2026.10.31LabsWhite paper

How much time AI really saves

It is a fair question, because AI is sold on the time it promises to save. The AI True Cost Index measures how much of that time is saved once failures, checking and rework are counted.

AI True Cost Index

For every hour AI is expected to save, it actually saves about 25 minutes

Every hour AI is expected to save1 hourWhat actually happens to that hour25 min actually saved25 minactually saved20 min lost to failed tasks20 minlost to failed tasks15 min checking and fixing15 minchecking and fixingEvery hour AI is expected to save1 hourWhat actually happens to that hour25 min actually saved25 minactually saved20 min lost to failed tasks20 minlost to failed tasks15 min checking and fixing15 minchecking and fixingEvery hour AI is expected to save1 hourWhat actually happens to that hour25 min actually saved25 minactually saved20 min lost to failed tasks20 minlost to failed tasks15 min checking and fixing15 minchecking and fixing

AI True Cost Index, average of the February 2026 and November 2025 results, rounded to whole minutes. An average conversation is expected to save about 2.9 hours, here scaled to one hour. Expected saving: time a task takes a person alone minus time with AI, from the Anthropic Economic Index; checking rate from Workday.

About this paper

AI is sold on the time it saves. Few public sources measure that time. By the measure built here, AI promises an hour and in reality saves about 25 minutes, less than half of the promise. This paper builds its estimates from the usage data Anthropic and OpenAI publish about their own products.

These records have gaps. The model judges its own success, the times are estimates, and each source covers one vendor. The paper names each gap where it matters and keeps the detail in the notes and the methodology section.

Working from imperfect evidence beats guessing, so the paper uses what exists.

What the data shows

  1. AI promises an hour and in reality saves about 25 minutes, less than half.
    The AI True Cost Index puts the return at 39% to 43% of the time AI promises to save. Failures take 19 to 22 minutes of each promised hour. Checking and rework take 14 to 15.
  2. What people use AI for decides the return.
    Writing returns up to half of its promised time, software about a third. Office work returns the most hours, although it succeeds less often than writing.
  3. When agents do the work, people become the bottleneck.
    At Anthropic, agents worked six hours for every employee hour. About 80% of their interactive session time passed waiting for people.
  4. Use is spreading to older age groups.
    People aged 18 to 24 sent 41% of ChatGPT messages in mid-2024 and 30% in mid-2026. Younger people ask AI for guidance more often, older people for information.
  5. Markets differ in what they use AI for, far less in what comes back.
    The US uses AI most for personal advice and the EU more for software. Each market’s mix moves the index by less than one point.

Of every promised hour, 24 to 26 minutes come back

The AI True Cost Index measures success in hours. Between 39% and 43% of the promised time reaches the person.

Anthropic’s data on Claude judges whether each conversation succeeded. It also estimates how long the task would take a person alone and how long it takes with AI.1 A success counts the same whether it replaced five minutes of work or five hours. The index converts each success into time.

The promised saving assumes every conversation succeeds. The returned saving counts only the conversations that succeed. A further 37% of what remains goes to checking and rework, the share found in a Workday survey.2 The index divides returned hours by promised hours. The methodology section at the end gives the formula and its assumptions.

In February 2026 an average conversation promised 2.91 hours and returned 1.24, an index of 43%. In November 2025 it promised 2.81 hours and returned 1.11, an index of 39%.

Where the hour goes

Of every promised hour, 24 to 26 minutes come back

February 2026February 2026, come back: 26 minutes26 mincome backFebruary 2026, lost to failed tasks: 19 minutes19 minlost to failed tasksFebruary 2026, checking and fixing: 15 minutes15 minchecking and fixingNovember 2025November 2025, come back: 24 minutes24 mincome backNovember 2025, lost to failed tasks: 22 minutes22 minlost to failed tasksNovember 2025, checking and fixing: 14 minutes14 minchecking and fixingFebruary 2026February 2026, come back: 26 minutes26 mincome backFebruary 2026, lost to failed tasks: 19 minutes19 minlost to failed tasksFebruary 2026, checking and fixing: 15 minutes15 minchecking and fixingNovember 2025November 2025, come back: 24 minutes24 mincome backNovember 2025, lost to failed tasks: 22 minutes22 minlost to failed tasksNovember 2025, checking and fixing: 14 minutes14 minchecking and fixingFebruary 2026February 2026, come back: 26 minutes26 mincome backFebruary 2026, lost to failed tasks: 19 minutes19 minlost to failed tasksFebruary 2026, checking and fixing: 15 minutes15 minchecking and fixingNovember 2025November 2025, come back: 24 minutes24 mincome backNovember 2025, lost to failed tasks: 22 minutes22 minlost to failed tasksNovember 2025, checking and fixing: 14 minutes14 minchecking and fixing

AI True Cost Index, computed from the Anthropic Economic Index, releases of November 2025 and February 2026.

