AI-first company

Tool

AI ROI calculator: will AI pay back?

Enter your own numbers. Tick the mistakes you are making and the practices you follow. See what turns the return on AI from a loss into a gain. The tool shows which choices matter most. It does not forecast your return.

By
Last reviewed
Sources
80
Corrections
Report an error

Your numbers

Each field starts with an example figure. It does not come from a study. Replace it with yours. Amounts are in US dollars because the list prices are. If you enter the hourly cost and the licence in your own currency, read the results in it.

What AI is used for

Only kinds of work with a measured figure are listed: writing short professional texts, 40% less time (Noy and Zhang, 2023); customer support, 13% less time (Brynjolfsson, Li and Raymond, 2025); analysis and consulting work, 10.9% less time (Dell’Acqua and others, 2026); writing software, 20.7% less time (Cui and others, 2025). Each was measured on one group of people doing one kind of work.

Example figure Used only when you choose a kind of work.

Company size

Set the number of people for a company of

Size sets the number of people and nothing else. No study shows that company size changes the return.

Example figure Everyone who gets a licence.

Cost of the AI tools

Example figure Anthropic lists Claude Team at $20 a seat a month, billed yearly. Microsoft lists its Copilot add-on for businesses at $18. Both checked 13 October 2026 (Anthropic price list, Microsoft price list).

Example figure For comparison: a two-person French firm reports about 60 hours to set up its AI tools Reported, not measured (France Num case page).

Your people’s time

Example figure Wages plus what the employer pays on top, per hour worked.

Example figure

Example figure After holidays and other leave.

Example figure No study gives this share. It is yours to set.

Example figure Time people spend checking and fixing what AI produced.

Mistakes

Tick the ones that describe your company. Under each one you see which line of the result it changes, and how.

Time freed The saving is only a few per cent of everyone’s hours: what staff in company-wide rollouts said they saved. It replaces a figure measured on one kind of work.

Sources

Humlum and Vestergaard, 2025, Bick, Blandin and Deming (St. Louis Fed), 2025, Dillon and others, 2025, Australian government Copilot trial, 2024.

Time freed You cannot count on any saving. The low estimate of time freed drops to zero, and the “Unverified” warning appears. In studies, what people believed and what the clock showed have pointed in opposite directions.

Sources

Becker and others (METR), 2025, Yu and others, 2026, UK Department for Business and Trade, 2025, Australian Treasury, 2025.

Time freed You may save no time at all. The low estimate of time freed drops to zero. Where this was measured, AI made some work slower or less accurate. No single figure fits all kinds of work.

Sources

Becker and others (METR), 2025, METR update, 2026, Dell’Acqua and others, 2026, UK Department for Business and Trade, 2025.

Time put to use The time saved may never turn into money. The low estimate of time put to use drops to zero. Studies that checked found time saved, with no change in earnings, hours or tasks.

Sources

Humlum and Vestergaard, 2026, Dillon and others, 2025, UK Department for Business and Trade, 2025, UK government Copilot experiment, 2025.

Time put to use Costs come first and gains later, so the result shows the first year apart from later years. Nobody knows how much later the gains come.

Sources

Brynjolfsson, Rock and Syverson, 2018, McElheran and others, 2025, Acemoglu, 2024, Yotzov and others, 2026.

Running cost Nobody has measured what this costs, so you get a warning, not a number. Only reported cases exist.

Sources

Fortune, April 2026, heise online, 2026, Fortune, August 2025 (Coinbase), CNBC, 2025, Fortune, August 2025 (Duolingo), BetterUp and Stanford survey, 2025, Holmström and Milgrom, 1991.

Running cost Rework eats into the time put to use. Your own checking figure is replaced by the only published one: employees told a survey by a software vendor that about 40% of the time AI saved them was lost to rework.

Sources

Workday survey, 2026, UK Department for Business and Trade, 2025.

Risk: what it cost others Lists what that cost other organizations, case by case.

Sources

12 cases, listed under “Risk: what it cost others”.

Practices

Tick the ones you follow. Where a practice is the opposite of a mistake, ticking one unticks the other.

Time freed The saving is the figure a study measured for the kind of work you choose. It applies only to the share of time that work takes. Choose the kind of work.

How to do it: Start with one costly process, not a company-wide rollout.

Sources

Noy and Zhang, 2023, Brynjolfsson, Li and Raymond, 2025, Dell’Acqua and others, 2026, Cui and others, 2025.

Time freed The low estimate of time freed is no longer set to zero for lack of measurement, and the “Not measured in your company” warning goes.

How to do it: Measure time saved per task against a baseline.

Sources

UK Department for Business and Trade, 2025.

Time freed The low estimate of time freed is no longer set to zero for untested work. The “Not tested” and “Possible slowdown” warnings go.

How to do it: Map tasks before deciding what to automate, For each kind of task, test whether a person with AI beats either one alone.

Sources

Dell’Acqua and others, 2026.

Time freed More of your staff use the tool. The low estimate of how many use it rises from 47% to 83%. A Danish survey found that difference where employers encouraged use. The “Usage targets” warning goes.

How to do it: Train staff on the tools you actually chose.

Sources

Humlum and Vestergaard, 2025, Dillon and others, 2025.

Time put to use Your own share of freed time put to use counts in both the low and the high estimate, and the “Not decided” warning goes.

Sources

Australian government Copilot trial, 2024, Humlum and Vestergaard, 2025, Jiang and others, 2025.

Time put to use You get a note and its sources, not a number. Studies show that it helps, but not by how much.

How to do it: Redesign the workflow instead of adding a tool to it.

