Back to Explore
NewsVibeFix editorialUpdated Nov 21, 2026

Devin teams up with McKinsey: the AI coding race shifts from tools to trust

Cognition and McKinsey formed an alliance on 16 November 2026 to sell AI coding transformation to large enterprises: Devin does the work, McKinsey rewires the organisation. McKinsey's survey: the top quarter of organisations more than doubled productivity, while 30% went backwards — the tool does not decide the outcome, the way of working does.

Engineers gathered around a whiteboard in discussion, laptops on the table

In short: Devin does the work, McKinsey makes the work count

On 16 November 2026, Cognition, the company behind Devin, and consulting firm McKinsey & Company announced an alliance to sell "AI-written code" to large enterprises in the US and Europe. The division of labor is clear: Devin takes on engineering tasks, while McKinsey helps clients pick projects, set baselines, and change how teams work, so what the AI produces shows up in business results. The primary source is the official release issued that day (BusinessWire, 16 November 2026, 17:00 UTC); the core figures and quotes below all come from it, read in full.

This alliance is worth writing about not because another company is selling an AI coding tool, but because it admits something many sellers will not: handing an AI agent to engineers does not necessarily raise productivity — it can lower it. The release quotes a McKinsey survey from the second quarter of 2026: among 334 product and engineering leaders, the top quarter of organizations reported productivity gains of more than two times, while 30% reported declining productivity. Same tool, some doubled, some went backwards. The difference is not the tool; it is whether the organization changed with it. That is why this alliance exists.

The numbers first: the top quarter doubled, three in ten slowed down

Look at the survey numbers in the open. 334 product and engineering leaders, surveyed in Q2 2026. The best 25% said AI more than doubled their productivity; 30% said productivity fell. The remaining roughly 45% are not broken out in the release — they sit in the middle, neither doubling nor declining, or changing little. That distribution itself is informative: the effect of AI coding tools is not evenly spread, it is polarized. Buy the same tool and the outcome can be completely different.

The release's explanation is that the difference increasingly comes down to whether a company redesigns how its engineering organization works, not whether it has access to AI. That matches the direction of other independent research from the same period, but one caution: the survey is McKinsey's own, the sample is 334 leaders, and the answers are self-reported productivity changes, not independently measured delivery data. Self-reports have slack — leaders may overstate success and understate failure, or credit AI for gains that came from other reforms. So the numbers are good for understanding "why some win and some lose", not as a promise that "using AI doubles you". The alliance's pitch is built on exactly that gap: it sells not the tool, but the service that fills the gap.

McKinsey also cites another study to back the pitch: its Rewired research, covering 20 leading transformations, found that companies focusing on one to three high-value domains such as software saw an average EBITDA uplift of about 20% and broke even within one to two years. Apply the same discount: 20 "leading" cases are selected successes, not a random sample; the EBITDA uplift is an average, and it bundles consulting, organizational change, and AI tooling together, so it cannot be attributed to Devin alone. The release puts the two figures side by side for a reason: the alliance sells transformation outcomes, not tool subscriptions.

What the alliance actually sells: Devin works, McKinsey rewires

According to the release, the alliance works in three phases: help clients choose the projects that matter most, set a baseline for cost, speed, and quality, then scale new ways of working so finished projects deliver measurable business value. Devin's role is the "doing": it takes on engineering work across multiple projects, connects to the systems and security processes a company already uses, and brings the work back for review. Engineers set direction, solve the hardest problems, and decide what ships. McKinsey's role is the "rewiring": reimagining operating models, team structures, talent, and technology, with training, hands-on coaching, and ongoing support for everyone from individual engineers to senior executives.

One detail matters: Devin's platform is model-agnostic, matching different models to different kinds of work. The release claims this delivers frontier-level performance for 60% less, with room to adopt new models as they improve. The 60% is a vendor figure, unverified independently, with no stated baseline — cheaper than what? The most expensive model only, or than humans? The release does not say. Treat it as a direction — model routing can save money — not as a precise promise. The question a buyer should ask is: how was the 60% calculated, against which model and which task mix?

The alliance's other pitch is compliance and trust. The release notes that Cognition works with six of the eight global systemically important banks (G-SIBs) and three of the six largest health insurers — among the most regulated industries there are. Getting into those customers says Devin's security and audit processes passed their bar. Russell Kaplan, Cognition's co-founder and president, says in the release that the company is already working with regulated enterprises around the world, and that those clients want to know three things: what was done, what it would have cost without AI, and what completing the project makes possible. The alliance exists to answer exactly those three questions — the same three a tool-only vendor cannot answer.