The 37% for checking is the least certain number in the index. Set it anywhere between 30% and 45%, and the result still lands between a third and a half of the promised time.3

Failures cost more than checking

Failures take 19 to 22 minutes of every promised hour. Checking and rework take 14 to 15.

Claude judged 70% of conversations successful in February 2026 and 67% in November 2025.1 A failed conversation returns nothing and still costs its time with AI. The model grades its own work, so the true failure rate may lie higher or lower.

Reported checking costs vary widely between studies.4 In one survey, review time exceeded the time AI saved. In another, nearly half of the scientists who saved time lost more than a quarter of it to checking. Against this evidence, Workday’s 37% is a moderate assumption.

Felt time and clock time can diverge. In a 2025 randomised study by the research group METR, experienced programmers took 19% longer with AI, but afterwards believed it had made them 20% faster.5 METR now considers that result likely outdated. Every time figure in the index is a model estimate, and no clock measurement checks it.

When agents do the work, people become the bottleneck

At Anthropic, agents worked six hours for every employee hour. Most of their interactive session time passed waiting for people.

The index describes chat, where a person asks and AI answers. Anthropic’s report on its own employees shows what changes when agents take on more of the work.6 By August 2026 the agents directed by employees worked six hours for every human hour. About 80% of their interactive session time passed idle, waiting for a person.

Reading everything the agents produced would take more than half the working week. Writing instructions for agents, having agents check each other and monitoring them grew from 10% of sessions in autumn 2025 to 22% in spring 2026. The median employee spent 20% more time in meetings than a year earlier.

The agents’ code held up. It was removed or reverted at no different rate than code people wrote themselves. The report covers one company that builds the AI it uses, so how far its findings apply elsewhere remains open.

What people use AI for decides the return

Writing returns up to half of its promised time and software a third. Office work returns the most hours.

Grouping the 1,008 kinds of request in Anthropic’s April and May 2026 data by how people use AI gives five segments. The same five appear in the November and February data.7

  • Personal advice, 32% of conversations. People ask about health, emotions, money, purchases and hobbies, and 85% of it is private.
  • Handing over office work, 25%. Presentations, reports, summaries and analysis, 70% of it for work. People hand over the whole task in 45% of these conversations, more than in any other segment.
  • Writing that goes out, 17%. Business letters, posts, marketing copy and CVs, 76% of it for work. People refine the text together with AI in 52% of conversations, more than anywhere else.
  • Building software, 14%. In about half of these conversations AI writes code and the person feeds the errors back.
  • Learning and study, 12%, of which 81% is study.
What people use AI for

Over half of all conversations are work, spread across three uses, and a third are personal advice

Personal advice: 32% of conversations32%Personal adviceHanding over office work: 25% of conversations25%Handing over office workWriting that goes out: 17% of conversations17%Writing that goes outBuilding software: 14% of conversations14%Building softwareLearning and study: 12% of conversations12%Learning and studyPersonal advice: 32% of conversations32%Personal adviceHanding over office work: 25% of conversations25%Handing over office workWriting that goes out: 17% of conversations17%Writing that goes outBuilding software: 14% of conversations14%Building softwareLearning and study: 12% of conversations12%Learning and studyPersonal advice: 32% of conversations32%Personal adviceHanding over office work: 25% of conversations25%Handing over office workWriting that goes out: 17% of conversations17%Writing that goes outBuilding software: 14% of conversations14%Building softwareLearning and study: 12% of conversations12%Learning and study

Share of conversations by segment; segments are breinz’s grouping of the Anthropic Economic Index, April and May 2026.

Writing returns 41% to 50% of its promised time across the two releases. Software returns 33% to 35%, the least of all.

