Americans use ChatGPT at work most to make decisions - 1 in 5 requests

Americans who use ChatGPT for work turn to it most often to make decisions and solve problems. That activity accounted for 21.6% of US work-related requests to ChatGPT in the first quarter of 2026, according to data OpenAI published on October 26, 2026, ahead of thinking creatively at 15.6% and writing up or recording information at 11.8%. Each request is a single prompt typed into ChatGPT, which OpenAI's data counts as a "message."

Germany looks much the same, with decision-making at 19.4%. The release, titled "ChatGPT's expanding world of work" and published by OpenAI's Economic Research team, sorts work requests into 165 categories drawn from ONET, the occupational database sponsored by the US Department of Labor. It covers seven quarters, from Q1 2025 through Q3 2026, and nearly 100 countries. Its headline finding is that work use is spreading out: across 74 countries tracked in every quarter, the ten most common activities fell from an average of 56.0% of work requests in Q1 2025 to 51.1% in Q3 2026. In both the US and Germany, requests for advice and for creative work took a bigger share over the year to Q1 2026, while a category ONET uses for programming and software work shrank. "Interacting with computers" dropped from 11.7% to 6.7% of US work requests, and from 12.7% to 8.4% of German ones.

In Short

OpenAI published data on what people ask ChatGPT to help with at work, and in the US the biggest use, about one in five requests, is making decisions and solving problems. Over the past year, more requests in the US and Germany were about getting advice and coming up with ideas, and fewer were about computer and software tasks. The figures show what people ask about, not how much work ChatGPT does, and they leave out company accounts and OpenAI's coding product, Codex.

What OpenAI published

The page is dated October 26 and filed under Research, with the tags ChatGPT and Economic Research. Its subtitle reads: "OpenAI Economic Research releases data for researchers to explore questions on AI and labor more deeply."

The opening paragraph sets out the scope. "Today, we are releasing OpenAI Signals data that explores 165 work activities across nearly 100 countries, revealing how consumers use ChatGPT for work," according to OpenAI. The examples it gives run "from preparing instructional materials and developing marketing content to programming and analyzing business data."

OpenAI frames the release as raw material for outside study rather than a finished report. "Researchers can use this data to explore how ChatGPT use for work differs across countries, track changes over time, and match this data to other economic indicators," the page states. It closes with an invitation: "We invite researchers to explore the data, connect it with other sources, and investigate what these patterns mean for workers and businesses."

A citation block at the bottom of the page names the dataset "OpenAI Signals v3.0" and lists Aaron Chatterji, Thomas Cunningham, David J. Deming, Zoe Hitzig and Drew Johnston among its authors; the author list is cut off on the page itself. The citation points to OpenAI's Signals data download page, labelled "Data download and methodology." According to that page, all Signals data is licensed under Creative Commons CC BY 4.0, which permits reuse and adaptation with attribution, and the methodology sits in a separate data dictionary. The page also links to a separate Signals data page.

Signals itself is not new. OpenAI describes it as a hub for data, research and analysis from its Economic Research team, and the hub lists ten earlier reports published between May and September 2026, among them "How AI is expanding what people do at work" in July and "Work at the Frontier" in September. A February 2026 document, "OpenAI Signals: Measuring AI adoption, protecting privacy, and empowering decisions," laid out the program's design before any of those reports appeared. The same day as the work release, OpenAI also published a separate research post titled "Working with academic communities to build scientific understanding."

How the classification works

The 165 categories are Intermediate Work Activities, or IWAs. "IWAs are commonly-used categories of tasks for the kinds of work people do across occupations," according to OpenAI. The page says the new data "draws on O*NET, a standardized framework that groups occupation-specific tasks into broader categories of work activities shared across occupations," and argues that a common framework "helps researchers compare ChatGPT use with other types of occupational and employment data."

ONET builds its activity taxonomy in layers. According to ONET's data dictionary for release 31.0, its file linking broad work activities to IWAs contains 332 rows, and every IWA is linked to exactly one broader generalized work activity. By PPC Land's calculation, OpenAI's 165 reported categories amount to just under half of O*NET's 332 IWAs. The rest fall into a bucket the charts label Other IWA.

