Software engineering is one of the highest-impact arenas for agentic AI: it sits at the centre of the modern economy, and it is unusually well suited to systems that can plan, write, run and verify their own work. Code can be edited, tested and corrected, which is why these agents are already in daily use – implementing features, repairing failing builds, migrating legacy systems and opening review-ready pull requests. The payoff is shorter cycle times, more output per engineer and less time lost to boilerplate and maintenance. That combination of leverage and real production use makes agentic software engineering one of the most promising applications of AI in the enterprise.
But realising that opportunity requires more than deploying new technology. Companies need to rethink how they prioritise engineering work, measure productivity and maintain accountability as AI agents take on more responsibility.
Cognition, the company behind Devin, an autonomous AI software engineering agent, is working with enterprises navigating this shift. Reflecting the growing appetite for AI-powered software engineering, Cognition recently raised $2 billion at a valuation of $48 billion. McKinsey Senior Partner Anita Balchandani spoke with Cognition co-founder and President Russell Kaplan about the evolution of autonomous engineering, what distinguishes successful enterprise deployments and how leaders can translate advances in AI into measurable business outcomes.
From coding assistants to autonomous engineering
Anita Balchandani: Cognition was founded on the belief that AI could take on software engineering tasks autonomously. How has that vision evolved as Devin has moved from an early concept into some of the world’s largest organisations?
Russell Kaplan: From day one, the bet was that AI would work as a co-worker, not an autocomplete tool. That’s why it’s called Devin and not something that sounds like a tool.
Devin was meant to be a full agent with its own machine, working in the same systems as your team: Slack, Jira, your repositories, your continuous integration tools. You hand it a real task and get back work an engineer can review, test and ship.
When we launched in early 2024, that was an aspiration as much as a product. The models were just good enough to start, and early Devin needed a lot of hand-holding. It did best on contained work like a bug fix or a migration. Shipping early let us learn from real customers sooner than anyone else.
Today you can work with Devin as a co-worker inside companies with decades of code, connected systems and strict access and review rules. It tests and checks its own work. A growing share of Devin sessions aren’t started by a person at all. They’re triggered automatically, and Devin goes and finds the work that needs doing.
Engineers direct far more work than before, and they’re still accountable for what ships.
Anita Balchandani: AI models are evolving rapidly, while enterprises are making long-term investments in how their engineering teams operate. How does Cognition’s model-neutral approach help companies navigate that uncertainty?
Russell Kaplan: A large organisation will invest years in changing how its engineers work. In that time, models will get better, and the right model for one kind of task may differ from the right model for another.
We call ourselves the Switzerland of AI. Devin runs on frontier models from the major labs and on our own SWE family of models, and we evaluate them openly to pick the best mix of capability and cost for the work.
That matters because the harder part of enterprise adoption is building a reliable way to assign work, grant appropriate access, review results and measure output.
An organisation needs those practices to endure. As models improve, we can bring the improvement into Devin, and the customer can keep building on the way its teams already work.
From engineering productivity to business value
Anita Balchandani: Companies are investing heavily in AI-powered engineering, but greater productivity doesn’t automatically translate into business value. What distinguishes those getting it right?
Russell Kaplan: It starts with choosing what the organisation wants to accomplish. Leaders name the outcome, assign an owner and establish what it would cost and how long it would take the current way.
A modernisation project might be holding back a product launch. A security backlog might be tying up a team the business needs elsewhere.
At EBANX, growth depends on entering new markets, and every market means integrating dozens of local payment providers. Engineering capacity was the limit on how fast they could expand. So they tested Devin on a payment provider integration, the exact work holding the business back.
Comarch did the same with a 20-year-old, ten-million-line ERP system. They set the goal of rebuilding it within 18 months and ran a controlled proof of concept scored on quality, speed and total cost of ownership. Once the outcome is set, the engineering team can decide which parts of the work Devin should take on. Engineers still set direction, understand the system and review what ships. Their role becomes more focused on making good decisions and checking the result. If an agent completes more tasks but the important project does not move, leadership needs to change what it is asking the agent to do. The measure is whether the work helped the organisation achieve the outcome it chose.