Return by use

Writing returns the most of its promised time, software the least

Share of promised time backWriting that goes outWriting that goes out: November 2025 41%, February 2026 50%50%41%Learning and studyLearning and study: November 2025 44%, February 2026 46%46%44%Personal advicePersonal advice: November 2025 40%, February 2026 44%44%40%Handing over office workHanding over office work: November 2025 38%, February 2026 41%41%38%Building softwareBuilding software: November 2025 33%, February 2026 35%35%33%Share of promised time backWriting that goes outWriting that goes out: November 2025 41%, February 2026 50%50%41%Learning and studyLearning and study: November 2025 44%, February 2026 46%46%44%Personal advicePersonal advice: November 2025 40%, February 2026 44%44%40%Handing over office workHanding over office work: November 2025 38%, February 2026 41%41%38%Building softwareBuilding software: November 2025 33%, February 2026 35%35%33%Share of promised time backWriting that goes outWriting that goes out: November 2025 41%, February 2026 50%50%41%Learning and studyLearning and study: November 2025 44%, February 2026 46%46%44%Personal advicePersonal advice: November 2025 40%, February 2026 44%44%40%Handing over office workHanding over office work: November 2025 38%, February 2026 41%41%38%Building softwareBuilding software: November 2025 33%, February 2026 35%35%33%
November 2025February 2026

AI True Cost Index by segment; segments are breinz’s grouping of the Anthropic Economic Index, November 2025 and February 2026.

Success and hours rank the segments differently. Writing succeeds most often, and office work comes fourth. A person alone would need about four and a half hours for a typical office task, so office work returns the most hours per conversation before checking. Software comes last on both measures, because it fails most often and its tasks in these releases were the smallest.

Success vs time

Office work succeeds less often than writing but returns the most hours

Success rateHours back per conversationWriting that goes out84%1.8 hLearning and study77%2.2 hPersonal advice73%1.5 hHanding over office work67%2.7 hBuilding software63%0.9 hSuccess rateHours back per conversationWriting that goes out84%1.8 hLearning and study77%2.2 hPersonal advice73%1.5 hHanding over office work67%2.7 hBuilding software63%0.9 hSuccess rateHours back per conversationWriting that goes out84%1.8 hLearning and study77%2.2 hPersonal advice73%1.5 hHanding over office work67%2.7 hBuilding software63%0.9 h

Success as judged by Claude and hours returned before checking, by segment. Anthropic Economic Index, February 2026.

Steering AI may add a further cost where people hand over or iterate on work. If a fifth of conversations in office work and software counted as overhead of the kind seen at Anthropic, the index for all use would fall to 37% in February and 35% in November.8

Older people send a growing share of AI messages

The share of ChatGPT messages from people aged 18 to 24 fell from 41% to 30% in two years. Younger people ask for guidance more often, older people for information.

Anthropic’s data holds no ages. OpenAI publishes ChatGPT data by age group, from a sample of a million consumer messages a month.9 That data holds no success or time measure, so the index cannot be split by age.

Age

Messages from 18 to 24 year olds fell from 41% to 30% in two years

20242024, 18 to 24: 41.0%41.0%2024, 25 to 44: 48.8%48.8%2024, 45 and over: 10.2%10.2%20252025, 18 to 24: 35.5%35.5%2025, 25 to 44: 52.0%52.0%2025, 45 and over: 12.5%12.5%20262026, 18 to 24: 30.3%30.3%2026, 25 to 44: 53.9%53.9%2026, 45 and over: 15.8%15.8%20242024, 18 to 24: 41.0%41.0%2024, 25 to 44: 48.8%48.8%2024, 45 and over: 10.2%10.2%20252025, 18 to 24: 35.5%35.5%2025, 25 to 44: 52.0%52.0%2025, 45 and over: 12.5%12.5%20262026, 18 to 24: 30.3%30.3%2026, 25 to 44: 53.9%53.9%2026, 45 and over: 15.8%15.8%20242024, 18 to 24: 41.0%41.0%2024, 25 to 44: 48.8%48.8%2024, 45 and over: 10.2%10.2%20252025, 18 to 24: 35.5%35.5%2025, 25 to 44: 52.0%52.0%2025, 45 and over: 12.5%12.5%20262026, 18 to 24: 30.3%30.3%2026, 25 to 44: 53.9%53.9%2026, 45 and over: 15.8%15.8%
18 to 2425 to 4445 and over

OpenAI Signals, data files version 3.0. Share of messages by age group, world, averages of July to September 2024, 2025 and 2026.

Age changes what people ask for, though moderately. Practical guidance covers how-to advice, tutoring, ideas and self-care, and it takes the largest share at every age. Seeking information rises with age up to 64, then dips after 65. Writing peaks among 25 to 34 year olds. Technical help, which includes programming, stays at about 4% at every age. Younger people seem to use AI more as a tutor and adviser, older people more as a search engine.

