AI now writes 42% of the code developers commit, according to Sonar's January 2026 survey of more than 1,100 professional developers, and 96% of them don't fully trust it. That gap explains most of what's happening in AI-assisted software development in 2026: the tools are everywhere, the speed gains are harder to prove than the marketing suggests, and the hard work has shifted from writing code to checking it.
Where AI-assisted software development stands in 2026
Almost every team uses it already. Google's 2025 DORA report, published in September 2025 from a survey of nearly 5,000 technology professionals, put adoption at 90%, with a median of about two hours a day spent working with AI. Roughly two thirds said they rely on it heavily.
Use is also routine rather than experimental. Among developers who have tried AI coding tools, 72% use them every day, per the same Sonar survey.
Both numbers are self-reported, and Sonar sells code-verification tools, so read its framing with that in mind. Still, the direction is not in dispute. The open question is what all that usage buys.
Does AI actually make developers faster?
Sometimes, but the best controlled evidence is less flattering than the surveys. In a randomized trial METR published in July 2025, 16 experienced open-source developers worked through 246 real issues in repositories they knew well, with AI randomly allowed or banned on each one. Issues done with AI took 19% longer.
The issues were real work, averaging about two hours each, in repositories that averaged over a million lines of code. That makes the result harder to dismiss as a lab artefact, and it also makes it narrow, since it says nothing about newcomers or small projects.
The perception gap is the more interesting part. Before starting, the developers forecast a 24% speedup, and afterwards they still believed AI had made them about 20% faster.
The authors are careful about what that shows. The sample was small, the tools were early-2025 ones (mostly Cursor Pro with Claude 3.5 and 3.7 Sonnet), and they don't claim AI slows most developers. METR's page now marks the results as out of date.
The DORA report points the other way: more than 80% of respondents said AI improved their productivity, and AI adoption is now linked to higher delivery throughput, reversing the previous year's finding. Put the two together and you get a sensible rule. Developers' feelings about their own speed are a poor measurement, so measure your own delivery times before and after.
Which tasks AI handles well
Tasks where the AI explains or documents existing work come out ahead of tasks where it writes new logic. Sonar asked developers how effective AI was for each job, and the ranking is steady enough to plan around.
| Task | Share of users rating AI effective | What it tells you |
|---|---|---|
| Documentation | 74% | Best-rated use in the survey |
| Explaining existing code | 66% | Useful for onboarding and legacy systems |
| Generating tests | 59% | Good payoff because the output is easy to run and check |
| New code development | 55% | Used by 90%, yet barely over half rate it highly |
| Refactoring | 43% | Used by 72%, with the weakest ratings |
Note that the new-code and refactoring figures count only users who rated it extremely or very effective, so they are a stricter bar than the rest. Even allowing for that, the most popular use is not the best-rated one.
Start with documentation, explanations and tests. Hand over the risky logic last.
Why reviewing AI code is the slow part now
When a tool produces code faster than a person can read it, the review queue becomes the constraint. Only 48% of developers always verify AI-generated code before committing it, and more than a third say reviewing it takes more effort than reviewing a colleague's work.
Google's researchers call the pattern a trust paradox: people find AI useful without fully trusting what it hands back, which suggests they treat it as support for their judgement rather than a replacement for it. That's the healthy reading. The unhealthy one is that fewer than half of developers always run the check, which leaves AI-written code in production that no human has read closely.
The cost of skipped review doesn't show up on the day you merge. It shows up weeks later, in a bug nobody on the team remembers writing.
What separates teams that benefit from teams that don't
Photo by Markus Winkler on Pexels
Process quality decides it. DORA describes AI as a mirror and a multiplier: in cohesive organizations it lifts efficiency, and in fragmented ones it exposes the weak points. The report sorts teams into seven archetypes, from "harmonious" high achievers to teams stuck behind a legacy bottleneck.
Google also introduced a model of seven capabilities that mix technical and cultural factors, meaning the gains depend on practices around the tool as much as the tool.
The mechanism is easy to see. If releases already queue up behind a slow test suite or a manual sign-off, faster code generation just piles more work in front of the same gate. Teams building AI into the product itself, rather than only using it to write code, are in the territory our AI and automation services cover.
How to bring AI into a project without losing control
A staged rollout beats a blanket mandate. These four steps follow from the evidence above, and they work whether you build in-house or with a partner on an app or SaaS product.
- Begin with documentation, code explanation and tests, the three jobs developers rate highest.
- Pair every AI-assisted change with automated quality and security checks, so review doesn't depend on someone's memory.
- Record cycle time and defect counts before you start, then compare. METR's trial shows why a feeling of speed isn't evidence.
- Clear the slowest step in your delivery pipeline first, because that's where the extra output will wait.
Keep the human decision on architecture and on anything customers touch. The developers in Sonar's survey use AI for prototypes far more than for business-critical services, and that ordering is a reasonable instinct to copy.
Where vibe coding fits
Vibe coding is the loosest end of the spectrum: describing what you want and letting the model build most of it. It suits prototypes, which matches where developers say they use AI most.
We've written about how vibe coding speeds up an MVP and about how it compares with traditional development. The short version is that the looser the process, the more the verification step above matters.
Frequently asked questions
Is AI-assisted development used for business-critical software?
Yes, though less often than for low-stakes work. In the Sonar survey, 88% of developers use AI for prototypes, 73% for customer-facing applications and 58% for business- or mission-critical services. That shows usage, not safety, so the higher the stakes, the more review the code needs.
Is the 19% slowdown still the latest evidence?
No. METR's own page says the July 2025 results are out of date and points to a follow-up study from February 2026. Treat the 19% figure as a snapshot of early-2025 tools in one setting, not a verdict on AI today.
Do developers use AI every day?
Most of those who have tried it do. Sonar found that 72% of developers who've tried AI coding tools use them daily, while DORA measured a median of about two hours a day working with AI.
Cover photo by Matheus Bertelli on Pexels
