Delivery
AI coding tools: the typing was never the expensive part
What the 2025–26 evidence on AI coding tools shows, and the three controls to have in place before your change volume rises.
Computing America. Applications, integrations, data analysis and AI/ML engineering.
Send the build
In short
- In DORA’s 2025 data, AI adoption correlates positively with delivery throughput and negatively with delivery stability.
- Veracode’s 2025 tests across more than 100 models found 45% of test cases introduced an OWASP Top 10 vulnerability, and that rate stayed flat while the same models got measurably better at functional correctness.
- GitClear’s commit-corpus analysis finds the share of changes that touch code last modified more than twelve months ago fell from 1.7% in 2023 to 0.46% year-to-date in 2026.
- Most developers use these tools or plan to. The buyer's decision is whether tests that can fail a build, review by someone who didn't write the code, and a way back exist before change volume rises.
- METR’s 2025 trial found experienced developers 19% slower with AI, and its authors limit what that shows. In February 2026 METR reported selection effects in its late-2025 follow-up and said it is redesigning that study. We don't use either as an estimate of today's effect.
The studies we rely on below look at delivery and quality outcomes, and that is where they find the trouble. In DORA’s 2025 survey, AI adoption went with higher delivery throughput and lower delivery stability. Veracode found security stayed flat while correctness improved, and GitClear found the share of changes that touch old code fell. That matters most if you're deciding whether to buy engineering help.
In Stack Overflow’s 2025 survey (opens in a new tab), 84% of developers use or plan to use these tools, and the most-cited frustration, at 66%, is output that’s almost right but not quite. The question is what happens to that output after it’s written.
Higher throughput, lower stability¶
DORA’s 2025 report on AI-assisted software development (opens in a new tab) surveyed nearly 5,000 technology professionals. 90% use AI at work and more than 80% believe it has made them more productive. Unlike the previous year, it finds a positive relationship between AI adoption and both delivery throughput and product performance. It finds, again, a negative relationship with delivery stability.
“AI doesn’t fix a team; it amplifies what’s already there.”
A number we won't rely on¶
METR’s randomized controlled trial (opens in a new tab) is a widely cited figure in this debate: sixteen experienced open-source developers, 246 real tasks, 19% slower with AI. We won't use it as an estimate of today's effect, for the authors’ own reasons. The paper says it isn't evidence that AI fails to speed up most developers, that it doesn't generalize past experienced developers on code they know, and that it's a historical result.
task time with AI, as % of time without it
- Without AI100%
- Forecast before the work76%
- Self-estimate after the work80%
- Measured119%
Measured task time with AI was 119% of the time without it, while developers believed 80%; a 2025 result this article does not use as today's estimate.
SourceMETR, early-2025 study of 16 developers on 246 issues (arXiv 2507.09089): completion time reduced 24% (forecast) and 20% (self-estimate), increased 19% (measured).
In February 2026 METR reported selection effects in its follow-up study (opens in a new tab). That study began in August 2025 with 57 developers. METR cites “a significant increase in developers choosing not to participate in the study because they do not wish to work without AI.” It says the data gives “an unreliable signal of the current productivity effect of AI tools.” The post does not retract the 19% result. METR writes that “it is likely that developers are more sped up from AI tools now,” but that its data “is only very weak evidence for the size of this increase.”
Security didn't improve with capability¶
Veracode tested 80 coding tasks across more than 100 models (opens in a new tab) in Java, JavaScript, Python and C#. In 45% of test cases, the generated code introduced an OWASP Top 10 vulnerability. Java had a security failure rate over 70 percent, cross-site scripting went undefended in 86% of the cases where it was relevant, and log injection in 88%. Across larger and more sophisticated models, functional correctness improved and security stayed flat.
Veracode sells application security testing, so trust the shape of the result more than the exact figures. A working demo shows that the code works. It can't show the failure modes that carry contractual or regulatory weight.
Edits to old code are falling¶
GitClear’s analysis of 623 million code changes (opens in a new tab) measures the share of changes that remove or update code last touched more than twelve months ago. It fell from 1.7% in 2023 to 0.46% year-to-date in 2026, which is consistent with developers writing new code beside the old and leaving the old alone. This is correlation across a commit corpus with no controlled trial behind it, GitClear sells developer analytics, and its headline refactoring figure is quoted against two baselines in its own materials. Read it as a direction, not a measurement of your codebase.
Three controls to have before the volume rises¶
- 1.A test suite that can fail a build. A suite that only reports can be overridden under deadline.
- 2.Review by someone who didn’t write it, which now means someone who didn’t prompt it. Accepting output is faster than writing it, so more of it arrives.
- 3.A way back: a release you can undo without a meeting.
Keep one more thing: the reason. A decision record, a pull-request comment or three lines in a file will do. In a year you'll be able to read what the code does, and you'll need to know what it was supposed to do.
What it changes about buying¶
For a small internal tool (one screen, a few people in one building, nothing regulated, no outsider depending on the answer), build it yourself. That was true before these tools.
The boundary isn't size or sophistication. It's the first moment someone outside the team that built it depends on the answer, or the first time being wrong costs money. Past that line you're buying review, tests, evidence and accountability as much as code.
git log -50 --onelinegit show <commit>git blame -L <start>,<end> <commit>^ -- <file>Sources
- 1.State of AI-Assisted Software Development 2025 (opens in a new tab), DORA (Google Cloud),
- 2.2025 GenAI Code Security Report (opens in a new tab), Veracode,
- 3.The Maintainability Gap: 2026 AI code quality research (opens in a new tab), GitClear
- 4.Measuring the impact of early-2025 AI on experienced open-source developer productivity (opens in a new tab), METR,
- 5.We are changing our developer productivity experiment design (opens in a new tab), METR,
- 6.AI (opens in a new tab), Stack Overflow Developer Survey 2025
Next step
Send us the thing you already built.
Tell us what it does and who depends on it now. We will name which of the three downstream controls it is missing (tests that can fail a build, review by someone who did not write it, or a way back) and say whether that is worth fixing before it grows.
- Phone
- (214) 723-2510
- Reply
- A person replies, not a sequence: within one business day, from someone who would be on the engagement.
Packages
Start with a package
Four packages, each with its price on the page. Pick one and buy it online.
$199/month
Billed monthly
$182.42/month
Billed yearly
$499/month
Billed monthly
$457.42/month
Billed yearly
- Scammers stopped from sending email as you
- Your website and systems scanned every week
- Your email settings checked every week
- Everything in the 1 to 10 plan
- Covers up to 50 people
$199/month
after a one-time setup fee
- A site planned, designed and built for you
- Hosting and updates every month
$499/month
Billed monthly
$457.42/month
Billed yearly
$999/month
Billed monthly
$915.75/month
Billed yearly
- Your Google profile, fully set up
- A Google post every week
- 2 social posts a week
- Everything in Get found
- 4 social posts a week
- Your top 10 Google searches tracked
from$500/month
after a one-time setup fee
- Your data gathered from the tools you use
- A first analysis that answers your questions
- New findings every month
Related reading
AI3 min readWhich AI projects are worth funding: the three conditions
What to check before you fund an applied AI project, and what to build before the model.
Strategy3 min readLow-code limits: where the ceiling is, and the signs you are near it
Four constraints set a low-code platform's ceiling, and the signs you are near it are visible.
Delivery3 min readHow to evaluate a software vendor without reading code
Four requests that show whether a firm builds, tests, tracks its dependencies and learns from failure.
