Is AI Showing Up in Productivity and Jobs Data? A Cross-Check of the New Evidence

AI may be raising productivity in a small part of the US economy, but the data do not prove it yet. Output per worker in information, finance, and professional services grew much faster after 2022, and our recomputation from government data confirms most of that jump. Most of the gain came from flat headcount, not faster output, and the timing fits the 2022 interest-rate shock as well as it fits AI. The clearest effect on jobs is in tech and in entry-level hiring, not across the economy.

The Claim and the Counter-Claim

The claim comes from Eric Basmajian of EPB Research, whom the Financial Times cited on October 31, 2026. He grouped the three sectors with the most AI use: information, professional services, and finance. In his numbers, their labor productivity trend rose from 2.2% to 5.0% a year after 2022. The rest of the economy moved from 0.9% to 1.4%. The three sectors hold about 15% of jobs.

In the same FT column, John Burn-Murdoch tested the claim and found it weaker than it looks. His main points, in summary:

1. A Goldman Sachs note that week linked higher AI use by sector with steeper falls in employment. The FT found that the link depends on one data source and one time window. It goes away with a richer employment dataset, and it was stronger two years earlier.
2. The link between AI use, productivity, and falling employment mostly does not appear in other high-income countries.
3. Where it appears, it is mostly in the tech sector, not in finance or professional services.
4. The change starts in 2022, when AI meant chatbots, before agents existed.
5. The gain comes from flat or falling headcount, not faster growth in output.
6. Rate rises and uncertainty from 2022 cut hiring, and Big Tech made large lay-offs before production-ready AI.

We did not see the Goldman note, so we do not report its numbers. We tested the other points against primary data.

We Recomputed the Productivity Jump

Labor productivity here is real output per worker. Real output is a sector's value added after inflation, from the Bureau of Economic Analysis GDP by industry data (chained 2017 dollars, quarterly, through Q2 2026). Workers are payroll employment from the BLS Current Employment Statistics, averaged per quarter. We added the three sectors together and divided output by jobs. "Rest of private economy" is all private industries minus the three sectors.

The trend is the compound annual growth rate between two quarters. The formula is (end value / start value) to the power of (1 / years), minus 1. It leaves out hours, because BLS does not publish hours for finance and insurance. For the other two sectors, output per hour gives almost the same result.

The AI-sector number reproduces. With Q4 2015 to Q4 2019 as the "before" window and Q1 2022 to Q2 2026 as "after", we get 2.1% and 4.9%. Basmajian reports 2.2% and 5.0%. The three sectors are 15.0% of private payroll jobs in Q2 2026, which also matches.

The rest-of-economy number does not reproduce. We get 0.9% before and 0.4% after, not 1.4%. Part of the gap is method. Chained dollars do not add exactly, so subtracting sectors from the total is an approximation. His exact windows and labor measure are not public. The official BLS nonfarm business productivity series grew 1.8% a year before and 2.0% after, so the economy-wide speed-up is small either way.

Is the 5.0% number real? It depends on the window. The table shows how much.

Every window shows an acceleration, from about 1 point to about 4 points. The sectors also differ a lot. Information grew 10.9% a year after 2022 and drives most of the total. Professional services barely changed, from 2.4% to 2.9%. Finance moved from -0.6% to 3.3%.

The finance number needs care. National accounts measure part of bank output from the margin between loan and deposit rates, as Statistics Finland explains for its own accounts. When rates rise, measured bank output can rise without any change in work. That is the deflator problem that readers of the FT column raised.

Output or Headcount: Where the Gain Came From

Output per worker grows when output grows faster than jobs. So a productivity gain can come from more output, from fewer jobs, or from both.

A worked example shows the split for the three AI-heavy sectors:

1. Before 2020, real output grew 4.0% a year and employment grew 1.9% a year. Output per worker grew about 4.0 - 1.9 = 2.1%.
2. After 2022, real output grew 5.3% a year and employment grew 0.4% a year. Output per worker grew about 5.3 - 0.4 = 4.9%.
3. The total change is 4.9 - 2.1 = 2.8 points.
4. Faster output gives 5.3 - 4.0 = 1.3 points of it.
5. Slower hiring gives 1.9 - 0.4 = 1.5 points of it.