Sources

McKinsey survey, 2025, Dillon and others, 2025, Babina and others, 2024, Aghion and others, 2025, Czarnitzki, Fernández and Rammer, 2023.

Running cost Your own figures for checking, set-up and training count, and the “Not budgeted” warning goes.

How to do it: Budget for the work AI creates, Budget for people and process, not just tools.

Sources

Humlum and Vestergaard, 2025, France Num case page, Australian Treasury, 2025.

Risk: what it cost others The list of what went wrong at other organizations shrinks to a short note.

How to do it: Keep a person checking anything that goes to a customer or carries legal weight, Claim only what your AI has been tested to do.

Sources

Vaccaro, Almaatouq and Malone, 2024.

How the result is worked out

The result has four lines. The first three each give a low and a high estimate, and together they add up to the result. The fourth lists what went wrong at other organizations. It is never added in.

Time freed

Hours AI could free in the work it is used for.

$0 to $79,500 0 to 1,990 hours a year

Not tested Unverified

How this line is worked out

Not tested. You have not checked which tasks AI does well. Consultants using AI were 19 percentage points less likely to be right on a task the AI could not do well. In a UK government pilot, scheduling and making images took longer with AI. Sources: Dell’Acqua and others, 2026, UK Department for Business and Trade, 2025.

Unverified. You know the saving only from what staff say. In a trial, 16 experienced developers believed AI had made them 20% faster. Timed, they were 19% slower. In experiments on simple tasks, people expected AI to save 55.7 seconds and it saved 7.5. Staff are not always wrong: in a UK government pilot, timed sessions mostly agreed with users’ diaries, but not for data analysis. Sources: Becker and others (METR), 2025, Yu and others, 2026, UK Department for Business and Trade, 2025, Australian Treasury, 2025.

  • 20 Example figurepeople given AI
  • × 1,840 Example figurehours a year each: 40 hours a week for 46 weeks
  • × 2.8% to 5.4% Self-reportedof their hours: what chatbot users in Denmark (2.8%) and users of generative AI in the United States (5.4%) said they saved. Humlum and Vestergaard, 2025, Bick, Blandin and Deming (St. Louis Fed), 2025. A trial across 66 firms measured two hours a week saved on email among those who used the tool (Dillon and others, 2025).
  • × 47% to 100% Self-reportedof them using it. The low estimate is the share of Danish workers using chatbots where the employer did not encourage it (Humlum and Vestergaard, 2025). The high estimate is everyone.
  • Low estimate: 0 because nobody has timed the work before and after AI.

Time put to use

How much of that freed time your company actually uses.

$0 to $39,700 0 to 994 hours a year

Time saved counted as money Costs first, gains later

How this line is worked out

Time saved counted as money. The low estimate is zero. Studies that checked found time saved, with no change in earnings, recorded hours or tasks. A Danish study of pay records rules out any effect on earnings or hours above 2%. Sources: Humlum and Vestergaard, 2026, Dillon and others, 2025, UK Department for Business and Trade, 2025.

Costs first, gains later. Of the costs you entered, $3,200 falls only in the first year. Without it the result is −$4,800 to +$19,000 a year. New technology tends to cost before it pays: US factories that adopted industrial AI first became less productive. Nobody knows how long the wait is for a company like yours. In a survey of nearly 6,000 executives, 89% said AI had not changed their firm’s productivity in the past three years. Sources: Brynjolfsson, Rock and Syverson, 2018, McElheran and others, 2025, Acemoglu, 2024, Yotzov and others, 2026.

  • 50% Example figureof freed time put to use
  • Low estimate: 0 because nobody has decided what the freed time is for.
  • No study gives the share of freed time that becomes useful work. For comparison only: in an Australian government trial about 40% of participants said they moved time to higher-value work Self-reported (Australian government Copilot trial, 2024). In Denmark 80% of users said saved time went to other job tasks Self-reported (Humlum and Vestergaard, 2025), and 3% to 7% of reported savings showed up in workers’ pay Measured in a study (Humlum and Vestergaard, 2025). None of these is your share.

Running cost

Licences, set-up, training, checking and rework.

$8,000 to $23,900 the cost of 200 to 597 hours of their time

Survey figure, not yours

How this line is worked out

Survey figure, not yours. Checking is set to 40% of the time put to use. It is the only published figure: what employees told a survey by Workday, which sells workplace software. Set aside your own time for checking to replace it. Sources: Workday survey, 2026.

  • $4,800 Example figurelicences: 20 people × $20 a month × 12
  • + $3,200 Example figureset-up and training, first year only: 20 people × 4 hours × $40
  • + $0 to $15,900 Self-reportedchecking and rework: 40% of the time put to use, from a vendor survey of 3,200 full-time employees who use AI, at organizations with $100M+ revenue, North America, EMEA and APAC, November 2025 (Workday survey, 2026). The more time is put to use, the more checking it takes, so the low estimate of cost goes with the low estimate of time put to use.

Result, first year

Time put to use, in money, minus running cost.

−$8,000 to +$15,800 the value of −200 to 396 hours of their time

Risk: what it cost others

What went wrong at other organizations and what it cost them. Not part of the sum.

AI output went out unchecked, or you claimed more than you tested. Here is what that cost other organizations.

You have not said whether a person checks what goes out. Tick the mistake or the practice to settle this line.

A person checks what goes out, and you claim only what was tested. Doing both guards against cases like these. Checking is not a guarantee: a study that combined 106 experiments found that a person working with AI did worse, on average, than the better of the two working alone Measured in a study (Vaccaro, Almaatouq and Malone, 2024).

