# McKinsey: Top 25% Get 2x From AI Coding _McKinsey told AI Engineer the top 25% of firms get 2x gains by rebuilding sprints, teams and context, not by adding tools._ **Published:** 2026-11-09 **Source:** https://www.startuphub.ai/ai-news/artificial-intelligence/2026/mckinsey-top-25-get-2x-from-ai-coding --- Martin Harrysson and Prakhar Dixit told [AI Engineer](https://www.youtube.com/watch?v=21tuT-e4NVo) that a top 20 to 25% of large enterprises are now seeing 2x end-to-end gains from AI in software work, while most firms still report only 10 to 20% faster coding. Tools in the same process do not work. The PDLC, the software development life cycle that carries an idea from requirements through build and test to release, is what changed. Harrysson said many clients tried to plug agents into the old flow and stalled. Dixit said winners are dropping the two-week scrum sprint for one-day and two-day Kanban loops where agents run the night shift and humans validate, then tighten or loosen human review as confidence grows. Those loops started in build and test and are now spreading to product definition and release, which only works if artifacts like PRDs and specs are standardized so agents get the same shape of context every time. Teams are getting smaller and roles are collapsing. Dixit said he sees pods of two to four instead of large scrum teams, with product work moving toward a definer role and dev plus QA moving toward a builder role. On tech, he called tooling solved. The edge is the semantic layer, the curated company knowledge that tells agents what good means in this codebase, because too much context still lives in the heads of longtime experts. That is why curation now beats availability. One client found 1,700 downloadable skills for writing a single product requirement document and froze on choice. Speed has created a quality problem. A third McKinsey speaker said defects are rising even on a per-developer and per-PR basis as commits accelerate, so review skill has to grow as fast as generation speed. Her fix is golden examples, hand-picked samples that define good code, plus policies for maintainability and merge safety fed to agents at each checkpoint rather than checked at the end. Once build and test compress, she said, the bottleneck moves upstream to requirements, downstream to release, and sideways to risk, compliance and design. Her example was a bank app where button colors had to pass accessibility rules that no coding agent would know unless encoded. Measurement is still thin. Harrysson said few firms feel confident they track real output against token, tool and retraining cost, and he pointed to AI FinOps and to using AI to score which work actually drives value. His near-term cheat is simple arithmetic. Halve pod size, double team count, and leaders get fast acceleration before they need a better model. That matches outside engineering notes that in an AI-native SDLC, per the Forbes Tech Council piece, writing code by hand becomes an exception while engineers spend more time on interfaces, constraints and architecture. Watch whether your team can name its golden examples and its human checkpoints. If not, you are buying speed without control. --- Original analysis from [startuphub.ai](https://www.startuphub.ai), the #1 AI startup directory. © StartupHub.ai. All rights reserved. You may not republish this article in full without a license. Search engines and AI research tools may crawl and summarize for reference. Bulk reproduction or model training on this content requires a license. See https://www.startuphub.ai/terms.