Microsoft study finds developers spend just 14% of time coding

An ACM Queue essay, drawing on large-scale studies, interviews and field observations, sets out to correct eight myths about generative AI's effect on software engineering, arguing that marketing claims, anecdotal wins and misread research have outpaced the evidence and are steering AI adoption and measurement decisions the wrong way. The retrieved text covers four of those myths in detail before cutting off mid-sentence.

The first myth: that developers spend most of their day writing code. A 2025 study of more than 450 Microsoft engineers found they spend just 14 percent of their time writing code, in line with older findings that put coding time at 18 percent on a "good" workday and 11 percent on a "bad" one. Because so little of the job is typing in the editor, the article argues an AI assist that made coding twice as fast would, in theory, raise overall productivity by less than 15 percent, since the other 85 percent of a developer's time (design, meetings, reviews, legacy code, environment setup) remains untouched. A developer interviewed in a June 2025 Microsoft study made the same point directly: "coding is [one] aspect, but a lot of time is also spent on design and meetings," and "the number of points in my job work that are even touched by GitHub Copilot are relatively small."

The second myth: that lines of code, including AI-generated lines of code, a figure that companies including Microsoft have reported publicly, is a valid productivity metric. A 2014 study, "A statistical study of the relevance of lines of code measures in software projects," concluded the metric "fails to meet the specified validity tests and, therefore, has limited utility." The essay argues the metric persists anyway, and can push teams to game the numbers and prioritize coding volume over design quality, growing technical debt. It cites Bill Gates: "Measuring software productivity by lines of code is like measuring progress on an airplane by how much it weighs."

The third myth: that GenAI's effect on developer productivity is straightforward rather than highly context dependent. Research results here are mixed: some studies find large productivity gains, some find neutral effects, and a 2025 study of experienced open-source developers found AI tools increased implementation time by 18 percent on average. The 2024 Microsoft Report on AI and Productivity Research found bigger Copilot gains on familiar, well-understood tasks than on unfamiliar ones, and found that more years of professional development experience correlate with less confidence in writing effective prompts. One cited study found that rewriting a prompt while keeping its meaning unchanged produced different generated code in 46 percent of cases and changed code correctness in 28 percent of cases.

The fourth myth: that AI tools turn individual developers into "10x developers." A widely cited "55 percent productivity gain" figure is, per the article, context dependent, since controlled studies on isolated tasks do not capture the coordination and collaboration that team-based software delivery requires. The article also cites Georgetown computer science professor Cal Newport, quoted from a New Yorker piece, arguing that past productivity gains came from organization-level system redesign, not individual workers reoptimizing their own workflows: unlike Ford's assembly line, which took heavy investment and years of iteration, companies are now asking individual knowledge workers to do similarly complex optimization themselves while still doing their existing jobs.

Key facts

A 2025 study of more than 450 Microsoft engineers found developers spend just 14 percent of their time writing code.
An earlier study found coding time ranges from 11 percent on a "bad" workday to 18 percent on a "good" one.
A 2014 study found lines-of-code metrics "fail to meet the specified validity tests," yet Microsoft and others still publicly report AI-generated lines of code as a metric.
A 2025 study of experienced open-source developers found AI tools increased implementation time by 18 percent on average, one of several studies with mixed results on GenAI's productivity effect.
Rewriting a prompt while keeping its meaning unchanged produced different generated code in 46 percent of cases and changed code correctness in 28 percent of cases.

Why it matters

Organizations are spending real budget on GenAI coding tools and on metrics like AI-generated lines of code to justify that spend. The article argues both the promised productivity gains and the metrics used to track them are shakier than the marketing suggests, and says organizations have invested millions in licenses without a clear plan for capturing value from them.

Who it affects

Practitioners, team leads and engineering leaders who decide how to roll out and measure GenAI coding tools. The article frames itself explicitly as evidence for that audience rather than as general skepticism about AI.

How to use it

Do not judge AI adoption by lines of code generated: a 2014 validity study found the metric fails standard statistical tests, and gaming it can push teams toward more technical debt and lower design quality. Expect the biggest Copilot gains on familiar, well-understood tasks rather than unfamiliar ones, per the 2024 Microsoft Report on AI and Productivity Research, and expect real productivity gains to require organization-level workflow changes, not just handing licenses to individual engineers.

How solid is it

The piece leans on named, dated studies rather than anecdote: a 2025 survey of more than 450 Microsoft engineers (14 percent coding time), an earlier study splitting good and bad workdays (18 percent and 11 percent), a June 2025 qualitative study of Microsoft developers, a 2014 lines-of-code validity study, the 2024 Microsoft Report on AI and Productivity Research, a 2025 study of open-source developers, and a study on prompt-rewrite sensitivity. The retrieved text cuts off mid-sentence before finishing the fourth myth and before the remaining four of the eight promised myths are presented, so the article's fuller argument is not verifiable from this excerpt.

Risks and caveats

The evidence itself is mixed rather than uniformly damning: the article notes some studies find large GenAI productivity gains, others find neutral effects, and one 2025 study of open-source developers found AI tools increased implementation time by 18 percent on average. A widely cited "55 percent productivity gain" figure is flagged as context dependent, but the retrieved text does not restate what task or setting produced it. The developer quoted throughout the first section is not named or given a title.

“Measuring software productivity by lines of code is like measuring progress on an airplane by how much it weighs.”

— Bill Gates