What each age asks

Seeking information rises with age, practical guidance falls

18-2425-3435-4445-5455-6465+Age groupWriting, 18-24: 21.2%Writing, 25-34: 21.9%Writing, 35-44: 21.4%Writing, 45-54: 20.4%Writing, 55-64: 18.6%Writing, 65+: 17.0%17%WritingSelf-expression, 18-24: 10.1%Self-expression, 25-34: 8.7%Self-expression, 35-44: 8.5%Self-expression, 45-54: 8.8%Self-expression, 55-64: 8.8%Self-expression, 65+: 10.9%11%Self-expressionTechnical help, 18-24: 4.4%Technical help, 25-34: 4.2%Technical help, 35-44: 4.4%Technical help, 45-54: 4.4%Technical help, 55-64: 4.1%Technical help, 65+: 3.4%3%Technical helpPractical guidance, 18-24: 31.4%Practical guidance, 25-34: 30.0%Practical guidance, 35-44: 27.6%Practical guidance, 45-54: 26.3%Practical guidance, 55-64: 26.6%Practical guidance, 65+: 27.0%27%Practical guidance31%Seeking information, 18-24: 18.1%Seeking information, 25-34: 19.8%Seeking information, 35-44: 21.2%Seeking information, 45-54: 23.3%Seeking information, 55-64: 25.2%Seeking information, 65+: 22.6%23%Seeking information18%18-2425-3435-4445-5455-6465+Age groupWriting, 18-24: 21.2%Writing, 25-34: 21.9%Writing, 35-44: 21.4%Writing, 45-54: 20.4%Writing, 55-64: 18.6%Writing, 65+: 17.0%17%WritingSelf-expression, 18-24: 10.1%Self-expression, 25-34: 8.7%Self-expression, 35-44: 8.5%Self-expression, 45-54: 8.8%Self-expression, 55-64: 8.8%Self-expression, 65+: 10.9%11%Self-expressionTechnical help, 18-24: 4.4%Technical help, 25-34: 4.2%Technical help, 35-44: 4.4%Technical help, 45-54: 4.4%Technical help, 55-64: 4.1%Technical help, 65+: 3.4%3%Technical helpPractical guidance, 18-24: 31.4%Practical guidance, 25-34: 30.0%Practical guidance, 35-44: 27.6%Practical guidance, 45-54: 26.3%Practical guidance, 55-64: 26.6%Practical guidance, 65+: 27.0%27%Practical guidance31%Seeking information, 18-24: 18.1%Seeking information, 25-34: 19.8%Seeking information, 35-44: 21.2%Seeking information, 45-54: 23.3%Seeking information, 55-64: 25.2%Seeking information, 65+: 22.6%23%Seeking information18%18-2425-3435-4445-5455-6465+Age groupWriting, 18-24: 21.2%Writing, 25-34: 21.9%Writing, 35-44: 21.4%Writing, 45-54: 20.4%Writing, 55-64: 18.6%Writing, 65+: 17.0%17%WritingSelf-expression, 18-24: 10.1%Self-expression, 25-34: 8.7%Self-expression, 35-44: 8.5%Self-expression, 45-54: 8.8%Self-expression, 55-64: 8.8%Self-expression, 65+: 10.9%11%Self-expressionTechnical help, 18-24: 4.4%Technical help, 25-34: 4.2%Technical help, 35-44: 4.4%Technical help, 45-54: 4.4%Technical help, 55-64: 4.1%Technical help, 65+: 3.4%3%Technical helpPractical guidance, 18-24: 31.4%Practical guidance, 25-34: 30.0%Practical guidance, 35-44: 27.6%Practical guidance, 45-54: 26.3%Practical guidance, 55-64: 26.6%Practical guidance, 65+: 27.0%27%Practical guidance31%Seeking information, 18-24: 18.1%Seeking information, 25-34: 19.8%Seeking information, 35-44: 21.2%Seeking information, 45-54: 23.3%Seeking information, 55-64: 25.2%Seeking information, 65+: 22.6%23%Seeking information18%

OpenAI Signals, data files version 3.0. Topic mix of each age group derived from age by topic and topic shares, world, average of July to September 2026; Multimedia and Other not shown.

The two vendors’ data point to different uses, although their categories differ.10 Technical help fell from 13.5% of ChatGPT consumer messages in July 2024 to about 4% in 2026. Building software takes 14% of Claude conversations.

Markets use AI differently and get nearly the same return

The US uses AI most for personal advice and the EU more for software. Each market’s mix moves the index by less than one point.

