McKinsey and Cognition put AI coding returns to the test
A McKinsey survey of 300+ leaders found productivity more than doubled at 25% of organizations but fell at nearly 30%, as Devin and workflow redesign target AI returns.

By Sophie Lindqvist
In a McKinsey survey of more than 300 product and engineering leaders, 25% of organizations reported that AI had more than doubled productivity, while nearly 30% reported a decline.
That divergence is an organizational result, not a national labor-market rate. The survey provides no country, sector, wage or headcount breakdown. For an individual software engineer, it also shows why a headline about AI productivity cannot establish whether jobs, pay or team sizes will rise or fall.
What does the McKinsey-Cognition partnership cover?
Under a partnership announced Monday, McKinsey will advise clients on which engineering projects to target and how to redesign teams and workflows around AI. Cognition will provide Devin, which the company says can understand, plan, write and test code.
Devin launched in early 2024 and unsettled some software engineers who feared both workflow disruption and threats to their livelihoods. Since then, developers’ work across industries has shifted from writing code toward guiding AI output, although the source provides no layoff or replacement figure.
How much can AI coding agents do now?
Lari Hämäläinen, a senior partner in McKinsey’s technology, media and telecommunications practice, said coding tools became substantially more capable during the previous year. Until the final months of 2025, many remained most useful for narrow tasks such as finding technical guidance or syntax details.
More capable models have since allowed companies to hand off longer, more autonomous coding work. Hämäläinen said: “This is really a step-change phenomenon” that “requires a complete rethink” of software and product development. He also warned that technology might improve faster than organizations can redesign their processes.
Can the productivity gains survive a large rollout?
A small team can experiment with an AI coding tool, but Hämäläinen said scaling it across 2,000 or 20,000 developers presents a different problem. The survey’s sharp split suggests that successful experimentation does not automatically produce organization-wide gains.
The source identifies only the sample as more than 300 product and engineering leaders. It gives no exact count, country mix, field dates, weighting method or revision schedule, so the findings cannot show whether the productivity gains cluster by industry, company size or implementation method.
How do the partners intend to control cost?
Efforts to quantify AI returns and manage their expense have led Goldman Sachs, Cisco and JPMorgan to restrict access to particular tools or tighten how employees use them, according to the source.
Art Levy, Cognition’s vice president of global partnerships, said Devin can draw on multiple AI models. Companies can use more expensive frontier models for difficult engineering work and cheaper models for other tasks. His analogy captured the proposition: “You don’t need to drive a Ferrari to the grocery store.”
AI routing tools have become popular for managing those costs by assigning tasks according to each model’s capabilities and price. Levy said Devin can also help with large modernization projects that organizations had postponed. At one client, an enterprise resource planning update expected to take five years is being compressed into 18 months.
Levy’s measure of success now includes cost as well as completion: buyers ask not only whether an agent can finish the job, but whether it can do so economically. The 18-month project offers one example, though the source does not establish how representative it is.
What does this mean for software engineers?
The evidence supports a narrower conclusion than the partnership’s ambition. It documents changing software work and uneven organizational productivity, not measured displacement of engineers. For an individual, guiding AI output now belongs to the job; the source supplies no layoff count, wage movement, hiring total or replacement rate with which to judge employment risk.
Large-company restrictions also show that AI access will not be uniform. Tool selection, model routing and workflow redesign will determine whether employers can turn agent capability into sustained productivity. The next test is whether the 25% gain can be reproduced beyond pilot teams at organizations with 2,000 or 20,000 developers while keeping model costs low enough for buyers to accept.
Source: feeds.businessinsider.com
More from Sophie Lindqvist
Show full bio
Staff writer covering regulation, policy and their impact on Job market data and labor economics at Path & Pay.
583 articles
Go deeper
- JPMorgan's $2,000 AI Cap Signals the End of Unlimited Work Tokens
- 96% of Firms Invest in AI, Only Half Report AI-Literate Staff
- AI Could Push 11 Million US Workers Into New Careers by 2035
- 1.9 Million US Coders, But the Job Now Runs on Meetings
- TD Economics: AI Reshapes Jobs Rather Than Eliminating Them