The interactive chart does not show the 165 categories at first glance. It opens on broad groups - Planning and decisions, Information input, Developing and advising, Work output, Information processing, Communicating and interacting, and Administration - and, inside them, tiles carrying the names of ONET's generalized work activities, such as "Making decisions and solving problems" and "Getting information." Selecting a group drills down to the specific activities within it. One label differs from ONET's current wording: the category OpenAI's chart calls "Interacting with computers" appears in ONET as "Working with Computers," which ONET defines as using computers and computer systems, hardware and software, to program and write software, among other functions.

O*NET's own definitions help explain what the other labels capture. Making decisions and solving problems covers analyzing information and evaluating results to choose the best solution. Thinking creatively covers developing, designing or creating new applications, ideas, relationships, systems or products. Providing consultation and advice to others means giving guidance and expert advice to management or other groups on technical, systems- or process-related topics. Interpreting the meaning of information for others is defined as translating or explaining what information means and how it can be used.

How does a request end up in one of these bins? The page gives only an outline. "To protect users' privacy, researchers do not read individual conversations," according to OpenAI. "Automated systems classify messages after identifying information is removed. We publish aggregated statistics, withhold results for small groups, and add statistical noise to protect individual messages." The page does not say which large language model performs the classification, how a request is judged to be work-related, how many requests underlie each quarter, or how large the added noise is.

OpenAI's February 2026 Signals document went further on privacy. According to that report, the text of requests is never released, differential privacy adds calibrated noise to published statistics, and users who declare an age under 18 or who have opted out of training are excluded. The October page does not say whether those exclusions apply to version 3.0.

Germany and the United States, side by side

OpenAI's text does not single out any country. The German and American figures below come from the page's interactive chart, which shows the share of work requests to ChatGPT in each activity for a selected country and quarter.

The chart prints figures only on tiles large enough to hold them. In the US view for Q1 2025, two more activities carry numbers: "Training and teaching others" at 4.3% and "Analyzing data or information" at 4.2%. Neither is labelled in the other three views, so no year-on-year comparison is possible for them.

Eight activities can be compared across both countries and both quarters. All eight moved the same way in Germany as in the US. Five fell, three rose. The magnitudes differed, and by PPC Land's calculation the gaps are large enough to change the ranking in both markets.

Computer work loses ground

The biggest decline in both countries came in "Interacting with computers." The US share fell by 5.0 percentage points, from 11.7% to 6.7%, and the activity slipped from third to sixth among the labelled tiles. Germany's fell by 4.3 points, from 12.7% to 8.4%, dropping from second place to seventh. In Q1 2025 the American and German shares were one point apart; a year later the gap was 1.7 points.

OpenAI's page offers no explanation for the decline. It does define the boundaries of the data: "The data covers work-related messages on individual ChatGPT plans and does not include enterprise messages or Codex use." Codex is OpenAI's agentic AI product for software development. Any programming work that moved from the ChatGPT chat window to Codex during the period would therefore leave these figures without showing up anywhere else in them. OpenAI said in February that Codex had 1.6 million weekly active users, more than triple the number at the start of 2026. Anthropic described a comparable shift for its own products in March, when its Economic Index found that coding work was moving from Claude.ai toward the API, where Claude Code splits tasks into many smaller calls. Nothing on OpenAI's page connects the ChatGPT decline to Codex, and the dataset cannot establish it.

Advice and creative work gain

The largest gains sat in two categories. "Provide consultation and advice to others" rose from 5.5% to 9.0% in the US, a gain of 3.5 points, and from 7.1% to 11.7% in Germany, a gain of 4.6 points. In Germany that lifted the category from seventh to third among labelled activities, just ahead of "Documenting/recording information" at 11.4%.

"Thinking creatively" moved further in the US. It rose 3.8 points to 15.6%, consolidating its place as the second-largest labelled activity behind decision-making. Germany's share climbed by a smaller 1.8 points, to 13.1%. "Documenting/recording information" rose by just over a point in both markets.

"Making decisions and solving problems" remained the largest single tile in all four views, yet it lost share in both countries: 1.2 points in the US and 2.6 points in Germany. "Getting information," which O*NET defines as observing, receiving and otherwise obtaining information from relevant sources, fell by 1.7 points in the US and 1.6 points in Germany.

Germany's interpretation gap

The sharpest difference between the two countries sits in "Interpreting the meaning of information for others." In Q1 2025 it accounted for 9.5% of German work requests against 5.1% in the US. By Q1 2026 the shares were 8.6% and 4.0%, so the German figure was more than double the American one.