12 cases and what each cost
  • Air Canada A Canadian tribunal ordered Air Canada to pay a customer C$812.02 after its website chatbot gave wrong information about fares. Court or regulator record decisions.civilresolutionbc.ca
  • The law firm in Mata v. Avianca A New York court fined a law firm and two lawyers $5,000 for filing six court decisions that ChatGPT had made up. Court or regulator record storage.courtlistener.com
  • Morgan & Morgan A Wyoming court fined three lawyers $5,000 in total ($3,000 and two fines of $1,000) over eight made-up cases from their firm's own AI tool. Court or regulator record storage.courtlistener.com
  • Butler Snow After five made-up citations, a US law firm paid another firm's 28 lawyers to check more than 2,400 citations in 330 filings. Court or regulator record storage.courtlistener.com
  • A law centre and a barrister, heard with Primus Solicitors In an English case heard with Primus Solicitors', a law centre and a barrister were each ordered to pay £2,000 in wasted costs over five non-existent cases; the Primus filing had 18. Court or regulator record judiciary.uk
  • accessiBe The US consumer regulator ordered accessiBe to pay $1 million over claims that its AI plug-in made any website accessible. Court or regulator record ftc.gov
  • DoNotPay The US consumer regulator ordered DoNotPay to pay $193,000 for selling an untested 'robot lawyer'. Court or regulator record ftc.gov
  • Delphia and Global Predictions The US securities regulator fined two investment advisers $225,000 and $175,000 for advertising AI they did not have. Court or regulator record sec.gov
  • Rite Aid Rite Aid was banned from using facial recognition for surveillance for five years after deploying it without testing its accuracy. Court or regulator record ftc.gov
  • Victoria’s child protection department After a child-protection worker used ChatGPT for a court report, the regulator found nearly 900 staff had opened the site and ordered the department to block such tools. Court or regulator record ovic.vic.gov.au
  • Klarna Klarna's chief executive said the company had weighed cost too heavily in customer service and got lower quality; it began hiring human agents again. Reported, not measured customerexperiencedive.com
  • WPP WPP raised spending on its AI platform to £300 million a year; its 2025 income still fell and it cut its forecast to a fall of 5.5% to 6.0%. Reported, not measured wpp.com

These costs fell on others. Nobody knows how likely any of them is for you, so none is in the sum.

Every number on this page carries a label that says what kind of figure it is:

  • Measured in a study
  • Self-reported
  • Your input
  • Example figure

A measured figure belongs to the people and the work it was measured on. Each one says who they were and links to the study. Where studies show which way something goes but not by how much, you get a note and its sources, not a number.

Treat a company’s own figure with care. Of the company cases on this site that report a result, 58% rest only on figures from the company itself: the evidence gap.

What this tool will not show you, and why
  • A share of saved time that “becomes value”, taken from a study. No study gives one. The share is your input. The published figures next to it measure other things: what people said they did with the time, or what reached their pay.
  • A number for what usage targets, quotas or leaderboards do. Nobody has measured it. You get the reported cases. You also get what a survey of about 25,000 Danish workers found: where employers encouraged chatbot use, 83% of workers used them, against 47% elsewhere (Humlum and Vestergaard, 2025).
  • Pilot “failure rates” in the sums. Four are widely quoted, and none measures the chance that your project fails.
    • 95%. The share of organizations with no measurable profit-and-loss return, in a report based on interviews at 52 organizations. Success meant that users or executives “remarked” on a lasting effect (MIT NANDA report, 2025).
    • 42%. The share of 1,006 survey respondents who said their company had abandoned most of its AI initiatives before production (S&P Global survey, 2025).
    • 30%. A prediction made in 2024, with no published method (Gartner, 2024).
    • 80%. Often credited to RAND, whose report introduces it with “By some estimates”. RAND interviewed 65 data scientists and engineers about why projects fail and did not count failures (RAND, 2024).
  • The chance or the expected cost of something going wrong. See the fourth line, Risk: what it cost others. It lists what went wrong at other organizations and says why none of it is in the result.
  • One gain for “all staff”. Each measured gain belongs to one kind of work and one group of people. The figure for customer-support agents says nothing about your accountants.
  • A return that depends on company size. See the note under the size buttons in Your numbers.
  • An economy-wide forecast as your return. See the next section.

The conservative view

The economist Daron Acemoglu estimated in 2024 how much AI will add to the productivity of the United States in ten years. His sum has four numbers.

  1. 20% of tasks × 23% = 4.6% of tasks AI could do or help with 20% of work tasks, counted by the pay attached to them. 23% of those are worth automating within ten years. He borrows the 23% from a cost study of computer-vision systems (Svanberg and others, 2024) and calls that step “not an innocuous step”.
  2. 27% labour saving × 0.535 = 14.4% cost saving 27% is the average of two experiments: the writing tasks and the customer support used in this tool. Labour makes up 53.5% of the cost of those tasks. Adding an experiment with a coding assistant raises the 27% to 36%.
  3. 4.6% × 14.4% = 0.66% Total extra productivity over ten years. He calls it an upper limit. If harder tasks save 7% and not 27%, it is 0.53%. In GDP: 0.93% to 1.16% more over ten years. He adds that the sum leaves out the cost of reorganizing work, so it may still be too high.

Calculated by the study’s authors Source: Acemoglu, 2024. It is a ten-year figure for a whole economy, not a company’s return. A company that picks its tasks well can be far above it, and a company that picks none is at zero.

