McKinsey and Cognition partner to prove AI coding ROI
The partnership aims to help large companies move past AI experimentation to see real productivity gains in software engineering.

The short version
- McKinsey will advise clients on which engineering projects to target for AI integration.
- Cognition's Devin agent can independently understand, plan, write, and test code.
- One client used Devin to compress an enterprise software update from 5 years to 18 months.
- A McKinsey survey found that while the top quarter of organizations saw productivity more than double, nearly 30% saw it drop.
Why it matters
For businesses, the challenge is moving from small-scale AI experiments to large-scale deployment. This partnership highlights a shift from just 'using' AI to fundamentally redesigning how companies build software.
Proving the AI payoff
McKinsey & Company and the AI lab Cognition have announced a partnership to demonstrate the tangible value of AI in software engineering. While many companies have experimented with AI tools, McKinsey wants to help clients identify specific projects where these tools can provide the most impact. The firm will advise on which engineering tasks to target and how to restructure teams to work alongside new technology.
Cognition will provide its AI coding agent, Devin, as the primary tool for these initiatives. Unlike basic autocomplete features, Devin is designed to understand, plan, write, and test code autonomously. The partnership seeks to move beyond the hype of AI capabilities to show how these tools can actually improve the bottom line for large organizations.
This collaboration comes as many large corporations, including Goldman Sachs and JPMorgan, become more selective about AI access. These companies are looking for ways to manage the high costs of AI while ensuring the technology produces measurable results. The goal is to move from asking if AI can do the work to asking if it can do the work efficiently and affordably.
From syntax to autonomy
The capabilities of AI coding tools have shifted dramatically over the last year. According to Lari Hämäläinen of McKinsey, early AI tools were mostly useful for narrow tasks like looking up technical guidance or specific syntax details. These were simple interactions where the human did the heavy lifting and the AI provided small pieces of information.
More capable models have now made it possible to hand off longer, more autonomous coding tasks to the AI. This shift represents what Hämäläinen calls a step-change phenomenon in the industry. Because the technology is advancing so quickly, it can sometimes outpace the ability of organizations to redesign their internal processes to accommodate it.
For many developers, this change means a shift in their daily roles. Instead of writing every line of code by hand, engineers are increasingly moving toward guiding the output of AI systems. This transition requires a complete rethink of how software and products are developed from the ground up, rather than just adding a new tool to an existing workflow.
The challenge of scaling
Adopting AI at scale presents a significant hurdle for large enterprises with thousands of developers. A recent McKinsey survey of over 300 product and engineering leaders revealed a wide range of outcomes. While the top quarter of organizations saw their productivity more than double, nearly 30% of the participants reported that productivity actually dropped.
One common pitfall identified by McKinsey is the difficulty of moving from a small team experiment to a company-wide rollout. It is relatively straightforward to let a few people play with an AI tool, but it is a completely different challenge to integrate it across 2,000 or 20,000 developers. Scaling requires consistent workflows, security measures, and clear guidelines that many companies have not yet established.
To address these issues, companies are looking for ways to manage the costs associated with high-end AI models. Cognition's Devin can draw on multiple different AI models to handle various tasks. This allows companies to use expensive frontier tools for difficult engineering problems while using cheaper models for simpler tasks, a strategy known as AI routing.
Compressing project timelines
One of the most compelling arguments for AI coding is the ability to accelerate massive software modernization projects. These are often the types of projects that large companies put off because they are too complex or time-consuming. Cognition points to a specific client where work to update enterprise resource planning software was expected to take five years.
By using Devin, that same project is being compressed into just 18 months. This represents a significant leap in speed for updating the core systems that many large businesses rely on daily. It demonstrates that AI can handle the heavy lifting of large-scale updates that were previously considered too daunting to tackle quickly.
Art Levy of Cognition noted that the focus for companies is shifting from whether the job can be done to how cheaply it can be completed. By using AI routing to balance cost and capability, firms hope to modernize legacy systems without breaking their budgets. This approach aims to make software updates a routine part of business operations rather than a multi-year burden.
What to watch
- How many other large enterprises successfully replicate the 18-month compression timeline.
- Whether AI routing becomes the standard architecture for enterprise AI spending.
Sources and links
First reported by Business Insider. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.