Investor Gavin Baker wrote on X on 1 November that it was hard to look at Financial Times productivity charts “and not see some evidence of “ROI on AI.”” He added that “America is 72% of Anthropic revenue vs. 50%ish at most application software companies”. Anthropic’s S-1 puts the United States at $11,763 million of $16,286 million in revenue for the six months to 30 June 2026, or 72.2% by our arithmetic.
Ben Pouladian posted a chart the same evening, credited to BEP Research. It says information, finance and insurance, and professional, scientific and technical services produced real output up 19.8% since the fourth quarter of 2022 with payroll jobs down 1.0%. Our own count of Bureau of Labor Statistics payroll data matches the 1.0% fall, and shows where it came from: information lost about 325,000 jobs, or 10.5%, while professional and technical services added about 163,000.
The key insight: Payroll jobs in three AI-heavy sectors fell 1.0% between late 2022 and mid-2026, matching the figure on BEP Research’s chart, but by our count of BLS data the whole fall sits in information, down 10.5%, while professional and technical services grew 1.5%.
Baker posted at 19:19 UTC with three images: an FT chart titled “Productivity is surging in AI-adopting sectors in the US, but is AI the cause?”, a close-up of its US panel, and a screenshot of an EPB Research post. He described “a productivity surge post-Covid era overhiring that began to reverse before ChatGPT came out and then resumed its upward trend after ChatGPT.” He called the FT work “Great stuff from @jburnmurdoch even if I disagree with his conclusion.” His post had about 140,590 views when we read it on 2 November.
Pouladian quoted Baker at 19:34 UTC. His chart’s header reads: “Selected sectors since 2022 Q4: real output +19.8% | payroll jobs −1.0% | output per job +21.0%”. It puts the sectors’ output per job on a 5.1% annual trend since the end of 2022, up from 2.3% over 2005 to 2022, against 0.9% to 1.4% for the rest of the private economy. The chart labels this an “Original BEP construction, not an EPB replication”. EPB Research’s own figures, a 2.2% trend rising to 5.0%, are in our earlier piece on EPB’s sector split.
Pouladian also wrote: “Does this prove AI caused the acceleration? No.” His chart’s footnote says it is an “Observational comparison, not causal AI evidence” and that output per job is “Not hours-adjusted productivity”.
What Was Posted on X
Hard for me to look at these charts from the FT, especially the first one, and not see some evidence of โROI on AI.โ
Seems like a productivity surge post-Covid era overhiring that began to reverse before ChatGPT came out and then resumed its upward trend after ChatGPT.
Especially given the @a16z data on revenue per employee at their portfolio companies and the @EPBResearch data showing that the 3 industries that have most heavily adopted AI have gone from 2.2% productivity growth to 5% productivity growth.
And makes sense that it is showing up in the US first – America is 72% of Anthropic revenue vs. 50%ish at most application software companies.
Great stuff from @jburnmurdoch even if I disagree with his conclusion.
Gavin Baker’s post of 1 November, 19:19 UTC, sharing FT charts of output per worker in AI-adopting sectors: he says it is hard not to see some evidence of “ROI on AI” and that “America is 72% of Anthropic revenue”. Anthropic’s S-1 gives 72.2% for the latest half, by our arithmetic.
The AI ROI debate is getting harder to dismiss.
Since late 2022, information, finance and professional/technical services produced **19.8% more real output with 1% fewer payroll jobs**.
Output per job is growing at a 5.1% annual trend, up from 2.3% before 2023. The rest of the private nonfarm economy 1.4%.
Meanwhile, Anthropic reports over 6,000 customers spending at least $100K annually, versus over 600 a year earlier. Microsoft says automated tasks nearly doubled since July.
This is the shift weโve been writing about: **from answers to work.** Agents doing useful work on enterprise data, inside governed workflows.
Does this prove AI caused the acceleration No. But โwhereโs the ROI?โ deserves a better answer than another chatbot joke.
Ben Pouladian’s post quoting Baker 15 minutes later, with a chart credited to BEP Research: real output up 19.8% since late 2022 in three AI-heavy sectors, payroll jobs down 1.0%. He adds that this does not prove AI caused the acceleration; our BLS count puts the job fall in information.
Payroll jobs in information, finance and insurance, and professional, scientific and technical services, indexed to the fourth quarter of 2022, from BLS Current Employment Statistics; quarterly averages, index values and job changes are our arithmetic, and the latest months are preliminary.