Most other forecasts are far higher

Acemoglu United States
McKinsey Global Institute World, to 2040
OECD paper United States and other G7
Aghion and Bunel Same formula, their inputs
Goldman Sachs World
Extra productivity growth a year from AI, in percentage points. The forecasts do not all measure the same kind of productivity or cover the same countries. Acemoglu’s is his ten-year total divided by ten.
Show the numbers as a table
ForecastCoversPoints a year
AcemogluUnited Statesabout 0.07
McKinsey Global InstituteWorld, to 20400.1 to 0.6
OECD paperUnited States and other G70.25 to 0.6
Aghion and BunelSame formula, their inputs0.07 to 1.24, middle 0.68
Goldman SachsWorld1.5

Aghion and Bunel used Acemoglu’s own formula with their reading of the evidence and got about ten times his figure. The answer depends on the first two numbers in the sum. Two more headline numbers measure something else. The IMF’s “almost 40%” is the share of jobs exposed to AI (IMF blog, 2024). McKinsey’s $2.6 trillion to $4.4 trillion a year is the value of 63 uses it chose to study (McKinsey Global Institute, 2023).

Sources

Every figure this tool uses or shows, with what kind of figure it is. A link to this site leads to a page that says what the study or case did and found.

Figures used in the sums

Each is used in the sum when you tick its box or choose its kind of work.

  1. Noy, S. and Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187-192 (peer-reviewed). doi.org Measured in a study In an experiment on short professional writing tasks, ChatGPT cut the time taken by 40% and raised quality by 18%. On this site.
  2. Brynjolfsson, E., Li, D. and Raymond, L. (2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889-942 (peer-reviewed). Abstract read on arXiv:2304.11771. arxiv.org Measured in a study In a large customer-support operation, agents given an AI assistant resolved 15% more issues per hour on average. On this site.
  3. Dell’Acqua, F. et al. (2026). Navigating the Jagged Technological Frontier. Organization Science, 37(2), 403-423 (peer-reviewed). doi.org Measured in a study Consultants using GPT-4 did 12.2% more tasks 25.1% faster on work the AI does well, and were 19 percentage points less likely to be right on a task the AI could not do well. On this site.
  4. Cui, K. Z., Demirer, M., Jaffe, S., Musolff, L., Peng, S. and Salz, T. (February 2025 version). The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers. Working paper (version hosted by MIT Economics; not peer-reviewed at that version). economics.mit.edu Measured in a study Across three company-run randomised trials, developers given an AI coding assistant completed about 26% more tasks.
  5. Becker, J., Rush, N., Barnes, B. and Rein, D. (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. arXiv:2507.09089v2 (preprint, not peer-reviewed). arxiv.org Measured in a study In a randomised trial, experienced open-source developers took 19% longer on their own real tasks when allowed to use AI tools.
  6. Humlum, A. and Vestergaard, E. (2025). Large Language Models, Small Labor Market Effects. Becker Friedman Institute Working Paper 2025-56, May 2025 (working paper). bfi.uchicago.edu Self-reported Across a whole labour market, chatbot users reported saving 2.8% of their working hours on average. On this site.
  7. Bick, A., Blandin, A. and Deming, D. (27 February 2025). The Impact of Generative AI on Work Productivity. Federal Reserve Bank of St. Louis, On the Economy blog. stlouisfed.org Self-reported In a US national survey, generative-AI users said it saved them 5.4% of their work hours; across all workers that is 1.4%.
  8. Humlum, A. and Vestergaard, E. (2025). Large Language Models, Small Labor Market Effects. Becker Friedman Institute Working Paper 2025-56 (May 2025 version). bfi.uchicago.edu Self-reported Where Danish employers encouraged chatbot use, the share of workers using chatbots nearly doubled, from 47% to 83%, and reported time savings were larger.
  9. Workday (14 January 2026). New Workday Research: Companies Are Leaving AI Gains on the Table. Press release for 'Beyond Productivity: Measuring the Real Value of AI', fielded by Hanover Research (vendor survey). newsroom.workday.com Self-reported In a vendor survey, employees said nearly 40% of the time AI saved them was lost to fixing and checking its output.
  10. Anthropic. Claude pricing page (vendor price list), checked 13 October 2026. claude.com Vendor list price Anthropic lists its Claude Team plan at $20 per seat a month billed annually, or $25 billed monthly, with a premium seat at $100.

Time freed: shown for comparison

  1. METR (24 February 2026). We are Changing our Developer Productivity Experiment Design. METR blog. metr.org Measured in a study METR's follow-up found smaller slowdowns and says its own new data are unreliable, because developers no longer want to do tasks without AI.
  2. UK Government Digital Service (2 June 2025). Microsoft 365 Copilot experiment: cross-government findings report. gov.uk Self-reported In the UK government's Copilot trial, users reported saving 26 minutes a day on average. On this site.
  3. Dillon, E. W., Jaffe, S., Immorlica, N. and Stanton, C. T. (2025). Shifting Work Patterns with Generative AI. NBER Working Paper 33795 (working paper; three authors work at Microsoft Research). nber.org Measured in a study In a randomised trial across 66 firms, workers who used an AI assistant in their office software spent two fewer hours a week on email.
  4. Digital Transformation Agency (October 2024). Evaluation of the whole-of-government trial of Microsoft 365 Copilot: Summary of evaluation findings. digital.gov.au Self-reported The same trial found the perceived saving, about an hour a day, was limited to three tasks: summarising, first drafts and searching for information.

