AI productivity gains: where they help and where they don’t

AI productivity gains mean getting more useful output from the same resources, or completing the same work with fewer resources. With generative AI, the difference can come from time savings, better output quality or newly affordable work. Each needs its own measure. More generated content alone doesn't establish higher productivity.

Anthropic's engineers and researchers reported an average 50% productivity boost from using Claude in a study published in December 2025. They also estimated that 27% of their Claude-assisted work would otherwise have stayed undone: useful tools, documentation, tests and exploratory projects that hadn't justified the effort before.

Those are self-reported estimates. But that second finding makes the productivity question more interesting. An assistant can help someone finish an existing task faster. It can also change which tasks are worth attempting.

I asked Codex to compare the original studies and company reports behind these claims. This review is for people deciding where AI assistance belongs in their work: it compares task experiments, employee surveys, company accounts and economic estimates. The evidence supports useful gains on some jobs, alongside extra work and quality penalties on others.

The useful distinction is between completing a task faster and creating more value. Support workers can gain access to knowledge they need; experts may spend longer explaining context and checking suggestions. The sections below follow that difference from the individual task to business returns, with the benefit and the limiting condition kept together.

AI adoption inside Anthropic: more output, more work to supervise

Anthropic's October 2026 study compares 192 engineers who adopted agentic tools with 74 who hadn't, tracking changes over time. The analysis accounts for persistent differences between people and changes affecting a given month. It also examines later code deletions and reverted changes as indicators of quality, finding no difference in deletion or revert rates in those comparisons.

That is substantially more informative than a productivity questionnaire. It connects adoption to observed work and checks some possible consequences.

Adoption wasn't assigned randomly, though. Some autonomous-agent users already had rising output before adoption. The reported fivefold increase in commits per engineer over six months is a descriptive trend, rather than a clean estimate that AI alone multiplied productivity by five. Code checks also capture only part of software quality.

The upside is an organization getting more work done with tools built into its daily routines. The constraint is the human attention needed to supervise that work, which the report also identifies as a bottleneck. Its total financial return, and the effect in a very different company, remain open.

AI tools in support and coding: useful help with a defined job

A customer-support study published in the Quarterly Journal of Economics followed 5,172 employees as a company introduced an AI assistant. The preferred estimate was a 15.2% increase in successfully resolved issues per hour.

That outcome matters. A support team benefits from resolving customer problems, so the measure connects speed with a useful result. Less experienced workers generally gained more; the most experienced saw smaller gains and some quality trade-offs. The phased rollout gives researchers a comparison over time, although it wasn't a fully randomized deployment.

The fieldwork largely dates to 2020–21 and used a GPT-3-based assistant. It demonstrates a particular mechanism: making the knowledge behind good support more available to other workers. It doesn't describe what every current model will do.

There is evidence of AI productivity gains in ordinary software work, too. Three randomized field experiments, covering 4,867 developers, reported about 26% more completed tasks after access to a coding assistant. The pooled estimate was uncertain, and the full company-specific task and quality definitions weren't available in the material reviewed. It is evidence about the reported task measure, with earlier coding tools, rather than a measured 26% increase in software's business value.

The practical fit here is assistance that helps people complete a defined task and leaves a result that can be checked. The support findings suggest particular value when the tool brings useful knowledge to someone who has less experience. The trade-off is that the same advice may add little for an expert, or introduce quality problems. The measure needs to include the final result and the work required to correct it.

Professional writing: a useful draft still needs the right context

Writing tasks show another practical fit. In the March 2, 2023 working-paper version of Noy and Zhang's experiment, 444 professionals were assigned short writing jobs, and half were randomly given access to ChatGPT. The AI group took less time and received better ratings from evaluators who didn't know the treatment assignment. The tasks included press releases, short reports and delicate emails.

That supports generative AI as an aid to producing drafts under a clear brief. The experiment's small, self-contained assignments required little company-specific knowledge; they don't establish the same gain on documents that depend on a customer's history or an organization's unwritten constraints. The figures and methods here refer to that opened early version, not a current-model test.

