AI productivity gains: 10% or 10x? Three skeptical reads

The problem

Every AI tool pitch promises hours saved, and every buyer has to decide how much of that to believe. Three July 2026 pieces push back from different angles: a company analysis of engineering throughput across more than 400 firms, an essay arguing that faster output is not the same as economic productivity, and a personal story about a minimal setup beating an elaborate one. If you are deciding what to pay for or how much time to budget for adopting new tools, it helps to know which of these is measurement and which is opinion. This roundup sorts them by evidence type and says what each one can and cannot tell an individual worker.

Our take

These three sources are different kinds of evidence, and mixing them is how bad productivity claims spread. Only the first is a measurement: an analysis by DX of more than 400 companies reporting a median 7.76% rise in pull-request throughput and a mean of 13.1%, with the top decile near 44%. It is careful about its own limits, calling throughput an flawed proxy, and the author works at the company that ran the study. It also covers engineering organizations, not solo builders or non-developers. The second is a macro argument, partly built on comparison with earlier technology waves rather than current data. The third is a single anecdote about one very effective engineer. None of them tells you what a new tool will do to your week. Our read: plan for roughly 5 to 15% gains at team level, treat any promise of ten times as unsupported, and measure your own cycle time for a month before paying for anything.

The measurement: a large dataset with a narrow scope

Justin Reock, Deputy CTO at DX, reports in LeadDev that DX analyzed engineering velocity at more than 400 companies between November 2024 and February 2026, across technology, financial services, retail and healthcare firms with roughly 150 to 10,000 engineers. Over that time AI adoption rose 65%. Median pull-request throughput rose 7.76%, the mean 13.1%, and the 90th percentile approached 44%. His summary is that most organizations see a 5 to 15% throughput gain. He adds that coding is about 16% of how engineers spend their time and that PR throughput is the most useful, though flawed, proxy available.

The macro argument

Matt Scherer of the Open Markets Institute argues in Hard Reset that producing more output faster is different from raising economic productivity, which compares the value of outputs to inputs. He points to a reported $1.5 trillion of AI infrastructure spending and says app releases have doubled monthly while the number of apps with significant use has begun to fall. He concedes quality problems might be solved eventually and relies partly on the historical productivity paradox.

The anecdote

Alex Kotliarskyi describes a Facebook engineer who worked in a bare-bones editor with print-statement debugging and still shipped major features, while the author's heavily customized Vim setup did not make him faster. His point is that choosing the right problem matters more than tooling.

How to weigh them

Best for

• You are budgeting time or money for AI coding toolsPlan around a 5 to 15% throughput gain, not a multipleThe only dataset here reports a 7.76% median and 13.1% mean across 400+ companies.
• You are a solo builder or non-developerTrack your own cycle time for a month before and after a tool changeThe study covers engineering organizations, so it does not describe individuals.
• You are choosing between more tooling and a better problemSpend the first hour on problem selectionThe anecdote is one data point, but it matches the point that coding is only about 16% of engineers' time.