In April and May 2026 the US gave 40% of its Claude conversations to personal advice, against 34% in the EU and 29% in the rest of the world. Software took 10% in the US and in the UK and Ireland, 14% in the EU and 15% in the rest of the world. Writing that goes out stayed at 17% to 19% everywhere.1

Markets

The US leans on personal advice, the EU and the rest of the world on software

United StatesUnited States, Personal advice: 40%40%Personal adviceUnited States, Office work: 23%23%Office workUnited States, Writing: 18%18%WritingUnited States, Software: 10%10%SoftwareUnited States, Study: 10%10%StudyUK and IrelandUK and Ireland, Personal advice: 35%35%Personal adviceUK and Ireland, Office work: 22%22%Office workUK and Ireland, Writing: 19%19%WritingUK and Ireland, Software: 10%10%SoftwareUK and Ireland, Study: 15%15%StudyEuropean UnionEuropean Union, Personal advice: 34%34%Personal adviceEuropean Union, Office work: 22%22%Office workEuropean Union, Writing: 18%18%WritingEuropean Union, Software: 14%14%SoftwareEuropean Union, Study: 13%13%StudyRest of the worldRest of the world, Personal advice: 29%29%Personal adviceRest of the world, Office work: 26%26%Office workRest of the world, Writing: 17%17%WritingRest of the world, Software: 15%15%SoftwareRest of the world, Study: 13%13%StudyUnited StatesUnited States, Personal advice: 40%40%Personal adviceUnited States, Office work: 23%23%Office workUnited States, Writing: 18%18%WritingUnited States, Software: 10%10%SoftwareUnited States, Study: 10%10%StudyUK and IrelandUK and Ireland, Personal advice: 35%35%Personal adviceUK and Ireland, Office work: 22%22%Office workUK and Ireland, Writing: 19%19%WritingUK and Ireland, Software: 10%10%SoftwareUK and Ireland, Study: 15%15%StudyEuropean UnionEuropean Union, Personal advice: 34%34%Personal adviceEuropean Union, Office work: 22%22%Office workEuropean Union, Writing: 18%18%WritingEuropean Union, Software: 14%14%SoftwareEuropean Union, Study: 13%13%StudyRest of the worldRest of the world, Personal advice: 29%29%Personal adviceRest of the world, Office work: 26%26%Office workRest of the world, Writing: 17%17%WritingRest of the world, Software: 15%15%SoftwareRest of the world, Study: 13%13%StudyUnited StatesUnited States, Personal advice: 40%40%Personal adviceUnited States, Office work: 23%23%Office workUnited States, Writing: 18%18%WritingUnited States, Software: 10%10%SoftwareUnited States, Study: 10%10%StudyUK and IrelandUK and Ireland, Personal advice: 35%35%Personal adviceUK and Ireland, Office work: 22%22%Office workUK and Ireland, Writing: 19%19%WritingUK and Ireland, Software: 10%10%SoftwareUK and Ireland, Study: 15%15%StudyEuropean UnionEuropean Union, Personal advice: 34%34%Personal adviceEuropean Union, Office work: 22%22%Office workEuropean Union, Writing: 18%18%WritingEuropean Union, Software: 14%14%SoftwareEuropean Union, Study: 13%13%StudyRest of the worldRest of the world, Personal advice: 29%29%Personal adviceRest of the world, Office work: 26%26%Office workRest of the world, Writing: 17%17%WritingRest of the world, Software: 15%15%SoftwareRest of the world, Study: 13%13%Study

Segment mix by market; segments are breinz’s grouping of the Anthropic Economic Index, April and May 2026.

Anthropic publishes success and time only for the world, so the market index measures the effect of each market’s mix.11 On the February mix it gives 42.6% for Germany and the world, 42.8% for France and the UK and 42.9% for the US.

The UK and Ireland share a language with the US and lie in Europe, which allows a test.12 In what people ask for, the UK and Ireland look like the US. They request less software, more advice and almost no translation, as Americans do. In how they work and how much they study, they look like the continent. They hand over whole tasks less often than Americans and study more.