Software development and consulting: speed can come with a penalty

In a randomized study involving 758 BCG consultants, GPT-4 helped participants complete suitable tasks 25.1% faster and finish more of them. For a difficult management task outside the tested capabilities of the model, the AI groups were approximately 19 percentage points less likely to reach the correct answer than the control group.

The research was first released in 2023 and published in 2026; it concerns GPT-4 as it existed in June 2023. Its useful lesson is about matching a tool to a task and checking the outcome. A new publication date doesn't update the model in the experiment.

Experienced software developers: more context can mean more work

METR found a different problem in its 2025 experiment with experienced open-source developers. Sixteen developers worked on 246 real issues in repositories they knew well. Tasks were randomly assigned to allow or prohibit AI. With the early-2025 tools, completion took 19% longer, even though participants subsequently estimated that AI had made them faster.

The February 2026 follow-up complicates using that result as a verdict on newer tools. Participation and task selection changed, some developers avoided submitting work they didn't want to do without AI, and parallel agents made time harder to measure. METR considered the newer estimate unreliable.

These are useful examples of where assistance can get in the way. A plausible management answer can steer someone toward the wrong conclusion. A developer who already knows a repository may spend extra time explaining context and inspecting suggestions. That is a reason to compare completion time and correctness on the actual job, rather than assume the tool helps because it generates code quickly. The newer METR data still don't supply a reliable current effect size.

Generative artificial intelligence: useful ideas still need somewhere to go

Saving time on existing work is only one route to value. Another is getting access to knowledge or capabilities that were previously difficult to combine.

In P&G's Cybernetic Teammate experiment, participants worked on product-development problems alone or in pairs, with or without AI. In the published main analysis of 791 participants, individuals using AI produced better-rated proposals and reached the performance of teams working without it. Their proposals also brought technical and commercial perspectives together more effectively.

These were evaluated proposals from a workshop. The study didn't track the revenue of products launched from them. Its contribution is showing a plausible way AI can help someone work across a boundary that would otherwise require a colleague's expertise.

This makes AI a promising aid for exploring options and connecting areas of expertise. The limit is what happens after the proposal: someone still has to establish whether it works and whether anybody needs it. A higher-rated workshop idea hasn't yet faced production or customers.

The same distinction applies to the extra work reported by Anthropic's employees. A newly affordable test or dashboard can catch an error or improve a decision. It can also become another thing to review and maintain. A useful test is whether the new output gets used and makes something better. Counting only reduced payroll would miss that value; counting every new file would inflate it.

Business value: personal productivity gains can stall elsewhere

The 2026 McKinsey survey puts two measures next to each other: 80% of respondents reported personal productivity improvements, while 37% attributed a positive contribution to their company's earnings before interest and taxes, or EBIT, to AI.

These are shares of respondents reporting different outcomes. They don't mean a productivity increase of 80% or an earnings increase of 37%.

The survey covered 1,719 respondents in 97 countries, with fieldwork in May and June 2026. It leans toward larger businesses: 36% worked at companies with more than $1 billion in revenue. About a fifth said AI operating costs, including tokens, had constrained usage. This is useful evidence about reported experience, rather than an audit of AI's net contribution.

Other surveys show why a single success rate would be misleading:

The firms, questions and collection periods differ, so these percentages shouldn't be averaged. Together they point to an opportunity and a constraint: employees can use AI to improve their work, while integration costs, review capacity or a lack of demand for additional output can limit the business return. The surveys don't isolate which constraint explains each company's experience.

OpenAI's enterprise report announced in December 2025 adds a customer perspective. Its survey of 9,000 workers in nearly 100 companies reported 40–60 minutes saved per active usage day and newly manageable tasks for many respondents. These are users' estimates, alongside usage data. They help explain why people keep using the tools; they don't establish how much time ultimately became lower costs or additional revenue.

Support costs: a saving needs the full expense picture

Klarna provides a more concrete example. Its 2025 annual report lists customer-service and operations expenses of $240 million in 2023, $203 million in 2024 and $207 million in 2025. Separately, management estimated $39 million in AI support savings for 2024.