So the FT's point holds in part. About half of the gain is flat headcount. In information, it is most of the gain: output growth rose from 7.4% to 8.9% a year, while employment went from +1.1% to -1.8% a year. Payroll jobs in information fell from 3.06 million in Q2 2022 to 2.78 million in Q2 2026, a 9.1% drop. Over the same period, jobs in the rest of the private economy grew 5.1%.

Professional services is the opposite case. Its output growth slowed from 4.9% to 4.0% a year. Its productivity rose only because hiring slowed more. That is not the signature of a new technology that makes each worker produce more.

Big Tech: Revenue and Headcount Split in 2022

The FT showed that Big Tech revenue and headcount moved together until about 2022, then split. We checked the 10-K filings of Meta, Alphabet, and Microsoft.

From 2019 to 2022, combined revenue grew 67% and combined headcount grew 62%. From 2022 to 2025, revenue grew 48% and headcount did not change: 497,716 people at the end of 2022 and 497,685 at the end of 2025. Revenue per employee rose from $1.20 million to $1.78 million.

Two facts weaken a pure AI reading. First, Meta's headcount grew 92% from 2019 to 2022 while its revenue grew 65%, so part of the later flat line is a correction. Meta's 2022 10-K recorded 86,482 employees, and its headcount fell 22% in 2023. Second, revenue per employee rises whenever revenue grows faster than fixed costs. Software companies had that effect long before AI.

Who Uses AI: Two Adoption Measures Agree on the Ranking

If AI drives the productivity gains, the sectors with the most AI use should show the most change. Two very different datasets agree on which sectors use AI most.

The Census Business Trends and Outlook Survey (BTOS) asks firms every two weeks whether they used AI in any business function. In the period from September 21 to October 4, 2026, 24.6% of US firms said yes. Information was highest at 46.4%, then professional, scientific, and technical services at 43.8%, then finance and insurance at 40.5%. The Ramp AI Index counts firms with a paid AI transaction among more than 85,000 Ramp customers. It reported 55.9% for September 2026, with information at 79.9% and finance and insurance at 73.7%.

The ranking agrees. The level does not. Ramp's rate is 1.5x the Census rate in health care and 3.1x in accommodation and food. Three things explain most of the gap:

• The question. A St. Louis Fed analysis found that the wording of the question changes measured adoption a lot. When Census widened its question in November 2025 from "producing goods or services" to "any business function", the measured rate almost doubled.
• The sample. Ramp customers are firms that buy software through a corporate card and bill-pay platform. They are likely to be more digital than the average firm.
• The threshold. One paid subscription counts as adoption at Ramp. Ramp reports a median spend of $15.70 per employee per month in September 2026, while the top 1% of firms spent $9,567.53. Many adopters spend very little.

Usage data from the AI labs point to the same sectors. In the Anthropic Economic Index, computer and mathematical tasks were 35% of Claude.ai conversations in February 2026. Those occupations are about 3.4% of US workers, per Anthropic's first report. A Microsoft study of 200,000 Copilot conversations found that information work is the most common use. OpenAI's study of ChatGPT found that more than 70% of messages are not about work, and that work use is highest in well-paid professional jobs.

So adoption is concentrated where the productivity numbers moved. That supports AI as a cause. It does not settle the timing.

The Timing Problem: 2022 Came Before Most AI Use

The productivity break and the hiring slowdown both start in 2022. ChatGPT launched on November 30, 2022. In September 2023, only 3.7% of US firms used AI to produce goods or services, per a Census working paper on BTOS. The Fed's rate rises, by contrast, line up well. The federal funds rate was 0.08% in February 2022, 4.10% in December 2022, and 5.12% in July 2023.

Indeed software development postings peaked in February 2022 at 2.3x their pre-COVID level. By July 2023 they had fallen 65%. Since September 2023, the index has stayed between 61 and 81. Almost all of the fall happened while firms were still testing chatbots. All postings fell 19% over the same period.

