Dallas Fed: Postings for AI-Automatable Jobs Down 7-8% in Texas

Two-thirds of Texas firms told the Dallas Fed in May 2026 that they now use AI at work. Two years earlier it was 40 percent. The question a lot of Texas workers have been asking since then is blunt: is that adoption showing up in who gets hired? A new Dallas Fed analysis says yes, and it puts a number on it: postings for the occupations generative AI can automate have fallen by roughly 7 to 8 percent relative to less-exposed jobs, and the firms cutting back are not new AI-native startups. They are the same employers that were hiring before ChatGPT.

What the study actually measured

[Fact] The analysis, by Dallas Fed economists Samuel Dodini and Tucker Smith, was published on 1 September 2026. It links quarterly online job postings from Lightcast, which collects ads from more than 220,000 job boards and company sites, to an occupation-level "automation exposure" score. That score comes from Anthropic's task-based index, which maps O*NET tasks for roughly 1,000 occupations to the tasks people actually perform with Claude. The authors interpret it as the share of an occupation's tasks that generative AI can automate.

This matters for how you read the results. The exposure measure is not a forecast of what AI could theoretically do. It is a record of what one AI assistant is already being used for, at a price firms and consumers are willing to pay. That is the strength of the measure and also its main weakness, which I come back to below.

[Fact] The authors give one worked example: for a medical records technician, about 5 percent of the occupation's tasks are automatable under this index. The most exposed occupations are in software development, web design and other computer-heavy work, followed by managers, clerical workers and editors. If you are in one of those groups, the occupation pages for medical records specialists, software developers, web developers and editors show how the task mix breaks down on this site.

The headline: postings for exposed jobs fell, and kept falling

[Fact] The core estimate compares posting counts for more-exposed versus less-exposed occupations within the same industry, so a general hiring slowdown in, say, tech does not get counted as an AI effect. Scaled to a 10-percentage-point difference in automatable task share, postings for the more-exposed occupations were down about 5 percent by the end of 2023 and, according to the article text, about 8 percent by the first quarter of 2025.

I downloaded the chart data the Dallas Fed published alongside the article and re-read the series myself. [Estimate] The quarterly coefficients are −5.7 in Q4 2023, −6.9 in Q1 2025, −8.1 in Q3 2025, a low of −8.4 in Q4 2025, −8.0 in Q1 2026 and −7.1 in the latest quarter, Q2 2026. The "8 percent by Q1 2025" phrasing in the text is therefore a touch ahead of the data file, where the series first crosses −8 in Q3 2025. That is a small discrepancy between prose and spreadsheet rather than a different story, but it is worth knowing that the effect took until the second half of 2025 to reach the level the summary describes.

[Estimate] The uncertainty is also wider than the headline suggests. The standard error on the Q4 2023 estimate was 0.96 points; on the Q2 2026 estimate it is 1.82 points, almost double. The 95 percent interval for the latest quarter runs from roughly −10.7 to −3.5. The direction is not in doubt. The precise size is.

It is incumbents, not startups, doing the cutting

The most useful part of the study is the firm-level analysis. One tidy story about AI and hiring is that new "AI-native" companies simply start with fewer people, so the aggregate shift comes from firm turnover rather than from existing employers changing behaviour. The Dallas Fed tested that directly.

[Fact] Using a balanced panel of Texas firms present in the postings data across the whole period, with exposure defined by the occupations each firm was posting before ChatGPT, the authors find that more-exposed incumbent firms cut their postings by roughly 5 to 6 percent by mid-2024 and by 8 to 9 percent by early 2026. Those numbers are statistically indistinguishable from the occupation-level results. In other words, the surviving, established employers account for essentially all of the decline. The authors explicitly note that AI-native firm cohorts, which a JPMorgan Chase study found were more likely to adopt AI from the start, do not appear disproportionately responsible.