Anita Balchandani: Where does Devin still need a human to step in, and what shouldn’t a company hand it yet?
Russell Kaplan: Engineers remain accountable for direction and for what ships, and that is by design. Devin is strongest on work that can be clearly scoped and verified – a migration, a well-defined feature, a security fix that can be validated at runtime. Where the task is ambiguous, where the real challenge is deciding what to build rather than building it, or where a change carries consequences that no test suite fully captures – those are exactly the places a human should set direction and review the result. The teams that get the most from Devin aren’t the ones that hand it everything; they’re the ones that are deliberate about which decisions stay with people and which execution they delegate. Getting that division of labour right matters more than the raw capability of any single model.
Anita Balchandani: What should CEOs and boards demand as proof that their investment in AI-powered engineering is delivering real value?
Russell Kaplan: Start by refusing to accept activity as proof. Tokens, seats and lines of code tell you how much the tool was used, not what it produced.
A mechanical refactor can generate thousands of lines, while a hard bug investigation might change only a few. When we tested it, code volume turned out to be a weak predictor of how much engineering effort a task took.
The question a CEO should ask is simple: What work got done, and what would it have cost us without AI?
We built a system to answer it. For every completed Devin task, an estimator measures the equivalent human engineering hours, and we validate it against engineers’ own estimates of how long the same work would have taken them. We calibrated it to undercount, not overcount. We then convert those hours into dollars and compare them with what the customer spent.
We’re confident enough in that number to back it with the AI Productivity Guarantee: if Devin delivers less engineering value than an enterprise customer is paying for, we cover the gap with usage credits, up to $10 million.
Every AI vendor should be able to tell customers what they’re getting for their money. Engineering value proves the investment pays for itself. Business outcomes prove it was worth making.
Engineering value proves the investment pays for itself. Business outcomes prove it was worth making.
Scaling AI beyond the pilot
Anita Balchandani: Proving AI can work is one thing. Making it deliver value across an enterprise is another. What does it take to make that leap in practice?
Russell Kaplan: A pilot should start with a project that matters and an honest account of how the company handles it today.
What’s in the backlog, who owns it and what would it cost to finish with the current team and process? Access and security need to be settled before work starts. Those decisions give the pilot a fair test and give leadership something real to judge when it ends.
That’s where our forward-deployed engineers come in. They don’t hand over a login and leave. They embed with the customer’s team, find the high-value projects, pair with engineers on their real backlog, get Devin into the workflow and systems the team already uses, and run workshops on managing a fleet of agents well. They also measure what gets done.
At Comarch, our forward-deployed engineers were on site in Krakow from the first day of the pilot, working with Comarch’s best engineers and analysts to show that Devin could understand that legacy system well enough to rebuild it.
That’s how a pilot becomes something an organisation can repeat. A successful pilot should unlock value for the entire organisation, not just one team.
Anita Balchandani: As agents take on more of the work, leaders are asking what this means for their teams. How is the role of the engineer changing – and what does that imply for how companies hire and organise?
Russell Kaplan: The work shifts up a level. Engineers spend less time typing out implementations and more time on the decisions that determine whether the implementation is right: what to build, how the system should fit together, where the risks are and whether what came back meets the bar.
That makes judgment, system understanding and the ability to specify work clearly more valuable, not less. We see teams direct and review far more output per person than they used to, which means the constraint moves from how fast people can write code to how well they can decide what’s worth building and verify that it’s correct. The organisations handling this well are investing in those skills and treating their engineers as the people who direct a growing amount of work – rather than assuming the agent reduces the need for engineering talent.
Anita Balchandani: In your experience, what separates companies that see meaningful returns from AI in engineering from those that struggle to gain traction – and what should leaders prioritise in the first 90 days?
Russell Kaplan: They get specific early. They choose a bounded set of work with a real cost and agree on what completion looks like. They make sure Devin can access the repositories and systems it needs. And they put an engineering owner in place to direct the work and review the output.
Those steps sound basic, but they determine whether the team spends its time doing useful work or waiting on access, approvals and setup.