Where the Payroll Jobs Went
We took the BLS seasonally adjusted payroll counts for the three sectors and averaged each quarter’s three months. Combined, they went from 20.47 million jobs in the fourth quarter of 2022 to 20.27 million in the second quarter of 2026, a fall of about 199,000, or 1.0%, by our arithmetic. That matches the payroll figure on Pouladian’s chart.
The fall is not spread evenly. Information went from 3.10 million to 2.78 million jobs, down about 325,000, or 10.5%. Finance and insurance went from 6.71 million to 6.67 million, down about 37,000, or 0.6%. Professional, scientific and technical services rose from 10.66 million to 10.82 million, up about 163,000, or 1.5%. Information’s loss alone is larger than the three sectors’ combined net fall, by our arithmetic. The latest BLS months are marked preliminary.
Our count covers payroll jobs only. We did not recompute the 19.8% output figure, which Pouladian’s chart says comes from the Bureau of Economic Analysis’s GDP-by-industry data with BEP’s own aggregation.
Business Pill · The Jevons Paradox Explained
A short Business Engineer explainer on why cheaper, more efficient inputs can raise total use. It teaches the general idea only and says nothing about these payroll or output figures.
Why Baker Points to the US
The FT chart Baker shared has eight panels: the US, the UK, Canada, Australia, the EU27, France, Germany and Sweden. Each plots real output per worker in AI-adopting sectors, defined as tech, finance and professional services, against a 2012-19 trend. The US panel carries a ChatGPT marker; the other seven do not. Baker wrote that “makes sense that it is showing up in the US first”, pointing to Anthropic’s revenue mix.
The S-1 table on revenue by region, based on the customer’s billing address, supports his 72% figure. The US share was 70.5% in the six months to 30 June 2025, at $820 million of $1,163 million, by our arithmetic, so the share was already about seven in ten a year earlier. The S-1 also says customers with $100,000 or more of revenue on a trailing basis grew from more than 600 to more than 6,000 over that year, as we covered in Anthropic’s 6,000 $100K+ customers.
Baker’s comparison with “50%ish at most application software companies” is his own estimate; he did not name the companies. Billing address shows where customers are invoiced, not where the work is done.
The Structural Read
As we read the data, the payroll side of the output-per-job ratio is mostly an information-sector story: its loss of about 325,000 jobs is larger than the three sectors’ combined net fall, by our arithmetic.
As we read Anthropic’s S-1, a revenue base that was 72.2% American in the latest half is consistent with Baker’s point that any effect would show up in the US first, though billing address is not a measure of where work is done.
Ben Pouladian, on X
“Does this prove AI caused the acceleration? No.”
Three Implications
FOR READERS OF THE PRODUCTIVITY CHARTS Pouladian’s chart says output per payroll job is “Not hours-adjusted productivity”, and output includes non-payroll activity.
FOR THE AI REVENUE PICTURE Anthropic’s S-1 puts US revenue at $11,763 million of $16,286 million in the six months to 30 June 2026.
FOR THE JOBS PICTURE BLS counts show professional, scientific and technical services adding about 163,000 payroll jobs since the fourth quarter of 2022, by our arithmetic, while information lost about 325,000.
The framework separates two curves: “The first is what your underlying technology can do. The second is what your users can actually do with it.”
As we read it, sector output data measure the second curve, which is why a jump in model capability does not have to show up evenly across industries or countries. The framework is the owner’s argument and does not discuss these payroll or output figures.
None of the sources shows that AI caused the higher output per job; both Pouladian’s chart and the FT chart’s title raise that question. Output per payroll job is not output per hour, and our earlier piece on output per hour and adjusted TFP covers a measure on which economist Lars Christensen reads the recent gains as the economy running hot. The payroll data do not say why information jobs fell, and BLS figures are revised.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Baker’s 72% figure for Anthropic’s US revenue checks out against the S-1, and the 1.0% payroll fall on Pouladian’s chart checks out against BLS data. Inside that 1.0%, information lost 10.5% of its jobs while professional and technical services grew, by our arithmetic.
A note on sourcing. Quotes are from Gavin Baker’s and Ben Pouladian’s posts on X and the charts they shared. Output and trend figures are BEP Research’s; payroll counts are from BLS Current Employment Statistics, with the latest months preliminary; revenue by region is from Anthropic’s S-1. Quarterly averages, job changes and shares are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
