The news
The AI coding bottleneck has moved from writing code to checking it, according to Bain & Company's 2026 Global Technology Report, released on October 2. Bain found developers complete 21% more tasks with AI coding tools, but the time they spend reviewing those outputs has risen 91%, CIO Dive reported.
Bain surveyed nearly 300 senior technology leaders, according to CIO Dive. Developers are also juggling 47% more simultaneous workstreams. Leaders expect a 148% improvement in release-cycle speed and a 95% boost in developer productivity over the next one to two years, while the gains companies report today run between 20% and 27%.
Purna Doddapaneni, a Bain partner and author of the report, summed up the shift: "The bottleneck has moved from writing code to trusting it," as quoted by CIO Dive. She said every change still has to be understood, reviewed, tested and secured, and that most of that work still runs through people.
On October 5, GitHub, owned by Microsoft (MSFT), published ReviewBench, an open benchmark for judging AI code review agents. It contains 219 public pull requests, which are proposed code changes, drawn from 187 open-source repositories across 19 programming languages. GitHub said the benchmark measures how well review tools find real issues, with scores broken down by severity and category, and that the dataset, methods and an evaluation runner are publicly available.
The push to adapt came from computer scientist Peter Norvig as well. Speaking at The AI Conference in San Francisco on September 30, Norvig, a distinguished education fellow at Stanford's Institute for Human-Centered AI and former Google research director, argued that software engineering practices must change to suit increasingly capable coding agents, The Register reported.
The numbers
- More tasks completed with AI coding tools (Bain)
- 21%
- Increase in time spent reviewing outputs (Bain)
- 91%
- More simultaneous workstreams (Bain)
- 47%
- Tech leaders surveyed (Bain)
- Nearly 300
- Pull requests in GitHub ReviewBench
- 219
- Repositories / languages in ReviewBench
- 187 / 19
Why CEOs should care
For CTOs, the Bain data says buying more code-generation seats will not by itself speed delivery if review capacity stays flat. Ask your teams how long changes wait for review, how often AI-written changes are sent back, and where senior engineers spend their week. If review time is climbing, invest in automated review, testing and clear ownership of each change before adding more generation tools.
CFOs should check the gap between expectation and reality. Leaders in Bain's survey expect productivity gains of 95% within two years, against today's 20% to 27%. Budgets built on the higher figure risk disappointment. Tie AI tool spending to measured outcomes such as release frequency and defect rates rather than lines of code or tasks closed.
CISOs have a stake too, because review is where security flaws are caught. Norvig described a case in which code he wrote with OpenAI's Codex came with 6,000 temporary files that a colleague flagged, according to The Register. Benchmarks such as ReviewBench give security and engineering leaders a common way to compare AI reviewers before trusting them with that job.
The bigger picture
The software tools market is following the bottleneck. GitHub's decision to publish an open benchmark suggests AI code review is becoming a competitive product category, with vendors needing a shared yardstick. Bain's Doddapaneni said current workflows and team structures were designed to coordinate people, and that AI changes what needs coordinating, as quoted by CIO Dive.
Norvig, according to The Register, framed AI as another turning point like the moves from assembly to high-level languages, raising open questions about specifications, documentation, security, supply chains and how to monitor autonomous systems.
What’s next
Expect AI code review vendors to begin reporting ReviewBench results, giving buyers a way to compare tools. Engineering leaders planning 2027 budgets can use Bain's findings to rebalance spending from code generation toward review, testing and deployment capacity.
What “Fact-checked” means
Fact-checking means testing a story’s facts against the evidence before it is published. This story went through at least two separate checks before this version was published.
- What we checked
- Its names, figures, dates, job titles, quotes and who said what were checked against the story’s sources, including its main source where it could be opened. The headline was checked for accuracy and overstatement.
- How
- A first check reviewed the whole story. If it passed, a second, skeptical check went back to the sources to look for mistakes in the most important facts. If a check flagged the story, it was edited to fix the problems found, and a separate re-check then reviewed the whole story again.
- Who
- The checks are made by our newsroom, as steps kept separate from the writing, under rules set by our editor, Hussein Mukhtar. A story the checks still flag is held for the editor, who decides whether it is fixed, published or dropped.
- If something is wrong
- “Fact-checked” does not mean error-free. If a material error is found after publication, we correct the story and add a note saying what changed. Report an error
Companies in this story