Language

The UK and Ireland ask like Americans and work like Europeans

What people ask forSoftware requestsSoftware requests, US: 8.4%8.4%USSoftware requests, UK: 8.5%8.5%UKSoftware requests, Europe: 11.2%11.2%EuropeAdvice as the outputAdvice as the output, US: 14.4%14.4%USAdvice as the output, UK: 13.2%13.2%UKAdvice as the output, Europe: 11.4%11.4%EuropeHow people workWhole task handed overWhole task handed over, US: 31.8%31.8%USWhole task handed over, UK: 28.5%28.5%UKWhole task handed over, Europe: 28.5%28.5%EuropeStudyStudy, US: 10.8%10.8%USStudy, UK: 17.4%17.4%UKStudy, Europe: 15.3%15.3%EuropeWhat people ask forSoftware requestsSoftware requests, US: 8.4%8.4%USSoftware requests, UK: 8.5%8.5%UKSoftware requests, Europe: 11.2%11.2%EuropeAdvice as the outputAdvice as the output, US: 14.4%14.4%USAdvice as the output, UK: 13.2%13.2%UKAdvice as the output, Europe: 11.4%11.4%EuropeHow people workWhole task handed overWhole task handed over, US: 31.8%31.8%USWhole task handed over, UK: 28.5%28.5%UKWhole task handed over, Europe: 28.5%28.5%EuropeStudyStudy, US: 10.8%10.8%USStudy, UK: 17.4%17.4%UKStudy, Europe: 15.3%15.3%EuropeWhat people ask forSoftware requestsSoftware requests, US: 8.4%8.4%USSoftware requests, UK: 8.5%8.5%UKSoftware requests, Europe: 11.2%11.2%EuropeAdvice as the outputAdvice as the output, US: 14.4%14.4%USAdvice as the output, UK: 13.2%13.2%UKAdvice as the output, Europe: 11.4%11.4%EuropeHow people workWhole task handed overWhole task handed over, US: 31.8%31.8%USWhole task handed over, UK: 28.5%28.5%UKWhole task handed over, Europe: 28.5%28.5%EuropeStudyStudy, US: 10.8%10.8%USStudy, UK: 17.4%17.4%UKStudy, Europe: 15.3%15.3%Europe

Four measures from the Anthropic Economic Index, April and May 2026. UK stands for the UK and Ireland; Europe for continental Europe without them.

Language seems to shape what people ask for, and the European setting how they work. The test cannot separate language from culture, because the two travel together.

A view from one country: Poland

Seen from Poland, the same patterns take a local shape. Polish users code more and study less on Claude, write less on ChatGPT, and send fewer messages from people over 55. Their mix of requests barely changes the rate of return.

Poland shows how the patterns above look from inside one market. In all three Claude releases with segments, Polish users gave more of their conversations to software than the world did. Study took 6% to 10% in Poland against about 12% in the world.13

Poland: software

Poland gives software a bigger share than the world in every release

Share of conversations about softwareNovember 2025November 2025: Poland 23%, world 15%23%15%February 2026February 2026: Poland 15%, world 11%15%11%April and May 2026April and May 2026: Poland 17%, world 14%17%14%Share of conversations about softwareNovember 2025November 2025: Poland 23%, world 15%23%15%February 2026February 2026: Poland 15%, world 11%15%11%April and May 2026April and May 2026: Poland 17%, world 14%17%14%Share of conversations about softwareNovember 2025November 2025: Poland 23%, world 15%23%15%February 2026February 2026: Poland 15%, world 11%15%11%April and May 2026April and May 2026: Poland 17%, world 14%17%14%
WorldPoland

Software segment share, Poland and world; breinz’s grouping of the Anthropic Economic Index, November 2025, February 2026, April and May 2026.

The lean toward software matches the Polish economy. Poland exported USD 19.6 billion of ICT services in 2024, about twice what all of Africa exported in 2023. Telecom, computer and information services made up 17% of Polish services exports. The share of IT workers is ordinary, though. ICT specialists were 4.5% of Polish employment in 2025, below the EU average of 5.0%. Poland seems to build software for export more than it employs an unusual number of programmers.14

The Polish mix of requests gives an index of 42.3%, level with the other markets. Software returns the fewest hours per conversation. In February 2026 a Polish conversation returned about 5% fewer hours than the world average.

Language may add a cost the index does not see. A Polish word takes about 1.8 times as many tokens as an English one in common tokenizers, so the same request costs more and fills the model’s memory sooner. The index cannot show this, because its times are the model’s estimates of human work, and the data does not say which language Polish users write in. Quality does not obviously suffer. In a 2025 benchmark of 26 languages on long texts, Polish ranked first and English sixth.15

ChatGPT shows a different side of Polish use.9 Writing took 17.1% of Polish messages in July to September 2026, the lowest of the five markets compared, against about 21% in the US, Germany and the UK. Polish writing was the lowest in 2025 and early 2026 as well. Seeking information took 25.9%, close to France and above the US at 21.0%. Technical help took 4.1%, the same as in the US, Germany and the UK. The Polish lean toward software appears in Claude data only.

Polish use centres on people aged 25 to 44, who sent 61% of Polish ChatGPT messages in 2026, against 54% worldwide. People aged 55 and over sent 5.1%, against 7.7% in Germany, 8.3% in the UK, 9.1% in France and 10.3% in the US. The older share is growing fast, and the Polish age mix of 2026 is close to the US mix of 2025.