Comarch's 18 months: the only named case

The only named customer case in the release is Comarch, a European software company. It is using Devin to modernize ERP XL, a 20-year-old ERP system, on an 18-month timeline, and is working with McKinsey to extend that work across more teams. Tomasz Stachlewski, Comarch's CTO, says in the release that customers could not be asked to wait five years, so the team set an 18-month goal; Devin is helping engineers work through the 20-year-old system and get new functionality to customers sooner, while McKinsey helps redesign how teams work to embed Devin into engineering workflows, scale adoption, and measure business impact.

18 months versus 5 years is an eye-catching comparison, but be clear about what it is: a target timeline for a project still in progress, not a delivered result. The project is ongoing; whether the 18 months will hold, what the quality looks like, and how much rework happened along the way are not in the release. It works as an example of "how the alliance helps a client set goals"; it does not work as proof that "Devin compresses five years of work into 18 months". Likewise, the six banks and three insurers show the weight of the client list, not that each client achieved measurable results. A client list and results are two different things, and the release places them together — readers should separate them.

Saptarshi Ganguly, a Senior Partner at McKinsey, states the alliance's thesis in the release: "Doubling software development productivity takes more than deploying AI agents. It requires fundamentally rethinking how work gets done. The winners will be those that transform their teams, roles, and operating models as boldly as they embrace the technology." The subtext is that the alliance's real product is neither Devin nor a consulting deck — it is the transformation itself. The tool is the entry point; the transformation is the business.

McKinsey's Q2 2026 survey of 334 product and engineering leaders: the top quarter reported productivity gains of more than 2x, while 30% reported declining productivity. Comarch is using Devin to modernize its 20-year-old ERP XL system on an 18-month timeline instead of five years

Two numbers to discount: 60% cheaper, 20% EBITDA uplift

The release contains two numbers that make the best headlines and most need discounting. The first is "frontier performance for 60% less". The logic is model-agnostic routing: cheap models for easy work, strong models for hard work, and total cost comes down. The logic holds — routing across models is one of the most practical ways to save money today. But the 60% comes with no baseline and no task mix, so readers cannot tell which tasks it was measured on, or which benchmark "frontier performance" refers to. A percentage without a baseline is meaningless: it shows direction, not a budget figure.

The second is "average EBITDA uplift of about 20%, break-even in one to two years". Its source is McKinsey's Rewired research on 20 leading transformation cases. The problem is again the sample: 20 "leading" cases are winners, and failed transformations do not appear in such research; and the uplift in those cases is the packaged effect of consulting, organizational change, and technology, with Devin only one part of it. Reading 20% as "what you get when you buy the alliance" misreads the study. What it really shows is that in companies that transformed successfully, focusing on high-value domains did pay off handsomely — but the premise is "successful transformation", and successful transformation is precisely the service the alliance sells.

Both numbers share a trait: they are vendor figures with no independent verification. That is normal for an official release — its job is to announce the alliance, explain the division of labor, and set out the vision, not to publish an audit. The reader's job is to separate facts (the alliance exists, the division of labor is this, the survey numbers are these) from claims (60% cheaper, 20% returns, doubled productivity). Facts can be quoted; claims should carry their source and limits. That is how this article treats both numbers: stated, with source and limits attached.

How this differs from the Devin articles already published

This site has already published several Devin pieces, so this one's place needs stating. In November there was a Devin Budgets article, about the product feature that caps agent spending and scores efficiency — a feature explainer. In October there was a piece on Devin "dreaming" and an open standard for agent memory. Earlier still came pieces on Devin's price cuts and its revenue run-rate. None of those is about a product feature; this one is about a business deal: Devin plus McKinsey, selling "transformation outcomes from AI coding" to large enterprises. The feature pieces answer "what can Devin do"; this one answers "why should a large enterprise believe it pays off".

That boundary also defines the reader: not the individual developer setting a Devin budget, but the tech lead or engineering manager asking "is this AI rollout worth it for my team?" For an individual developer, the practical value of this piece is limited — the alliance serves large US and European enterprises, and its pricing and process are not built for you. But the survey numbers are relevant to you too: the top quarter doubled, three in ten went backwards. That distribution says the tool does not decide the outcome; the way of working does.

What it means for smaller teams: you can build half of what the alliance sells yourself

The alliance serves large enterprises, but the problem it names applies to smaller teams too: ship the tool without changing the process, and productivity may fall instead of rising. Small teams cannot hire McKinsey, but they can do the cheapest half of what the alliance sells themselves: pick projects, set baselines, change the process. Pick projects: start with one or two jobs where AI helps most — repetitive migrations, test backfills — instead of rolling out to everyone at once. Set baselines: record today's cost, speed, and quality before adopting the tool, or there will be nothing to judge it against later. Change the process: write down the rules for review, approval, and rollback, instead of letting agents push straight to the main branch.

