McKinsey and Cognition test AI coding’s uneven returns at scale
More than 300 leaders found AI had more than doubled productivity at a quarter of organizations but reduced it at nearly 30%, as McKinsey and Cognition tackle redesign for scaled coding agents.

Daniel Okafor
· 4 min read
By Daniel Okafor
The top quarter of organizations in a McKinsey survey of more than 300 product and engineering leaders said AI had more than doubled productivity; nearly 30% said productivity had fallen. On Monday, McKinsey & Company and AI lab Cognition announced a partnership to show that AI coding can justify its cost.
McKinsey will advise clients on which engineering projects to target and how to redesign teams and workflows around AI. Cognition will provide Devin, an AI coding agent (software that carries out coding tasks) that the company says can understand, plan, write and test code.
For employers, the partnership turns this into choices about work allocation, team structure, access and measurement. These are people-process decisions, not merely software purchases.
What changed in AI coding?
Devin’s early-2024 launch set off jitters among some software engineers over workflow disruption and possible effects on their livelihoods. In the years since, developers across industries have shifted from writing code to guiding AI output. The work has not disappeared so much as changed shape.
Lari Hämäläinen, a senior partner in McKinsey’s technology, media and telecommunications practice, said some of the biggest changes for developers had arrived in the past year. Until the final months of 2025, many AI coding tools were most useful for narrow jobs such as retrieving technical guidance or checking syntax. He described their usefulness as little more than that.
Since then, more capable models have allowed companies to hand off longer, more autonomous coding assignments, meaning more work can proceed with less hands-on intervention. Hämäläinen said: “This is really a step-change phenomenon” that “requires a complete rethink” of software- and product-development processes. Technology can improve faster than an organization can redesign its processes around it.
That gap explains the survey’s split: a successful pilot does not establish an enterprise operating model. Hämäläinen said scaling a tool across 2,000 or 20,000 developers was fundamentally different from letting a small group try it. At that scale, access, workflow integration and cost control become management decisions rather than pilot details.
For HR and line managers, the people-process work is to define how people and agents share tasks, establish baselines before measuring AI-assisted output, and revise performance expectations as coding roles shift toward guiding AI output. The changes belong in team design and performance management, not in a procurement note.
Employers also need to decide which projects receive access first; the survey does not support treating a successful pilot as a scaled result.
How can employers control costs and prove returns?
The search for measurable AI returns and manageable costs has already prompted Goldman Sachs, Cisco and JPMorgan to become more selective about who receives which tools and how employees use them. Their response exposes an uncomfortable procurement truth: a capable tool can still be too expensive to deploy indiscriminately.
Devin can draw on multiple AI models, according to Cognition. That lets a company reserve costly frontier models (high-end AI systems) for difficult engineering work and use cheaper models elsewhere. Levy used a blunt comparison: driving a Ferrari is not necessary for a grocery-store trip.
The point is not that every task deserves a premium model; each task needs an economic match.
AI routing tools (systems that assign work to models according to capability and cost) have become a popular way for companies to control spending. The decision now carries two tests: can the selected model complete the task at the required quality, and what will that completion cost?
For employers adding coding agents to large technical organizations, cost governance is not an IT afterthought. It determines whether the apparent productivity gain survives at scale.
Levy cited one client where an update to enterprise resource planning software (ERP, systems used to run finance, operations and records) was expected to take five years. Under the new approach, that work is being compressed into 18 months.
The example is a schedule claim from one client, not a survey average, but it shows how model selection can affect modernization timetables and team capacity.
“It’s no longer, ‘Can you get the job done?’ It’s, ‘Can you get the job done, and are you cheap at doing it?’,” Levy said. The second question is where model routing, workflow redesign and the budget meet. A productivity claim without cost evidence leaves the business case unfinished, and an unfinished budget argument is of limited use to an HR director.
What should employers test before scaling?
The partnership offers a capability, not a performance guarantee. Cognition’s claim that Devin can plan, write and test code raises four adoption questions for employers:
- Which engineering projects should receive access first, and which tasks should remain with people?
- Can longer autonomous assignments produce a measurable gain after teams and workflows change?
- Does a multi-model approach preserve required quality while controlling cost?
- How should employers evaluate engineers whose work includes directing AI output?
The survey already shows a split outcome: a quarter of organizations more than doubled productivity, while nearly 30% declined. The next phase of AI adoption will be judged less by demo speed than by whether redesigned teams can repeat the gains without adding cost. The Ferrari remains optional; the process redesign does not.
Source: feeds.businessinsider.com
More from Daniel Okafor
Show full bio
Senior reporter covering technology and product news in Recruiting and talent acquisition, learning and development at HR Briefing.
283 articles