Time put to use: shown for comparison

  1. Dillon, E. W., Jaffe, S., Immorlica, N. and Stanton, C. T. (2025). Shifting Work Patterns with Generative AI. NBER Working Paper 33795. nber.org Measured in a study The same trial found no change in how much work people did or what kind: the time saved did not show up as more output.
  2. Humlum, A. and Vestergaard, E. (2025, revised March 2026). Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI. NBER Working Paper 33777 (working paper). nber.org Measured in a study Two years after ChatGPT appeared, a Danish study linking surveys to pay records found no effect on earnings or recorded hours, and rules out effects above 2%. On this site.
  3. Humlum, A. and Vestergaard, E. (2025). Large Language Models, Small Labor Market Effects. Becker Friedman Institute Working Paper 2025-56 (May 2025 version). bfi.uchicago.edu Measured in a study In the Danish study, only 3% to 7% of workers' reported time savings showed up as higher earnings. On this site.
  4. Humlum, A. and Vestergaard, E. (2025). Large Language Models, Small Labor Market Effects. Becker Friedman Institute Working Paper 2025-56 (May 2025 version). bfi.uchicago.edu Self-reported Asked what they did with time saved by chatbots, 80% of Danish users said other job tasks, fewer than 10% said breaks or leisure, and 25% said more of the same task.
  5. Yotzov, I., Barrero, J. M., Bloom, N., Bunn, P., Davis, S. J. et al. (2026). Firm Data on AI. NBER Working Paper 34836, February 2026, revised March 2026 (working paper). nber.org Self-reported In a survey of nearly 6,000 executives in four countries, 89% said AI had had no effect on their firm's labour productivity over the past three years.
  6. Yotzov, I., Barrero, J. M., Bloom, N. et al. (2026). Firm Data on AI. NBER Working Paper 34836. nber.org Self-reported The same executives expect AI to raise productivity at their firms by 1.4% on average over the next three years.
  7. UK Department for Business and Trade (August 2025). The evaluation of the M365 Copilot pilot in the Department for Business and Trade. assets.publishing.service.gov.uk Measured in a study A UK government department that measured its Copilot pilot found time savings on some tasks and no evidence that they led to higher productivity. On this site.
  8. Digital Transformation Agency (October 2024). Evaluation of the whole-of-government trial of Microsoft 365 Copilot: Summary of evaluation findings. Evaluation by Nous Group. digital.gov.au Self-reported In the Australian government's Copilot trial, about 40% of participants said they were able to move saved time to higher-value work.
  9. Australian Centre for Evaluation, The Treasury (February 2025). Evaluation of a trial of generative artificial intelligence (Copilot) in The Department of the Treasury. evaluation.treasury.gov.au Self-reported In the Australian Treasury's Copilot trial, 85% expected the tool to speed up their work beforehand and 49% said afterwards that it had.
  10. McElheran, K., Yang, M.-J., Kroff, Z. and Brynjolfsson, E. (2025). The Rise of Industrial AI in America: Microfoundations of the Productivity J-curve(s). US Census Bureau, Center for Economic Studies Working Paper CES 25-27, April 2025 (working paper). www2.census.gov Measured in a study In US manufacturing, plants that adopted industrial AI first became less productive before gains appeared; the losses fell mainly on older businesses.
  11. Brynjolfsson, E., Rock, D. and Syverson, C. (2018). The Productivity J-Curve: How Intangibles Complement General Purpose Technologies. NBER Working Paper 25148; published in American Economic Journal: Macroeconomics, 13(1), 2021, 333-372 (peer-reviewed). nber.org Calculated by the study’s authors A new general-purpose technology needs unmeasured investment in processes and skills first, so measured productivity dips before it rises: the 'productivity J-curve'.
  12. Babina, T., Fedyk, A., He, A. and Hodson, J. (2024). Artificial intelligence, firm growth, and product innovation. Journal of Financial Economics, 151, 103745 (peer-reviewed). sciencedirect.com Measured in a study US firms that invested more in AI grew faster in sales, employment and market value, mainly by launching new products.
  13. Czarnitzki, D., Fernández, G. P. and Rammer, C. (2023). Artificial intelligence and firm-level productivity. Journal of Economic Behavior & Organization, 211, 188-205 (peer-reviewed, open access). sciencedirect.com Measured in a study German firms that used AI were more productive, and the link held when the authors corrected for which firms choose to adopt.
  14. Aghion, P., Bunel, S., Jaravel, X., Mikaelsen, T., Roulet, A. and Søgaard, J. (2025). How Different Uses of AI Shape Labor Demand: Evidence from France. AEA Papers and Proceedings, 115, 62-67 (lightly reviewed conference volume). ideas.repec.org Measured in a study French firms that adopted AI saw employment and sales rise afterwards, with different results for different uses.
  15. Jiang, W., Park, J., Xiao, R. and Zhang, S. (2025). AI and the Extended Workday: Productivity, Contracting Efficiency, and Distribution of Rents. NBER Working Paper 33536 (working paper). nber.org Measured in a study A US time-diary study found that people in jobs more exposed to AI worked longer hours and had less leisure, not more.
  16. McKinsey & Company (5 November 2025). The State of AI: Global Survey 2025 (consultancy survey). web.archive.org Self-reported In McKinsey's 2025 survey about 6% of organizations credited AI with 5% or more of operating profit, and they were nearly three times as likely to have redesigned workflows. On this site.