O*NET's definition of the category begins with translating. The page does not break the category down further, and it does not attribute the gap to language. It does offer a general caution about cross-country differences: "These differences might reflect variation in the kind of work done across countries or varying access to AI tools." Earlier research points in a similar direction without settling the question. Microsoft's 2025 study of Copilot conversations found that interpreters and translators had the highest overlap with frequent AI use, at 98% of their work activities.

Less concentrated within countries, slightly further apart between them

Two line charts carry the page's headline claim. The first, titled "Within countries," tracks the share of each country's work requests taken by its ten most common tasks. It starts at 56.0% in Q1 2025 and ends at 51.1% in Q3 2026, a drop of 4.9 points. Read from the chart, the decline ran through most quarters, with one small reversal: the share edged up from about 52.0% in Q1 2026 to about 52.4% in Q2 2026 before falling to 51.1%.

The second, "Between countries," measures how different national task mixes are from one another using average total variation distance. In plain terms, total variation distance is the share of one distribution that would have to be moved to make it identical to another; a figure of 23% means that, on average, roughly a quarter of one country's work requests would need reassigning to match another's mix. The page does not say whether the average is taken across pairs of countries or against a global mix. The measure rose from 22.2% in Q1 2025 to 23.0% in Q3 2026. Read from the chart, it dipped to about 21.8% in Q2 2025, reached roughly 23% in Q3 2025, then held near 22.3% to 22.4% for three quarters before returning to 23.0%.

OpenAI interprets the pair as follows: "Connecting this data to IWAs revealed that work-related use became less concentrated in the most common activities. That means that countries tended to specialize in their work over time, and different countries specialized in different types of work." The within-country chart, though, shows work spreading across more activities, not narrowing. The specialization claim rests on the between-country measure, which moved by 0.8 points over seven quarters and stood at roughly the same level in Q3 2025 as in Q3 2026. The page presents the question as open: "Exploring these questions is one example of many that researchers can pursue with this new data."

The concentration finding has a counterpart at Anthropic. Its March 2026 Economic Index found that the top 10 O*NET tasks on Claude.ai fell from 24% of conversations in November 2025 to 19% in February 2026. The two companies count different units - O*NET tasks in Anthropic's case, IWAs in OpenAI's - so the levels are not comparable. The direction is the same.

ChatGPT against the US economy

A second pair of charts compares US work requests to ChatGPT with the distribution of work across the US economy. On the left, "Work with ChatGPT" shows how US work requests split across activity groups in Q1 2025. On the right, "Work across the US economy" shows the share of economy-wide activity at a "fixed 2025-01-01 reference," meaning the economy side does not change as the ChatGPT side moves through the quarters.

The two look very different. Planning and decisions accounts for 35.8% of US work requests to ChatGPT against 9.3% of economy-wide activity, nearly four times as much. Work output holds 22.3% of ChatGPT work requests against 13.9% of the economy. Communicating and interacting runs the other way, at 7.0% of ChatGPT work against 12.8% across the economy. Information input (13.7%), Developing and advising (10.0%) and Information processing (9.0%) complete the ChatGPT side.

The largest single block on the economy side is Other IWA, at 43.1%. "The two distributions differ substantially, largely because about 43% of the employment-weighted task mix for mapped U.S. occupations falls into Other IWAs, outside the individually reported categories," according to OpenAI. "These include activities such as moving materials, operating vehicles, and maintaining health or medical records. These omitted categories did not meet our threshold for consistent representation in the original U.S. ChatGPT sample." The page does not state what that threshold is, or which employment data supplies the weights.

The comparison is shown for the United States only. No equivalent economy-wide view exists on the page for Germany or any other country.

What the dataset does not cover

The exclusions shape every figure above. The data covers individual ChatGPT plans. It leaves out requests sent from company Enterprise accounts and from Codex. OpenAI's page does not list which plans count as individual. Its February document had already noted that the initial release reflected only consumer usage.

Some company work runs through personal accounts. Akamai's October report, drawing on data from the browser security company LayerX, put the share of enterprise AI conversations running through personal identities rather than corporate-managed accounts at 47.11%, a practice it calls shadow AI. Work done that way would land in OpenAI's individual-plan data. Work done on a company's ChatGPT Enterprise account would not.

