Cognition and McKinsey Will Measure Devin's Returns: AI Failed to Boost Output for 30% of Teams
Cognition, the maker of the Devin AI agent, has formed an alliance with consultancy McKinsey to help companies see the return on AI in software development in hard numbers. In a survey of 334 technology leaders, the top quarter of organizations more than doubled productivity, while 30% saw no change or a decline. Devin takes on engineering tasks, while McKinsey helps choose projects and measure cost, speed, and quality. That is why adopting the agent should start with a baseline measurement.
AI-processed from Windsurf Blog; edited by Hamidun News
Nearly every development team now uses AI, but the returns vary widely. Cognition, the company behind Devin, an AI engineer that takes over part of the programming work, has formed an alliance with the consultancy McKinsey. The trigger was a McKinsey survey: in the top quarter of companies, development productivity rose by more than 2x, while 30% saw no change or even a decline. The idea behind the alliance is simple: Devin does the engineering work, while McKinsey helps choose projects, measure results, and restructure processes so that AI spending turns into finished projects and clear numbers in reporting.
Why AI Has Such Uneven Results
The McKinsey survey, conducted in the second quarter of 2026, included 334 leaders of product and engineering teams. The top quarter of organizations reported productivity gains of more than 2x. Meanwhile, 30% of respondents said productivity stayed the same or fell.
According to McKinsey, the difference increasingly comes down not to the choice of tool but to whether the company has restructured its engineering organization around AI. It is like buying professional kitchen equipment: if the chefs still work the old way, the stove will not make them cook faster. Roles, the order of tasks, and the way results are accepted all need to change.
Who Is Responsible for What
Devin works on several engineering projects at once. It connects to the security systems and processes the company already has, and returns finished work to engineers for review. People set the direction, solve the hardest problems, and decide what to release.
McKinsey takes on the organizational side:
- it helps select the projects that matter most to the business;
- it records baseline metrics for cost, speed, and quality so there is something to compare against later;
- it changes team structure, roles, and training, from junior engineers to senior executives.
Cognition stresses separately that Devin is not tied to a single model: it selects different models for different types of tasks. According to the company, this can deliver quality on par with leading models at up to 60% lower cost, and makes it faster to switch to new models when they are released. This is the company's own claim; the material contains no independent verification.
Regulated industries are showing interest in the solution. According to Cognition, it already works with six of the eight global systemically important banks in the US and three of the six largest American health insurers. McKinsey says it has helped more than 50 companies restructure around AI-driven development. In its Rewired study of 20 large transformations, companies that focused on one to three high-value areas, such as software development, achieved on average about a 20% increase in EBITDA (operating profit before taxes and amortization) and reached payback within one to two years.
Example: A 20-Year-Old ERP in 18 Months
The first illustrative client is Comarch, one of the oldest technology companies in Central Europe. It is using Devin to modernize ERP XL, its 20-year-old enterprise management system, and has set an 18-month timeline. In the words of CTO Tomasz Stachlewski, asking clients to wait five years was not an option. Devin helps engineers understand the legacy system and ship new features faster. In parallel, McKinsey is helping Comarch embed Devin into its workflows, extend its use to other teams, and measure the impact on the business.
What This Means for Practitioners
First, the alliance shows where the market for AI coding agents is heading. Selling the agent alone is no longer enough: enterprise buyers want to know what was delivered, what it would have cost without AI, and what a completed project actually yields. According to Cognition co-founder Russell Kaplan, these three questions are coming from customers.
Second, if you are introducing AI-driven development in your own organization, start with measurement. Record cost, speed, and quality before the agent goes live, or you will have no way to prove the effect later. Choose one to three projects with clear business value rather than handing the tool to everyone at once.
There are caveats. The material was published on Cognition's own blog and is a partner announcement, not an independent study. The figures on savings of up to 60% and the EBITDA gains are claims made by the parties involved and do not guarantee the same results for you. The terms of the partnership and its cost have not been disclosed; interested companies are directed to Cognition's enterprise team. Still, the direction is clear: those who change how their teams work as boldly as they adopt the technology will come out ahead.
- Nov 21, 202615 Claude Code Subagents Built a Shop-Floor System, but a Release Gate Agent Without Edit Rights Approves It
- Nov 21, 2026Coding Agents Built an LLM Engine for AWS FPGAs: 213 Tokens/s and 38% Cheaper Than CPU
- Nov 21, 2026Empatra Studio 0.29 Cuts Frontend Agent Costs by 59% and Runs Twice as Fast
- Nov 21, 2026Ego lite Gives Each AI Agent Its Own Workspace Space in the Browser
Key Questions Answered
Why AI Has Such Uneven Results?
The McKinsey survey, conducted in the second quarter of 2026, included 334 leaders of product and engineering teams. The top quarter of organizations reported productivity gains of more than 2x. Meanwhile, 30% of respondents said productivity stayed the same or fell. According to McKinsey, the difference increasingly comes down not to the choice of tool but to whether the company has restructured its engineering organization around AI.
Who Is Responsible for What?
Devin works on several engineering projects at once. It connects to the security systems and processes the company already has, and returns finished work to engineers for review. People set the direction, solve the hardest problems, and decide what to release.
What are the key details on Example: A 20-Year-Old ERP in 18 Months?
The first illustrative client is Comarch, one of the oldest technology companies in Central Europe. It is using Devin to modernize ERP XL, its 20-year-old enterprise management system, and has set an 18-month timeline. In the words of CTO Tomasz Stachlewski, asking clients to wait five years was not an option. Devin helps engineers understand the legacy system and ship new features faster.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.