Running cost: shown for comparison

  1. BetterUp Labs and Stanford Social Media Lab (2025). Workslop: The Hidden Cost of AI-Generated Busywork. Reported in Niederhoffer, K., Kellerman, G. R., Lee, A., Liebscher, A., Rapuano, K. and Hancock, J. T. (22 September 2025). AI-Generated “Workslop” Is Destroying Productivity. Harvard Business Review (vendor survey; magazine article, not peer-reviewed). betterup.com Self-reported In a BetterUp and Stanford survey, 40% of US desk workers said they had received 'workslop', polished-looking AI output that leaves the real work to the receiver, in the past month.
  2. UK Department for Business and Trade (August 2025). The evaluation of the M365 Copilot pilot in the Department for Business and Trade. assets.publishing.service.gov.uk Self-reported In the UK department's Copilot pilot, 22% of diary respondents had spotted made-up content; checking was inconsistent. On this site.
  3. UK Department for Business and Trade (August 2025). The evaluation of the M365 Copilot pilot in the Department for Business and Trade. assets.publishing.service.gov.uk Self-reported Some tasks took longer with Copilot: scheduling cost 0.6 hours more per task and generating images 0.5 hours more, by users' diaries. On this site.
  4. Humlum, A. and Vestergaard, E. (2025). Large Language Models, Small Labor Market Effects. Becker Friedman Institute Working Paper 2025-56 (May 2025 version). bfi.uchicago.edu Self-reported Chatbots created new job tasks for 8.4% of workers in the Danish study, including people who do not use the tools themselves. On this site.
  5. Australian Centre for Evaluation, The Treasury (February 2025). Evaluation of a trial of generative artificial intelligence (Copilot) in The Department of the Treasury. evaluation.treasury.gov.au Calculated by the study’s authors The Australian Treasury worked out that a mid-level official needs to move about 13 minutes a week to higher-value work to cover the Copilot licence.
  6. Microsoft. Microsoft 365 Copilot pricing page (vendor price list), checked 13 October 2026. microsoft.com Vendor list price Microsoft lists its Copilot add-on for small and mid-sized businesses at $18 per user a month paid yearly (shown as reduced from $21), on top of a Microsoft 365 plan.
  7. France Num (French government). Case page on SAM, a retirement advisory firm that automated its client files with AI. francenum.gouv.fr Reported, not measured A two-person French advisory firm reports what its AI set-up cost: about €3,000 for a server, nearly €650 a month in licences, and about 60 hours of set-up work. On this site.
  8. Fortune (9 April 2026). Report on Meta's 'Claudeonomics' employee token leaderboard and its removal, citing The Information and a Meta statement. fortune.com Reported, not measured An employee-built leaderboard at Meta ranked staff by AI tokens used; Fortune calculated that the top user's 30-day total could have cost more than $1.4 million at list prices.
  9. heise online (29 May 2026). Too much tokenmaxxing? Amazon stops internal AI ranking. Summary of a Financial Times report, quoting Amazon senior vice-president Dave Treadwell. heise.de Reported, not measured Amazon dropped an internal AI-usage ranking after employees ran AI agents on needless tasks to climb it, according to a Financial Times report.

Risk: what it cost others

  1. Customer Experience Dive (9 May 2025). Report on Klarna reinvesting in human customer-service staff, quoting chief executive Sebastian Siemiatkowski's remarks to Bloomberg. customerexperiencedive.com Reported, not measured Klarna's chief executive said the company had weighed cost too heavily in customer service and got lower quality; it began hiring human agents again. On this site.
  2. Moffatt v. Air Canada, 2024 BCCRT 149. British Columbia Civil Resolution Tribunal. decisions.civilresolutionbc.ca Court or regulator record A Canadian tribunal ordered Air Canada to pay a customer C$812.02 after its website chatbot gave wrong information about fares. On this site.
  3. Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y. 22 June 2023). Opinion and order on sanctions. storage.courtlistener.com Court or regulator record A New York court fined a law firm and two lawyers $5,000 for filing six court decisions that ChatGPT had made up. On this site.
  4. Wadsworth v. Walmart Inc., No. 2:23-cv-118 (D. Wyo. 2025). Order on sanctions. storage.courtlistener.com Court or regulator record A Wyoming court fined three lawyers $5,000 in total ($3,000 and two fines of $1,000) over eight made-up cases from their firm's own AI tool. On this site.
  5. Johnson v. Dunn, No. 2:21-cv-1701 (N.D. Ala. 23 July 2025). Sanctions order. storage.courtlistener.com Court or regulator record After five made-up citations, a US law firm paid another firm's 28 lawyers to check more than 2,400 citations in 330 filings. On this site.
  6. Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin). judiciary.uk Court or regulator record In an English case heard with Primus Solicitors', a law centre and a barrister were each ordered to pay £2,000 in wasted costs over five non-existent cases; the Primus filing had 18. On this site.
  7. US Federal Trade Commission (22 April 2025). FTC approves final order requiring accessiBe to pay $1 million. Press release. ftc.gov Court or regulator record The US consumer regulator ordered accessiBe to pay $1 million over claims that its AI plug-in made any website accessible. On this site.
  8. US Federal Trade Commission (February 2025). FTC finalizes order with DoNotPay that prohibits deceptive ‘AI lawyer’ claims, imposes monetary relief. Press release. ftc.gov Court or regulator record The US consumer regulator ordered DoNotPay to pay $193,000 for selling an untested 'robot lawyer'. On this site.
  9. US Securities and Exchange Commission (March 2024). SEC Charges Two Investment Advisers with Making False and Misleading Statements About Their Use of Artificial Intelligence. Press release 2024-36. sec.gov Court or regulator record The US securities regulator fined two investment advisers $225,000 and $175,000 for advertising AI they did not have. On this site.
  10. US Federal Trade Commission (19 December 2023). Rite Aid banned from using AI facial recognition after FTC says retailer deployed technology without reasonable safeguards. Press release. ftc.gov Court or regulator record Rite Aid was banned from using facial recognition for surveillance for five years after deploying it without testing its accuracy. On this site.
  11. Office of the Victorian Information Commissioner (September 2024). Investigation into the use of ChatGPT by a Child Protection worker. ovic.vic.gov.au Court or regulator record After a child-protection worker used ChatGPT for a court report, the regulator found nearly 900 staff had opened the site and ordered the department to block such tools. On this site.
  12. WPP plc (30 October 2025). Third Quarter Trading Update 2025; and WPP Annual Report on Form 20-F for 2024. wpp.com Reported, not measured WPP raised spending on its AI platform to £300 million a year; its 2025 income still fell and it cut its forecast to a fall of 5.5% to 6.0%. On this site.