The figures are shares, not volumes. The page reports what percentage of work requests falls into each activity. It does not say how many requests were counted, how many people sent them or what share of all ChatGPT use is work-related. OpenAI's 2025 usage study, written with researchers from Duke and Harvard, found that non-work requests rose from 53% to more than 70% of consumer use between June 2024 and June 2025. A falling share for one activity can therefore coexist with rising absolute use of it, if total work requests grow fast enough.

Coverage also narrows between the headline and the chart. The release spans "nearly 100 countries," but the concentration figures are calculated across the 74 countries observed in every quarter. And with noise deliberately added to protect individual requests, small shares carry more uncertainty than large ones - a point that matters most for the smaller tiles and the smaller countries.

Why this matters for marketers

The page names "developing marketing content" among the work activities people bring to ChatGPT, one of only four examples it gives. That puts marketing work inside the dataset, though the 165 categories describe activities rather than job titles, so the data cannot isolate marketers from everyone else.

Surveys of marketers fill some of that gap, on a much smaller base. Basis's fourth annual "AI and the Future of Marketing" report, covering 184 agency, brand and publisher professionals, found that ChatGPT use fell to 74.5% of respondents from 88.6% a year earlier, while Claude rose to 72.3% from 17.9%. Research (85.9%) and ideation (83.2%) were the most common uses. Basis measured marketers specifically, while OpenAI's figures cover all work requests, so the two cannot be compared directly.

The release says nothing about advertising. It does, however, describe the same broad population that ChatGPT's ad business is built on. On October 21, five days before the work data appeared, OpenAI extended ChatGPT ads to 26 more countries, bringing the total to 89, with ads shown on the Free and Go tiers. In Europe, where Germany sits, OpenAI told users in August that Free and Go users in the European Economic Area and Switzerland would see ads from later that month. The country-level activity mix in Signals is not an audience-targeting tool, and nothing on the page suggests OpenAI intends it as one.

Regulators have a stake in the numbers too. The European Commission designated ChatGPT a Very Large Online Search Engine on August 31 after OpenAI declared 159.1 million average monthly active users in the EU. That designation carries obligations under the Digital Services Act, among them researcher data access. The Signals release is a voluntary publication and the page does not mention the DSA.

ChatGPT is also one assistant among several. Comscore's US panel data showed ChatGPT's share of AI prompt volume falling from 70% in January 2026 to 50% in June, as Gemini and Claude gained. A shift in the ChatGPT task mix could reflect people moving some kinds of work to other assistants as much as people changing how they work.

Anthropic, Microsoft and Google have each published usage data mapped to occupational categories, and each uses a different method. Anthropic's first Economic Index, in February 2025, found AI used for at least a quarter of tasks in 36% of occupations. Microsoft's study of about 200,000 Copilot conversations scored occupations on AI applicability. Google's ATLAS report in July 2026 drew on 14.65 million interactions and found usage in 68% of detailed occupations, with a median of 21% of their tasks performed with Gemini. OpenAI's release adds a quarterly series across nearly 100 countries under an open license. The page does not compare its results with any of the others.

None of these datasets measures employment. The labor-market question sits elsewhere: Stanford Digital Economy Lab research, using payroll data on more than 25 million US workers, found that employment for workers aged 22 to 25 in the most AI-exposed occupations fell 16% between October 2022 and September 2025. OpenAI's stated aim is for researchers to "match this data to other economic indicators." Whether the shift from computer tasks toward advice and creative work in ChatGPT shows up in hiring, wages or output is a question the release leaves to them.

Timeline

• January 1, 2025 - Fixed reference date for OpenAI's US economy-wide activity distribution; Q1 2025 is the first quarter of the Signals work data
• February 10, 2025 - Anthropic launches its Economic Index, finding AI used for at least a quarter of tasks in 36% of occupations
• July 22, 2025 - Microsoft publishes the final version of its Copilot study based on about 200,000 conversations
• September 15, 2025 - OpenAI's usage study shows non-work requests above 70% of consumer ChatGPT use
• February 2026 - OpenAI publishes its Signals methodology document, describing differential privacy and consumer-only coverage
• February 27, 2026 - OpenAI reports 900 million weekly ChatGPT users and 1.6 million weekly Codex users
• March 24, 2026 - Anthropic's fifth Economic Index finds the top 10 tasks on Claude.ai fell from 24% to 19% of conversations
• July 23, 2026 - Google publishes ATLAS v1.0, finding Gemini usage in 68% of detailed occupations
• July 26, 2026 - A paper in Human Resource Development Review cites Stanford data showing a 16% employment fall for workers aged 22 to 25 in AI-exposed occupations
• August 15, 2026 - OpenAI tells Free and Go users in the EEA and Switzerland that ads are coming
• August 31, 2026 - European Commission designates ChatGPT a Very Large Online Search Engine with 159.1 million EU users
• September 22, 2026 - Comscore reports ChatGPT's US AI prompt share fell from 70% to 50% between January and June
• September 30, 2026 - End of Q3 2026, the last quarter in the Signals work data
• October 19, 2026 - Basis finds ChatGPT used by 74.5% of surveyed marketers, Claude by 72.3%
• October 21, 2026 - ChatGPT ads reach 89 countries; Akamai reports 47.11% of enterprise AI conversations run on personal identities
• October 26, 2026 - OpenAI publishes "ChatGPT's expanding world of work" and OpenAI Signals v3.0, covering 165 work activities in nearly 100 countries from Q1 2025 to Q3 2026