The expense figures are recorded financial results. The amount attributed to AI is a management estimate. Other efficiency measures, changing volumes and the mix of work could also affect the expense line. The report doesn't turn those figures into a fully isolated calculation of AI returns after every associated cost.

That distinction also helps with the next period: Q2 2026 support and operations costs rose to $57 million from $48 million a year earlier. Rising costs alone don't settle whether a system is useful when the business and workload are changing. A meaningful comparison needs the volume and quality delivered for that spending.

Time savings and aggregate productivity: where the gains can go

What measured productivity means

The Bureau of Labor Statistics distinguishes labor productivity from total factor productivity. Labor productivity relates output to hours worked. Total factor productivity relates output to combined inputs, including labor, capital and, in some measures, purchased inputs. A worker producing more per hour may also be using more computing or other resources. That is why a task speedup, annual productivity growth and a net financial return answer different questions. None is a shortcut to measuring AI's contribution to the entire economy.

How AI productivity gains are estimated from conversations

Anthropic's November 25, 2025 analysis of Claude conversations estimates task duration with and without AI across 100,000 transcripts. The model estimates roughly 80% less task time overall. Selected examples show why those numbers need their measurement boundary:

Within those selected conversations, Claude estimated 80% less time for financial analyst tasks such as interpreting financial data, 87% less time for writing invoices, memos and similar documents, and about 95% less time for compiling information from reports. The comparison concerns the tasks people brought to Claude, not every part of those occupations.

These are model estimates for selected conversations, not independently timed, accepted work. The method omits work outside the chat, including refinement and validation. It is useful for finding promising tasks; it can overstate end-to-end savings.

The report extrapolates to 1.8% additional annual US labor productivity growth over a decade, assuming broad use of the observed capabilities. It weights tasks by estimated time shares and wage-based economic importance. That scenario illustrates possible scale; it doesn't incorporate an adoption path or measure the economy's realized gain.

Survey estimates can connect tasks to the economy

A November 2025 Federal Reserve Bank of St. Louis update pools surveys from February, May and August 2025. Including nonusers, respondents reported time savings equivalent to 1.6% of all work hours. Feeding those estimates into an aggregate production model suggested up to 1.3% higher labor productivity since ChatGPT's introduction.

This extends the question beyond enthusiastic customers of one supplier. But the modeled productivity effect still starts with self-reported time savings. Industry-level productivity growth was positively correlated with reported savings; the authors explicitly say that relationship cannot establish causation. The update also revised earlier adoption estimates after changing the survey-question sequence. Old survey numbers shouldn't quietly become current productivity statistics.

A reported average needs a population and a version date

An earlier St. Louis Fed analysis, published February 27, 2025, illustrates the denominator problem. In its November 2024 survey, AI users reported savings averaging 5.4% of work hours. For a 40-hour week, that works out to 2.2 hours. It wasn't an average saving of 2.2 hours for every worker: including nonusers, the reported saving was 1.4% of all work hours.

Savings also varied within the user group. Among people who had used generative AI in the previous week, 20.5% reported saving four or more hours. That describes a subset, not a typical four-hour saving.

That earlier article originally reported that 28% of U.S. workers used generative AI at work in August 2024, with usage roughly stable in November. The later update reweighted earlier surveys after finding that question order affected answers; its August 2024 work-adoption estimate is 33.3%. The lesson is practical: keep the date, population and survey version attached to a productivity statistic. These historical figures describe those surveys, not current usage or a guaranteed saving.

Task time and working-day time have different denominators

Consider a hypothetical task that occupies one fifth of a working day. Cutting its duration in half frees one tenth of the day before any extra checking or coordination. The task took 50% less time; the whole day didn't.

What happens next depends on the job. The person could resolve more customer issues, improve a proposal, finish earlier or reduce overtime. They could also spend the time checking a larger pile of generated work. These are different outcomes, with different beneficiaries.

DORA's 2025 research found AI adoption positively associated with software-delivery throughput, while the relationship with stability remained negative. It is observational evidence, but it identifies a practical place to look: faster production can expose weaknesses in testing and delivery. The system around the assistant affects how useful its output becomes.

