and productivity and how this relates to AI exposure, AI adoption, and other trends.
Unfortunately, and remarkably, the UK is unlikely to publish any good official data on either question in the near future. The ONS recently recommended that, “at the current time, users focus on the experimental RTI approach as the best estimate of how productivity is changing”. Yet, the RTI (real-time information) data—based on HMRC records on the number of employees on PAYE payrolls—lack the necessary granularity to answer the questions above.
1. The RTI labour productivity statistics provide no sectoral granularity
One of the most obvious ways to assess the impact of AI on productivity and jobs is to look at whether industries that have higher AI adoption also have higher productivity growth or faster employment decline, as I did in one of my earlier posts. The ONS’s traditional productivity statistics—where hours worked are based on the Labour Force Survey (LFS)—have commendable granularity, with significant sub-sector detail. In contrast, the RTI productivity statistics provide two very blunt metrics: output per worker and output per hour for the UK as a whole. No sectoral granularity.
2. A “proxy productivity” metric is possible, and somewhat informative
The ONS does publish, separately, sector-level data on PAYE employment and gross value added (GVA). Dividing sectoral GVA by sectoral PAYE employment gives us a proxy for labour productivity, and does so much faster than official productivity statistics. Initial PAYE payrolls data is typically published with less than a one month lag, and GVA statistics come out with about a two month lag. Labour productivity, in contrast, is only reported quarterly, with a roughly 3 month lag.
However, as you would expect, the proxy productivity measure comes with some big caveats. Most importantly, it ignores the effect of self-employed people who do not show up in HMRC statistics. Secondly, it only provides information about output per employee, not my preferred productivity metric, output per hour. In its RTI statistics, the ONS adjusts for these factors, but only at the aggregate level and only with the mentioned 3 month lag.
Nevertheless, I have started using the “proxy productivity” measure as precisely that: a rough indication of what might be happening at the sectoral level closer to real time.
Comparing the proxy against official ONS output-per-hour statistics across 19 sectors and 32 quarters of pre- and post-pandemic data (2015–2025, excluding 2020–21), the proxy tracks the level of productivity across sectors reasonably well (median R² = 0.72). For year-on-year growth — the more relevant measure for understanding trends — performance is more variable, with a median R² of 0.49.
The proxy is most reliable for health and social care, education, transport, finance, and energy (R² above 0.83 for each), and also performs well for manufacturing (R² = 0.74). It is least reliable for hospitality and arts, other services, real estate, and information and communication (ICT). This is consequential: the ICT sector has been the largest single contributor to UK labour productivity growth in recent years but the proxy explains only around a quarter of the variation in annual growth (R² = 0.26).
Unsurprisingly, the proxy productivity metric is weakest in sectors with high self-employment — where workers are invisible to PAYE data — such as construction, arts and recreation, and other services (e.g., hairdressers or personal trainers); and in sectors with highly variable hours per employee, such as hospitality, where nearly a third of workers are on zero-hours contracts, making headcount a poor proxy for hours actually worked. In real estate, the proxy is likely misleading due to the inclusion of imputed rents in the GVA figure. As for ICT, PAYE headcount has been more volatile than hours worked: IR35 reclassification added bodies to payroll in 2022–23 without adding hours, then in 2024–25 headcount fell even as hours held up.
3. Proper productivity data will be needed to decipher AI’s impact
The chart below illustrates how the proxy productivity metric can be used. The bottom left hand pane simply shows what you would expect: absent big swings in output (GVA), productivity looks to have increased most in sectors which also reduced their headcount the most—ICT, hospitality, and retail. Similarly, productivity grew the least in sectors which increased headcount the most (even though for water, waste and energy utilities, and agriculture, a big drop in output also contributed).
The right hand panes are what we are more interested in: did employment decline, or productivity increase, more in sectors with higher AI adoption? Based on this proxy data, the answer is a clear ‘no’. There is no meaningful linear relationship between AI adoption and change in payrolled employees or AI adoption and productivity growth. Specifically, employment declined, and productivity increased, most in both ICT—a very high AI-adoption sector—and hospitality—a fairly low AI-adoption sector. Frustratingly, these are among the sectors for which the productivity proxy is least reliable, making it difficult to draw firm conclusions.
We could look at the LFS-based metrics (and I have), but as mentioned, they are fairly out of date (available until Q4 2025). More fundamentally, the ONS has now recommended not to use them. So, at present, we have a timely measure with little detail, and detailed measures that the ONS itself is increasingly reluctant to rely on. This is not a satisfactory position if we are serious about understanding AI’s impact on jobs and productivity.
The obvious solution is for the ONS either to improve the quality of the traditional productivity statistics, so they can be relied upon again, or to provide RTI-based productivity estimates at a much more granular sectoral level. Academic and other researchers will increasingly complement these efforts with firm-level analyses of AI adoption and productivity. But such studies tend to be available only with a significant lag. We therefore need both: timely official statistics and more detailed research evidence.