Related PPC Land coverage

• ChatGPT usage reaches 700 million users as personal applications dominate platform growth - OpenAI's 2025 usage study, which first mapped ChatGPT work requests to O*NET activities and found non-work use above 70%.
• Experienced Claude users succeed 10% more - and the gap is widening - Anthropic's March 2026 Economic Index, including the fall in top-10 task concentration and the shift of coding toward the API.
• Artificial intelligence now used in 36% of occupations, study finds - The February 2025 launch of Anthropic's Economic Index and its occupation-level findings.
• Microsoft study shows AI impacts 75% of major occupations in marketing communications - Microsoft's applicability scores from about 200,000 Copilot conversations, with interpreters and translators at the top.
• Google finds AI touches 68% of jobs but only 21% of their tasks - Google's ATLAS v1.0 report mapping 14.65 million Gemini interactions to occupations and tasks.
• 47% of enterprise AI chats run on personal accounts, Akamai report says - Akamai's repackaged LayerX data on work conducted through personal AI accounts.
• Advertisers in 26 more countries gain ChatGPT ads, bringing the total to 89 - OpenAI's October 21 expansion of ChatGPT ads and its 100,000 cumulative advertisers claim.
• ChatGPT Free and Go users in Europe face ads from later this month - The August 2026 notice extending ChatGPT ads to the EEA and Switzerland on the Free and Go tiers.
• ChatGPT faces EU risk rules after declaring 159.1 million users - The Digital Services Act designation that brings audits, ad repositories and researcher data access.
• ChatGPT loses 20 points of US AI prompt share to Gemini and Claude - Comscore's Q2 2026 panel data on prompt share across seven AI assistants.
• ChatGPT drops to 74.5% of marketers as Claude climbs to 72.3%, Basis finds - A survey of 184 marketing professionals on which AI tools they use and for what.
• OpenAI hits 900M users - OpenAI's February 2026 user, subscriber and Codex figures alongside its $110 billion funding round.
• Paper cites Stanford data: young workers lost 16% of AI-exposed jobs - Employment evidence on early-career workers in AI-exposed occupations.

Summary

Who: OpenAI's Economic Research team, with a dataset citation naming Aaron Chatterji, Thomas Cunningham, David J. Deming, Zoe Hitzig and Drew Johnston among its authors. The data describes people using individual ChatGPT plans for work in nearly 100 countries, including the United States and Germany.

What: OpenAI released OpenAI Signals v3.0, sorting work-related requests to ChatGPT into 165 O*NET Intermediate Work Activities by country and quarter. Making decisions and solving problems is the largest single work use, at 21.6% of US requests and 19.4% of German requests in Q1 2026. Across 74 countries, the ten most common activities fell from 56.0% to 51.1% of work requests. In the US and Germany, "Interacting with computers" lost the most share between Q1 2025 and Q1 2026, while "Provide consultation and advice to others" and "Thinking creatively" gained. The data excludes company Enterprise accounts and Codex.

When: Published October 26, 2026, covering Q1 2025 through Q3 2026.

Where: On OpenAI's website, with downloadable data on the Signals data pages under a CC BY 4.0 license.

Why: OpenAI says the release gives researchers a way to study how AI use at work differs across countries, track it over time and connect it to other economic indicators. For marketers, it is a country-level view of what ChatGPT's individual-plan users do at work, the same population that ChatGPT ads now reach in 89 countries, though the page draws no link to advertising.