Background: usage targets, what users say, and checking

  1. Becker, J., Rush, N., Barnes, B. and Rein, D. (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. arXiv:2507.09089v2. arxiv.org Measured in a study The developers in the METR trial expected AI to make them 24% faster, and afterwards believed it had made them 20% faster, while the measured result was 19% slower.
  2. Yu, S., Cheng, M., Jabbar, A., Sucholutsky, I., Collins, K. M., Jurafsky, D. and Hawkins, R. D. (2026). The efficiency-gain illusion: People underestimate the rate of AI use and overestimate its benefits on simple tasks. arXiv:2605.22687 (preprint, not peer-reviewed). arxiv.org Measured in a study In three pre-registered experiments on simple tasks, people expected AI to save 55.7 seconds and it saved 7.5.
  3. UK Department for Business and Trade (August 2025). The evaluation of the M365 Copilot pilot in the Department for Business and Trade. assets.publishing.service.gov.uk Measured in a study In the UK department's pilot, users' diaries reported time saved on data analysis, while a timed test found Copilot users did Excel analysis more slowly and worse. On this site.
  4. Fortune (25 August 2025). Report on Coinbase chief executive Brian Armstrong's AI coding mandate, quoting his remarks on the Cheeky Pint podcast. fortune.com Reported, not measured Coinbase's chief executive said he gave engineers a week to start using AI coding tools and fired some who did not.
  5. CNBC (7 April 2025). Shopify CEO says staffers need to prove jobs can't be done by AI before asking for more headcount. Report on Tobi Lütke's memo, which he posted on X. cnbc.com Reported, not measured Shopify's chief executive told staff that AI use would be part of performance reviews and that teams must show AI cannot do a job before asking for more people.
  6. Fortune (18 August 2025). Duolingo CEO admits his controversial AI memo ‘did not give enough context’ and insists the company never laid off full-time employees. fortune.com Reported, not measured Duolingo's chief executive said months later that his 'AI-first' memo 'did not give enough context'. On this site.
  7. Holmström, B. and Milgrom, P. (1991). Multitask Principal-Agent Analyses: Incentive Contracts, Asset Ownership, and Job Design. The Journal of Law, Economics, and Organization, 7 (special issue), 24-52 (peer-reviewed). econpapers.repec.org Calculated by the study’s authors The standard economic result on measuring the wrong thing: when only part of a job is measured and rewarded, effort moves to the measured part.
  8. Vaccaro, M., Almaatouq, A. and Malone, T. (2024). When combinations of humans and AI are useful: A systematic review and meta-analysis. Nature Human Behaviour, 8(12), 2293-2303 (peer-reviewed). doi.org Measured in a study Across 106 experiments, a human working with AI did worse on average than the better of the two alone. On this site.

Failure rates this tool leaves out

  1. Challapally, A., Pease, C., Raskar, R. and Chari, P. (July 2025). The GenAI Divide: State of AI in Business 2025. MIT NANDA (project report, not peer-reviewed; copy hosted by MLQ.ai). mlq.ai Reported, not measured The MIT NANDA report's '95%' is the share of organizations in its sample getting no measurable profit-and-loss return from generative AI; it also says only 5% of custom enterprise AI tools reached production.
  2. Challapally, A. et al. (July 2025). The GenAI Divide: State of AI in Business 2025. MIT NANDA. mlq.ai Reported, not measured The same report defines success as a tool that users or executives 'remarked' had a marked and lasting effect, which is a judgement and not a measurement.
  3. S&P Global Market Intelligence (30 May 2025). Generative AI experiences rapid adoption, but with mixed outcomes: Highlights from VotE: AI & Machine Learning. Summary of 'Voice of the Enterprise: AI & Machine Learning, Use Cases 2025' (vendor research; full report behind a paywall). spglobal.com Self-reported In S&P Global's 2025 survey, 42% of companies said they had abandoned most of their AI initiatives before production, up from 17% a year earlier.
  4. Gartner (29 July 2024). Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025. Press release. gartner.com Reported, not measured Gartner predicted in 2024 that at least 30% of generative-AI projects would be abandoned after proof of concept by the end of 2025.
  5. Ryseff, J., De Bruhl, B. F. and Newberry, S. J. (2024). The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed. RAND Corporation, RR-A2680-1. rand.org Reported, not measured RAND interviewed 65 data scientists and engineers about failed AI projects; the first cause they named was misunderstanding the problem, including models 'optimized for the wrong metrics'.