A Danish study linking surveys to employment records found no measurable average effect on wages or recorded hours through the end of 2024, despite reported changes in tasks and use of saved time. A wage increase is one possible destination for a gain. Better service, more work within the same hours, lower prices or reduced strain are others. The study's data end in 2024, so it doesn't settle the effects of today's agents.

This is also where large economic estimates need care. Australia's Tech Council report estimates annual productivity gains worth $136,000 for an average small or medium business moving from no AI use to broad adoption. In its table, that is the difference between $517,000 and $381,000 in modeled annual value added. It is a scenario about economic production, with adoption and productivity assumptions, rather than a measured profit increase for each business. Evidence that AI helps elsewhere doesn't verify that particular amount.

Measuring AI productivity: where it helps, where it gets harder

The findings suggest a practical way to weigh AI productivity gains against their costs and drawbacks. These are conditions to check, rather than permanent rules about what current AI can or cannot do. An observed failure and a missing measurement are different things.

Define the accepted result

A practical assessment starts with a specific piece of work and a clear definition of a good result. For support, that might mean a correctly resolved issue. For software, a change that passes review and works reliably. For a report, findings that survive checking and help someone make the intended decision.

AI assistance: count human review and the full cost

Then count the work around it: preparation, retries, human review, corrections, integration and maintenance, as well as the model or subscription bill. Check whether faster creation causes a queue somewhere else. If the gain is a new capability, establish who uses it and what it improves.

A useful cost measure is total task cost divided by the number of fully accepted results. Compare the same kind of work with and without AI assistance, including the same quality standard. A cheap draft that needs expensive repair can lose; a more expensive run that consistently delivers usable work can win. This is a proposed assessment method, not a new experiment performed for this article.

Compare specific tasks before scaling AI use

For software development, separate writing code from code review, merging pull requests and releasing a working change. For customer support, count correctly resolved problems alongside time spent handling escalations. For professional services, decide whether a completed proposal is the output being purchased or an intermediate step toward a decision.

Keep the baseline comparable: the same type of task, similar difficulty, appropriate access to information and the same acceptance standard. Record actual human time, including work done outside the AI conversation. Elapsed time matters too, especially when parallel agents create a queue for one reviewer. A reported reduction in hours worked means something different from more accepted work within unchanged hours.

This helps separate two questions: whether AI systems improve individual tasks, and whether the organization can use that improvement. Large language models can make complex tasks cheaper to attempt, while the value of the result still depends on whether it is correct and useful. For a new task that wasn't previously attempted, a time comparison may be impossible; assess the capability and its cost directly. There is no need to invent a no-AI completion time.

Decide who benefits from higher productivity

The financial question comes after that. Avoided overtime or outside fees can become a cash saving. Additional capacity can create revenue when there is demand for it. A shorter, less exhausting day can be worthwhile on its own. The same freed hour can't be counted simultaneously as eliminated payroll and extra production.

What can we learn from this?

The studies give us good reasons to try AI on specific jobs. They also give us reasons to measure what happens after the first output appears.

Start where the benefit is concrete. AI-assisted drafting and summarizing can save time when the brief and source material are clear. Support assistants can make useful knowledge easier to apply. These benefits depend on enough context and a way to verify the result; help with one part of a job doesn't establish that the whole job can be handed over.

Follow the work to an accepted result. Faster drafting is useful when review and delivery can keep up. If checking, corrections or maintenance consume the saving, a faster model response hasn't improved the workflow. Count those steps before scaling up.

Decide what the gain is for. Extra capacity can improve service, make room for neglected work or create a less exhausting day. Cash savings and revenue need their own evidence. A worthwhile benefit doesn't have to be a profit increase, but it does need to be something someone can actually use.

The decision to keep using AI should rest on that result. More output is a promising start; useful, accepted work is the stronger reason to continue.

Research note: This article draws on a targeted review of published studies and company reports prepared with Codex, with sources checked through October 23, 2026. It is not a new workplace experiment or a systematic meta-analysis. Publication dates and fieldwork dates are distinguished where they affect the interpretation; company claims and self-reported estimates remain attributed to their sources.