[Fact] Incumbents also changed what they asked for. Firms whose pre-ChatGPT job mix was 10 percentage points more automatable posted jobs with about 2 percentage points fewer automatable tasks afterwards, which the authors describe as nearly a 50 percent reduction relative to the mean. [Estimate] Reading the chart data, that composition shift is −2.4 percentage points in Q2 2026 and it has moved in one direction almost every quarter since Q4 2022. The 50 percent figure implies the typical firm's automatable task share in the sample was only around 4 to 5 percentage points to begin with, so a 2-point change is large in relative terms.

How big is this for Texas as a whole?

[Fact] Combining the exposure scores, AI usage rates and Texas's industry mix, the authors estimate that generative-AI exposure reduced total Lightcast postings in Texas by about 1.8 percent in 2024 and 2.6 percent in 2025.

Two or three percent of postings is not a collapse. The authors say as much: the aggregate effect "thus far has been modest." But it is concentrated, and that concentration is the point. [Fact] In the Lightcast data, fewer than half of typical job ads require more than two years of experience and very few require more than five. [Estimate] From the published distribution, ads requiring two years or less make up about 52 percent of the total, and ads requiring more than five years only about 9 percent. A hiring pullback in online postings therefore lands mostly on people trying to get in, not people already inside.

That is the bridge to the graduate story. [Claim] The authors argue that the unusually high unemployment rate among recent college graduates over this period is where the employment and earnings effects of generative AI are likely to appear first, and they report that Texas administrative data point the same way, including some current students changing what they study. The article does not publish those administrative results in detail, so I treat that part as the authors' characterisation rather than a number I can check. This site covered the Dallas Fed's earlier work on the entry-level "bottom rung" in a separate post, and its executive survey on staffing plans in another.

What this study cannot tell you

Three limits, stated plainly.

First, the exposure score is built from Claude usage. It measures what one vendor's product is used for, in the population that uses it, and it will move as that product and its users change. An occupation with a low score today is not necessarily safe; it may just be one that Claude users have not reached yet, or one where a different tool dominates.

Second, Lightcast postings are a measure of online hiring demand, not of employment. The authors themselves flag that farming, construction, building maintenance and personal service jobs are underrepresented. A posting decline is not a layoff, and nothing in this study says exposed workers lost jobs they already had.

Third, the pre-period is not flat. [Estimate] In the chart data, the same exposed occupations were posting about 2 percent less than their peers through 2020 and early 2021, before recovering to the Q3 2022 baseline. The post-ChatGPT decline is far larger and far more persistent than that pandemic-era dip, but a reader who wants to argue that exposed occupations are simply more cyclical has a foothold there, and the article does not address it.

There is also a counter-reading of the composition result. Firms posting fewer automatable tasks could mean AI is doing those tasks. It could also mean firms are rewriting job ads to sound less routine, because that is what attracts applicants in 2026. The transcripts of what those workers actually do after hiring are not in this dataset.

What to do with this if you are in an exposed occupation

The pattern in this study is not "AI replaces you." It is "the employer that would have posted your entry-level role posts it less often, and describes it differently." That has a practical implication: the cost of the transition is falling on the first job, not the fifteenth. If you are already employed in an exposed occupation, the evidence here is about the door narrowing behind you, not the floor giving way. If you are trying to enter one, the postings that remain are asking for a task mix with less of the routine work, which is worth reading as a description of the job you are actually being hired to do. The Dallas Fed's own related analysis of good jobs that do not require a degree is covered here.

Sources

• Dodini, S. and Smith, T. (2026), "Job postings show early signs of AI automation impact," Federal Reserve Bank of Dallas, Dallas Fed Economics, 1 September 2026. https://www.dallasfed.org/research/economics/2026/0901
• Chart data (xlsx) published with the article: https://www.dallasfed.org/-/media/documents/research/economics/2026/0901data.xlsx
• Underlying data sources named by the authors: Lightcast job postings; Anthropic's task-based AI exposure index; Texas Business Outlook Survey, May 2026.

Quotations from the Dallas Fed article are brief and used for commentary. The views in the article are those of its authors and not necessarily of the Federal Reserve Bank of Dallas or the Federal Reserve System.

AI-assisted analysis: this post was drafted with AI assistance and reviewed by a human editor before publication. The re-reading of the chart data and the derived figures marked [Estimate] are this site's own calculations, not the Dallas Fed's.