Possibly—but 32.6% is not a stopwatch result showing that every software engineer writes code 32.6% faster. The figure is a market-implied estimate from an NBER working paper, based on how company share prices reacted to AI-related news between November 2022 and December 2025. A separate randomized field-study program found that developers given an AI coding assistant completed 26.08% more tasks on average, with a reported 10.3% standard error. Those results address different questions and should not be treated as interchangeable.
What the 32.6% estimate actually measures
The 32.6% figure comes from a 2026 National Bureau of Economic Research working paper, as reported by The Register and Computerworld. Its wording is precise: from November 2022 through December 2025, AI increased the market’s expected present value of software-engineering productivity by the equivalent of a permanent 32.6% increase.
That is a forward-looking valuation estimate. It captures what investors believed AI could do for software-heavy companies, not a direct measurement of how long individual developers took to complete tickets.
How the market estimate was constructed
The researchers compared firms with different shares of payroll devoted to software engineering. They then examined whether those firms’ stock prices rose more when an index of AI-related stocks rose. As Chen Lian of the University of California, Berkeley, described the approach, the study asks whether companies with larger software-engineering payroll shares experience larger stock-price increases when the AI stock index rises.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
#1 Best Overall
An economic model translates that relationship into a permanent productivity-equivalent figure. Because share prices incorporate expectations about future earnings, the result can include anticipated gains, anticipated costs and changing investor sentiment. It is therefore sensitive to the model’s structural assumptions as well as to market optimism or pessimism.
What you should not infer
- It does not show that every engineer became 32.6% faster.
- It does not identify a particular product, such as GitHub Copilot or Claude Code, as producing a guaranteed 32.6% improvement.
- It is not a controlled comparison of AI users and non-users completing the same coding tasks.
- It does not establish that a company’s realized productivity gain will match the figure priced into shares.
What randomized developer experiments found
A separate field-study program run across Microsoft, Accenture and an anonymous Fortune 100 company randomly gave some developers access to an AI coding assistant while others served as controls. Across 4,867 developers, the combined estimate was a 26.08% increase in completed tasks. The study reports a 10.3% standard error, and the size of the effect varied across the individual experiments.
Who appeared to benefit most
The effects were larger for less experienced developers. That pattern does not mean AI removes the need for senior engineers: experiment-level results were noisy, and the participating organizations and tools represent specific workplace conditions rather than every software team.
What “completed tasks” means here
The randomized study counts completed developer tasks under the tested workplace conditions. It is an operational outcome for the participating teams, not a valuation of all future software-engineering output. The result can include changes in task selection, workflow and the kinds of work developers finish, in addition to time spent writing code.
Why 32.6% and 26.08% are different numbers
| Dimension | Market-implied estimate | Randomized field evidence |
|---|---|---|
| Measurement | Stock-price responses mapped through an economic model | Completed tasks in randomized AI-assistant access experiments |
| Period and population | November 2022–December 2025; firms in the market sample described in the working-paper reports | 4,867 developers at Microsoft, Accenture and an anonymous Fortune 100 company |
| Outcome unit | Expected present value of software-engineering productivity | Completed developer tasks |
| Uncertainty | Depends on investor expectations and structural assumptions; the contemporaneous summaries do not give a full interval | Combined estimate reported with a 10.3% standard error; experiment-level effects varied |
| Interpretation | What investors priced into company values | What participating developers produced when given the tested assistant |
The percentages can sit in the same conversation without being a head-to-head test. One is an economy- and firm-level expectation over several years; the other is an average treatment effect on a defined workplace outcome. Comparing them as if they measured the same “speed boost” would overstate what either study proves.
How to use these findings without making a performance guarantee
For engineering leaders
Treat the market result as evidence that investors expect meaningful software-engineering value from AI, not as a budget assumption. Treat the randomized result as a reason to run a controlled pilot, not as a promise that your team will gain 26.08%.
- Define the outcome before enabling an assistant: for example, tasks completed, cycle time, review rework or incident-free releases.
- Record baseline performance and separate comparable work. A change in ticket mix can look like a productivity gain even when throughput is unchanged.
- Use a control group or staggered rollout where practical, so seasonal demand and process changes are not attributed to the tool.
- Track quality and safety alongside throughput, including escaped defects, rollback frequency, security findings and review time.
- Break results out by experience level and task type. The field studies found larger effects for less experienced developers, but that pattern may not hold for every codebase or assistant.
- Report uncertainty and the observation window with the result. A one-time pilot outcome is not the same as a permanent productivity change.
For individual developers
An assistant may help most with particular tasks—such as navigating an unfamiliar codebase, drafting routine code or generating tests—but the cited studies do not establish a universal gain for every workflow. Judge a tool by your own completion time, review burden and defect rate rather than by attaching the 32.6% headline to your personal output.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.The practical verdict
AI-assisted software development has credible evidence behind it, but the evidence supports a narrower conclusion than the headline suggests. The 32.6% figure is a forward-looking stock-market inference covering November 2022 to December 2025. The 26.08% figure is a randomized estimate of completed-task growth among 4,867 developers in three organizations, with a 10.3% standard error. Together they indicate substantial potential, while leaving the size of the realized gain dependent on the tool, team, task mix, experience level and measurement method.
Recommended Free Tools
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