This does not rule out AI, because firms can cut hiring in expectation of future AI. But the rate shock and the end of the COVID hiring boom explain the timing better. AI may explain why software hiring has not come back.

Entry-Level Jobs: Where the Studies Agree and Disagree

The strongest job signal is for young workers. The Stanford Digital Economy Lab uses ADP payroll data. It finds that employment of workers aged 22 to 25 in the most AI-exposed occupations was about 19% below its expected level in June 2026, up from 15% a year earlier. The change comes mostly from fewer hires, not more separations. It is concentrated in jobs where AI tends to automate tasks, not where it assists workers.

The studies disagree on two points:

• Outside tech. The Stanford authors report that the gap remains when they exclude tech firms and computer occupations, and when they control for exposure to rate rises. The FT reports that the entry-level gap in software appears in six countries, and that most of it closes outside tech. The two use different data and different groups of jobs, so both can be partly right.
• Economy-wide effect. The Yale Budget Lab and Brookings compared how fast the mix of occupations has changed since ChatGPT with the arrival of computers and the internet. They found no discernible disruption. The Stanford authors agree that they see no widespread displacement.

Ramp reports that firms that adopt AI grow employment by 10%, including entry-level hiring. Adopters may simply be growing firms, so this does not prove cause either.

The Cross-Check: Which Explanation Does Each Dataset Support?

Four explanations compete. AI raises output per worker. The 2022 rate shock cut hiring. Firms corrected COVID overhiring. Measurement effects, such as price indexes and interest margins, move measured output.

The international row is the FT's result. We did not reproduce it, because current OECD industry data were not available to us. The St. Louis Fed also warns that adoption rates across countries are hard to compare, because the surveys ask different questions.

No single explanation fits all the evidence. The rate shock and the overhiring correction explain the timing. AI fits the sector pattern, the Big Tech split after 2022, and the entry-level data. The most likely answer is that the 2022 shock started the change, and that AI has made firms slower to rehire in the jobs it can do.

For Whom Is AI Showing Up?

The evidence is strongest for tech. Software output scales: one more feature can serve millions of users with no extra staff. So when AI makes engineers faster, a company can grow revenue without growing headcount. The Big Tech filings and the information sector data both show that.

Finance and professional services are different, a point the FT also makes. Much of their output is tied to client relationships. If AI makes each audit or legal matter faster, the firm can serve more clients or bill less time. It cannot easily grow output without more client demand. That fits the data: professional services output growth slowed, while its AI adoption rate is one of the highest.

What This Means for a Fortune 2000 Company Deploying Agents

The macro data cannot predict one company's results, but they show which measures mislead. Use these steps:

1. Measure output per team, not headcount. Count work items closed, cycle time, and error rate before and after an agent goes live. Headcount falls for many reasons.
2. Start where output scales. Put agents first on software, document processing, reconciliation, and reporting.
3. In client work, plan for capacity first. If the work is tied to relationships, use the time saved for more clients or deeper service. Do not expect fewer staff in the first year.
4. Protect the entry-level pipeline. If agents do junior tasks, decide how new staff will learn. Our post on why replacing junior jobs with AI is a bad business strategy covers the trade-off.
5. Check vendor and survey numbers against your own. Good datasets differ by 2x on adoption. Use your own logs and spend data. Our guide to measuring AI coding assistant productivity shows a method for engineering teams.

Turn AI Use Into Output You Can Measure

The national data show that AI gains appear first where companies change the work and measure the output. OpenNash puts forward-deployed engineers on your team and works with Fortune 2000 companies to do that.

1. Pick the work where more output does not need more client time, and set the baseline measures.
2. Get the data ready and connect the agent to the systems where the work happens.
3. Build evals first, so that each agent release is tested for accuracy before it reaches production.
4. Ship to production with cost controls and review steps for client-facing work.
5. Train the team to run, measure, and improve the agents.

At the end, you own the agents, the evals, the dashboards, and the baseline data. Book a call with an OpenNash forward-deployed engineer or email us at hello@opennash.com. In the first session, we find the workflows where an agent can raise output per team the most, and estimate the ROI from your own numbers.