The line in the release — "engineers set direction, solve the hardest problems, and decide what ships" — applies to small teams too, and matters more there. Large enterprises have process as a safety net; small teams do not, and an engineer's judgment is the last line of defense. The alliance's model institutionalizes that line: Devin works, a person reviews, the organization backs it up. Small teams lack the third layer, which makes the first two non-negotiable. Put the other way: if your team already does "the agent works, a person reviews seriously, and every task has a baseline comparison", you already have the core of what the alliance sells — the rest is scale and consulting packaging.

One last yardstick for judging where your own team sits in the distribution: after the last task AI did for you, can you say offhand what it cost, how much time it saved, and whether there was rework? If you can answer all three on the spot, you are probably on the doubling side; if you cannot answer any of them, be careful — the tool may be quietly slowing you down rather than helping. The alliance's service, at its core, is helping large enterprises build the ability to answer those numbers on demand. Small teams do not need to buy that service, but they need the habit.

What each side wants from the alliance

The alliance runs both ways, and each side gets something it lacks. Cognition gets a trust endorsement and a ticket into large enterprises. Devin's technical ability is established; what it lacks is the thing large buyers care about most when purchasing: someone accountable for the outcome. A single AI tool vendor struggles to answer a board's question — "we spent this much, where is the return?" McKinsey fills that gap: it brings not just a methodology but credibility in driving change inside large companies. For Cognition, the alliance upgrades "selling a tool" into "selling an outcome", with deal size and renewal rates likely to follow.

McKinsey gets a new lever for the AI era. Traditional consulting sells people's time and experience, and AI coding tools are turning part of that experience into a product. If McKinsey only sells methodology, clients can buy the tool directly; if it only pushes tools, it has no product of its own. The alliance gives McKinsey a product it can land its methodology on — Devin — so transformation stops being advice written in slides and becomes tasks running in a workflow every day. For McKinsey, this is a move to stay relevant in the AI era.

This "tool plus consulting" combination was already a pattern in 2026, not a first. AI tool vendors teaming up with consultancies is becoming a standard play: the tool does the work, the consultancy makes people believe the work is worth it. When reading this kind of alliance news, the question to ask is not "did they team up?" but "will any client publish its baseline data afterwards?" Baseline data is the only test of whether an alliance delivers — without a baseline, every return figure is just a claim. Comarch's 18 months is a start, but one case is not enough; the story only stands when more clients come forward with their own books.

Conclusion: the AI coding race is shifting from tools to "how do you make people believe"

Seen in the November landscape, this alliance marks a shift: competition in AI coding is moving from "whose model is stronger, whose feature list is longer" to "who can make people believe it pays off". Devin's model-agnostic routing, McKinsey's transformation methodology, Comarch's 18-month target — all answer the same question: how do you prove the money was worth it? The old answer was benchmarks and demos; the new answer is baselines, review, and business results. That shift is good for buyers — it forces sellers to show something measurable instead of only a vision.

But the other side is visible too: the evidence is still thin. The only named case is still in progress, the two most eye-catching numbers are vendor figures, and the survey is self-reported. Whether the alliance delivers on its promise depends on Comarch hitting its 18-month target on time, and on more clients being willing to publish their baseline data. Until then, this article's judgment is: the direction is right — tool plus transformation is indeed the correct recipe for AI coding adoption — but its effectiveness has no independent evidence yet. Treat it as a signal worth tracking, not a conclusion already proven.

Views 5Comments 0

Comments (0)

ME
0/1000
Loading comments...

Sources

Browse projectsPublish your project

Related articles

A scientist in a white coat and safety glasses examines a sample under a microscope in a bright laboratory, holding a tablet
News
Great at Assembling Molecules, Weak at Picking Drugs: OpenAI's MedChemBench Grades AI's Medicinal Chemistry

On 20 November 2026, OpenAI released MedChemBench: 99 medicinal-chemistry problems from 117 papers and patents, covering 761 molecules. GPT-6.2 Astra scores 84.5% overall — but 95.4% on assembling structures and only 50.4% on choosing compounds. The total flatters; the components tell the truth.

Industry TrendsModel UpdatesTesting & Quality
A developer with glasses types on a laptop late at night at a wooden desk, a large monitor behind showing colourful code, blurred city lights outside the window
News
Run the Middle Layers Three Times: Inkling 2 Mini Puts a 26B Coding Model on Your Laptop

On 20 November 2026, Thinking Machines Lab released Inkling 2 Mini: a 26.4B-parameter looped transformer that runs its middle 12 layers three times per token, fits in about 20GB, and runs offline on a laptop. It leads open-weights models of its size on DeepSWE and SWE-bench Verified — with the caveats printed in the same announcement.

Product LaunchModel UpdatesAI Coding