The conservative view and other forecasts

  1. Acemoglu, D. (2024). The Simple Macroeconomics of AI. NBER Working Paper 32487, May 2024. Published as: Economic Policy, 40(121), January 2025, 13-58, doi:10.1093/epolic/eiae042 (peer-reviewed journal; figures below quoted from the NBER text). nber.org Calculated by the study’s authors About 20% of US work tasks, weighted by wages, are exposed to AI in Acemoglu's calculation.
  2. Acemoglu, D. (2024). The Simple Macroeconomics of AI. NBER Working Paper 32487, May 2024. Published as: Economic Policy, 40(121), January 2025, 13-58, doi:10.1093/epolic/eiae042 (peer-reviewed journal; figures below quoted from the NBER text). nber.org Calculated by the study’s authors Acemoglu assumes that 23% of exposed tasks are worth automating within ten years, a figure he borrows from a study of computer-vision tasks.
  3. Svanberg, M., Li, W., Fleming, M., Goehring, B. and Thompson, N. (2024). Beyond AI Exposure: Which Tasks are Cost-Effective to Automate with Computer Vision? MIT FutureTech working paper, January 2024 (working paper, not peer-reviewed). futuretech-site.s3.us-east-2.amazonaws.com Calculated by the study’s authors A cost model of computer-vision tasks found that only 23% of the exposed wages were worth automating at the firm level, because of large up-front costs.
  4. Acemoglu, D. (2024). The Simple Macroeconomics of AI. NBER Working Paper 32487, May 2024. Published as: Economic Policy, 40(121), January 2025, 13-58, doi:10.1093/epolic/eiae042 (peer-reviewed journal; figures below quoted from the NBER text). nber.org Calculated by the study’s authors Multiplying the two shares, Acemoglu puts the part of the economy that AI will actually change within ten years at 4.6%.
  5. Acemoglu, D. (2024). The Simple Macroeconomics of AI. NBER Working Paper 32487, May 2024. Published as: Economic Policy, 40(121), January 2025, 13-58, doi:10.1093/epolic/eiae042 (peer-reviewed journal; figures below quoted from the NBER text). nber.org Calculated by the study’s authors On tasks where AI is used, Acemoglu takes the labour-cost saving to be 27%, the average of two experiments.
  6. Acemoglu, D. (2024). The Simple Macroeconomics of AI. NBER Working Paper 32487, May 2024. Published as: Economic Policy, 40(121), January 2025, 13-58, doi:10.1093/epolic/eiae042 (peer-reviewed journal; figures below quoted from the NBER text). nber.org Calculated by the study’s authors Because labour is only part of the cost of a task, Acemoglu converts the 27% labour saving into a 14.4% saving on total cost.
  7. Acemoglu, D. (2024). The Simple Macroeconomics of AI. NBER Working Paper 32487, May 2024. Published as: Economic Policy, 40(121), January 2025, 13-58, doi:10.1093/epolic/eiae042 (peer-reviewed journal; figures below quoted from the NBER text). nber.org Calculated by the study’s authors Acemoglu's headline result: AI raises US productivity by no more than about 0.66% in total over ten years.
  8. Acemoglu, D. (2024). The Simple Macroeconomics of AI. NBER Working Paper 32487, May 2024. Published as: Economic Policy, 40(121), January 2025, 13-58, doi:10.1093/epolic/eiae042 (peer-reviewed journal; figures below quoted from the NBER text). nber.org Calculated by the study’s authors For 'hard-to-learn' tasks, where there is no clear right answer to learn from, Acemoglu assumes a 7% labour saving in place of 27%, which lowers his ten-year productivity estimate to about 0.53%.
  9. Acemoglu, D. (2024). The Simple Macroeconomics of AI. NBER Working Paper 32487, May 2024. Published as: Economic Policy, 40(121), January 2025, 13-58, doi:10.1093/epolic/eiae042 (peer-reviewed journal; figures below quoted from the NBER text). nber.org Calculated by the study’s authors In GDP terms Acemoglu's estimate is roughly 0.9% to 1.2% in total over ten years.
  10. Acemoglu, D. (2024). The Simple Macroeconomics of AI. NBER Working Paper 32487, May 2024. Published as: Economic Policy, 40(121), January 2025, 13-58, doi:10.1093/epolic/eiae042 (peer-reviewed journal; figures below quoted from the NBER text). nber.org Calculated by the study’s authors Acemoglu notes that his figures ignore the cost and delay of reorganizing work, so even his small estimate may be too high for the first ten years.
  11. Goldman Sachs (5 April 2023). Generative AI could raise global GDP by 7%. Goldman Sachs Research article summarising Briggs and Kodnani (investment-bank research, not peer-reviewed). goldmansachs.com Calculated by the study’s authors Goldman Sachs Research forecast that generative AI could raise global GDP by 7% and productivity growth by 1.5 percentage points a year over ten years.
  12. McKinsey Global Institute (14 June 2023). The economic potential of generative AI: The next productivity frontier (consultancy report, not peer-reviewed). web.archive.org Calculated by the study’s authors McKinsey Global Institute estimated that generative AI could add 0.1 to 0.6 percentage points a year to labour productivity growth through 2040.
  13. McKinsey Global Institute (14 June 2023). The economic potential of generative AI: The next productivity frontier. web.archive.org Calculated by the study’s authors The same McKinsey report put the value of 63 generative-AI use cases at $2.6 trillion to $4.4 trillion a year.
  14. Aghion, P. and Bunel, S. (June 2024). AI and Growth: Where Do We Stand? Note published by the Federal Reserve Bank of San Francisco (not peer-reviewed). frbsf.org Calculated by the study’s authors Aghion and Bunel reran Acemoglu's formula with their own reading of the evidence and got 0.68 percentage points of extra productivity growth a year, about ten times his figure.
  15. Filippucci, F., Gal, P. and Schief, M. (2024). Miracle or Myth? Assessing the macroeconomic productivity gains from Artificial Intelligence. OECD Artificial Intelligence Papers No. 29, November 2024 (working paper). oecd.org Calculated by the study’s authors An OECD paper estimates that AI adds 0.25 to 0.6 percentage points a year to productivity growth over ten years.
  16. Georgieva, K. (14 January 2024). AI Will Transform the Global Economy. Let’s Make Sure It Benefits Humanity. IMF Blog, summarising IMF Staff Discussion Note SDN/2024/001 (Cazzaniga et al.). imf.org Calculated by the study’s authors The IMF's headline figure is about exposure, not growth: almost 40% of jobs worldwide, and about 60% in advanced economies, are exposed to AI.