Then they look at completed work alongside cost. Which tasks moved? How much human review was needed? What happened to the project timeline? The next use case should follow from those answers.
In the first 90 days, I would rather see a company establish one working pattern it can carry into another team than spread a thin pilot across an entire organisation.
The European opportunity
Anita Balchandani: Cognition is expanding its European presence, with new offices in London and Madrid, as businesses face growing pressure to modernise complex, often decades-old technology systems. What’s driving demand for autonomous engineering in the region, and how is that shaping your expansion strategy?
Russell Kaplan: The decision came from demand that was already there.
European enterprises run some of the largest, oldest and most regulated codebases in the world. They face constant change and strict security requirements, and many carry modernisation work that has been put off for years. That’s exactly where autonomous software engineering pays off, and European customers were adopting Devin well before we had an office here.
We see it across industries. In financial services, we work with customers including Santander, Hg Capital and The Access Group. In manufacturing and enterprise software, Mercedes-Benz and Comarch, and in services, Endava and WPP.
Our strategy is to be where our customers are. London is our EMEA headquarters, and Madrid anchors Southern Europe. The core of each office is our forward-deployed engineers, who work on site with customer teams to find the highest-value work, get Devin into their systems and measure what it delivers.
As we accelerate our European footprint, we’ll keep adding locations so we can stay close to the work.
Anita Balchandani: For enterprises operating in highly regulated environments, how can they harness the potential of autonomous engineering while maintaining the security, governance and control they need?
Russell Kaplan: We begin with the requirements of the organisation in front of us. Which systems can an agent access? How does the company handle its code and data? What review and audit processes apply to changes? And what deployment model will let Devin work with the customer’s network and development tools? Those are practical design decisions that the security and engineering teams need to make together.
For some organisations, private systems and network requirements affect the deployment they choose. For others, the immediate question is how to bring agents into an existing approval process. Our local teams can work through those details with customers. Getting the governance right at the start gives an organisation a clearer path to using agents on more consequential work.
What enterprises should expect from AI partners
Anita Balchandani: With so many AI providers competing for enterprise investment, how can business leaders distinguish those that can deliver lasting value from those simply promising it?
Russell Kaplan: Expect them to prove it, on your work and with their money behind it.
Four things matter.
First, results on your own codebase, not a demo.
The strongest customers we work with run real evaluations. Comarch gave several vendors the same scoped refactoring of its ERP system, staffed it with the same team and scored everyone on quality, speed and total cost of ownership. WPP had already rolled out an AI coding assistant to 500 engineers and seen none of its delivery metrics move. Then it ran side-by-side evaluations before choosing again.
Second, accountability for value, not activity.
Tokens and lines of code show usage, not return. A vendor should be able to tell you what work got done and what it’s worth. That’s why we offer the AI Productivity Guarantee: if Devin delivers less engineering value than you pay for, we make up the difference in credits, up to $10 million.
Third, independence.
Models change every few months. A vendor tied to one model or one cloud makes its roadmap your roadmap. You should be able to use the best model for each job as the field moves, without starting over.
Fourth, people on the ground.
Most of the value comes after the contract is signed: finding the right work, getting through security and access, changing how teams operate. Our forward-deployed engineers sit with customer teams to do exactly that. If a vendor can’t show up, measure what they delivered and stand behind it, they shouldn’t be one you work with.
Anita Balchandani: Finally, the ultimate proof point is your own company. How much of Cognition’s software is now built by Devin, and what has that let your team do that it otherwise couldn’t?
Russell Kaplan: Devin now writes roughly 95 percent of the code we ship, and it did not get there gradually. In the past six months, our total code output grew about 7x.
The speed shows up in ways customers recognise. A feature that would have queued for a sprint comes back the same day as a review-ready pull request. When writing the code is no longer the bottleneck, the real question becomes what is worth building.
That is where our engineers’ time goes now. Each one directs several Devins, and their job is to decide what to build, how the system should fit together, and whether what comes back meets the bar. It is the same division of labour we described earlier, running at full scale inside our own company.
It also changes how we hire. We look for people who are good at judgment calls, not just good at writing code. The companies seeing the best results with Devin are building the same muscle.