Poland: age

Polish messages come mostly from people aged 25 to 44

Share of messages18 to 2418 to 24: Poland 22.8%, world 30.3%22.8%30.3%25 to 4425 to 44: Poland 61.0%, world 53.9%61.0%53.9%45 to 5445 to 54: Poland 11.0%, world 9.3%11.0%9.3%55 and over55 and over: Poland 5.1%, world 6.5%5.1%6.5%Share of messages18 to 2418 to 24: Poland 22.8%, world 30.3%22.8%30.3%25 to 4425 to 44: Poland 61.0%, world 53.9%61.0%53.9%45 to 5445 to 54: Poland 11.0%, world 9.3%11.0%9.3%55 and over55 and over: Poland 5.1%, world 6.5%5.1%6.5%Share of messages18 to 2418 to 24: Poland 22.8%, world 30.3%22.8%30.3%25 to 4425 to 44: Poland 61.0%, world 53.9%61.0%53.9%45 to 5445 to 54: Poland 11.0%, world 9.3%11.0%9.3%55 and over55 and over: Poland 5.1%, world 6.5%5.1%6.5%
WorldPoland, higherPoland, lower

OpenAI Signals, data files version 3.0. Share of messages by age group, Poland and world, average of July to September 2026.

Use keeps spreading. Statistics Poland (GUS) found that 39.8% of people aged 16 to 74 used AI tools in the three months before its 2026 survey, 17.1 percentage points more than a year earlier. The share of firms using AI rose from 8.7% to 17.8%.16

Methodology

The findings rest on a series of decisions breinz took while building the analysis. Each decision chose between options for a stated reason. Together they form the breinz methodology behind the AI True Cost Index.

The question. The paper asks how much time AI really saves. It studies patterns of behaviour, not tools or vendors, so Claude and ChatGPT are compared only where they differ. The numbers are the base for reasoning, and the paper reports what they show about people.

The sources. The Anthropic Economic Index is the only public source that records, for each kind of request, whether AI succeeded and how long the task takes with and without AI. It carries the main analysis. OpenAI’s Signals data adds what Anthropic lacks, the age of users. Outside studies cover what neither vendor measures, the time spent checking output and the work of managing agents. Only data large enough to be reliable is used. Polish regions were dropped for that reason, and countries with fewer than about 1,500 conversations a week enter only in groups.

The unit. The data holds totals, not people, so the unit of analysis is the kind of request, 1,008 of them in the newest release. Segments are a way of reading these requests. Each kind of request is described by how people work with AI, what they use it for, its task and its output. A weighted clustering groups kinds of request with similar profiles. Outputs under 1% of conversations were merged into broader groups or left out, and translation was kept only for the language comparison. Five segments emerged, held across repeated runs and reappeared in the two earlier releases.

Time as the measure. Anthropic judges success conversation by conversation, and a success counts the same whether it saved five minutes or five hours. Automation is worth what it saves in time, so breinz made time the measure and converted success into hours. The promised saving is the time a person alone would need minus the time with AI, as if every conversation succeeded. The returned saving counts only the share that succeeds, and every kind of request counts as often as it occurs.

Index =Σ w · (p · Talone − TAI) · (1 − 0.37)Σ w · (Talone − TAI)

Here w is the share of conversations of each kind of request, p its success rate, and T the two time estimates.

The hidden costs. Three costs sit between the promise and the return, and each was treated according to its evidence. Failures come from the model’s own judgement and enter the index directly. Checking and rework come from a Workday survey that put them at 37% of the time saved. That figure enters the index too, and values from 30% to 45% were tested. The gap between felt time and clock time rests on one study of programmers that its authors now consider outdated, so it stays outside the index. The overhead of managing agents comes from one company, so it appears only as a scenario.

The assumptions. Where the data is silent, the index assumes. A failed conversation lasts as long as a successful one, because the data gives one time with AI per kind of request. After a failure the person does the task alone, and that time counts as ordinary work. Checking takes the same share in every segment.

The markets. The US is treated as a separate market, on the premise that most large language models are trained mainly on American English. The EU, the UK and Ireland, and the rest of the world are compared with it. The UK and Ireland serve as a test, since they share the language of the US and the setting of Europe. Anthropic publishes success and time only for the world, so each market’s index applies world rates to its own mix of requests and measures the effect of that mix alone.

The ages. OpenAI publishes the age mix of each country’s messages and of each topic. Combining the two shows what each age group asks for. Each year is compared in the same season, July to September, and all figures are shares of messages, not of people.

Poland. Poland is read as one country’s view of the general patterns. Each release holds about 9,000 to 10,000 Polish conversations a week, which puts the estimated sampling error of Polish shares near one percentage point.

The releases. The index uses November 2025 and February 2026, the only releases that judge success. The segments come from April and May 2026, the only release with output types. Anthropic changed its method in April 2026, so hours from different releases are never combined.

Limits

Each usage source covers one vendor and counts conversations or messages, so heavy users weigh more than light ones. Success and both times in Anthropic’s data are the model’s estimates. The rework rate comes from one survey and applies equally to every segment.

Anthropic’s newest release lacks a success measure, so the index rests on November 2025 and February 2026. Anthropic changed its method in April 2026, so hours from different releases stay apart. OpenAI’s data covers consumer accounts and adults only, and it holds no success or time measure. Its age figures are shares of messages, so they describe use, not the number of users. The report on agents describes one company.

Within these limits, writing, study and personal advice return more than office work in both releases, and software returns the least. Across the tested assumptions, the overall return stays between about a third and a half of the promised time.

Notes and sources

  1. Anthropic Economic Index. Anthropic samples conversations on Claude.ai. Claude classifies each one, and only totals are published. Each conversation gets a kind of request, a way of working, a purpose, a success judgement and two time estimates. The November 2025 and February 2026 releases each rest on a million conversations and are the only ones that judge success. Anthropic, Anthropic Economic Index, 2026.
  1. Workday surveyed 3,200 employees and found that 37% of the time AI saved went to rework. Workday, New Workday Research: Companies Are Leaving AI Gains on the Table, 2026.
  2. With rework at 30% of saved time, the index is 47% in February 2026 and 44% in November 2025. With rework at 45%, it is 37% and 34%.
  3. End users in a Foxit survey saved 3.6 hours a week and spent 3 hours 50 minutes reviewing AI output. In a study by Google, Google DeepMind and MIT, scientists saved 6.9 hours a week, and 46% of those who saved time lost more than a quarter of it to checking. Accounting Today, Time saved by AI partially canceled out by time spent checking AI, 2026. Research Professional News, AI saves scientists seven hours a week, Google paper finds, 2026.
  4. In February 2026 METR wrote that developers are likely more sped up by AI now than in early 2025. Its newer estimates, a speedup of about 18% for returning developers and about 4% for new ones, both have confidence intervals that include zero, and METR calls them very weak evidence. METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, 2025. METR, We are Changing our Developer Productivity Experiment Design, 2026.
  5. Technical individual contributors spent 16% more time in meetings. Reading all agent output would take more than half the week even for non-technical employees. Anthropic, How employment is changing at Anthropic, 2026.
  6. The segments are breinz’s own grouping. The methodology section describes how they were built.
  7. The overhead scenario is an assumption for illustration. It counts a fifth of conversations in office work and software as time spent steering AI.
  8. OpenAI Signals, data files version 3.0. It covers consumer ChatGPT plans only, without enterprise use or Codex. It samples a million messages a month from July 2024 to September 2026, from accounts with a declared age of 18 or more. Classifiers sort the messages, and noise is added to protect privacy. All figures are shares of messages. OpenAI, Signals consumer data, 2026.
  9. OpenAI’s technical help covers programming, mathematics and data analysis. Anthropic’s software segment is breinz’s own grouping. The two categories overlap without matching, and the two services reach different users.
  10. A market index built this way shows how much a market’s mix of requests moves the return. It cannot show whether AI succeeds more often in one market than in another.
  11. Software requests in the language test count one kind of request, so they take a smaller share than the software segment.
  12. In November 2025 and February 2026 only about 60% of Polish conversations could be placed in a segment, so the direction holds more firmly than the exact levels.
  13. Poland’s ICT services exports were USD 19.6 billion in 2024, by World Bank data. Africa’s were USD 9.2 billion in 2023, by a Tralac report; the years differ, so the comparison is approximate. Telecom, computer and information services made up 17% of Polish services exports in 2024. ICT specialists were 4.5% of Polish employment in 2025, against an EU average of 5.0%. Trading Economics, Poland ICT service exports (World Bank data), 2026. Ecofin Agency, African ICT service exports grow by 5.4% from 2022 to 2023, 2024. Polish Economic Institute, Economic Weekly 40/2025, 2025. European Commission, Poland snapshot: digital skills, 2026.
  14. Across six tokenizers (Llama, Gemma, GPT-4o, DeepSeek, Qwen), Polish averages 2.25 tokens per word against 1.23 for English, a ratio of 1.6 to 2.2 depending on the tokenizer. Claude’s tokenizer was not tested. Volodymyr Ovcharov, The Tokenizer Tax Across 25 European Languages, 2026. Kim, Russell, Karpinska and Iyyer, One ruler to measure them all: Benchmarking multilingual long-context language models, 2025.
  15. GUS, Społeczeństwo informacyjne w Polsce w 2026 r., 2026.