McKinsey Says 20x Developer Productivity Is Coming. Most Teams Can't Even Measure 1.2x.
McKinsey just published an excerpt from the second edition of Rewired that opens with a scene worth sitting with: three engineers walk into a London bank at 8 a.m. Overnight, nearly a hundred AI agent teams refined a cross-border payment system, tested failure paths, and shipped updates. Waiting for the humans is a stream of pull requests, test evidence, and risk flags - more progress in 12 hours than a traditional team might make in a month.
That agent factory is real, running at a global bank, and delivering 10 times the speed at half the cost.
The article maps the progression as four levels of developer support: engineers writing everything themselves, engineers assisted by gen AI tools like Claude and Copilot, agentic workflows automating entire steps, and finally small teams supervising an agent factory that delivers whole applications - the level McKinsey pegs at 20x leverage.
Most companies, McKinsey says, are at level two. Call it 1.2x: the tools are bought, the boost is real but modest, and nobody is quite sure how big it actually is.
The question buried in the middle of the article
The excerpt lists three questions every C-suite should be asking. The second is the one we'd frame the entire piece around: how do you actually know whether your organization is going up the software development productivity curve?
Because here's the pattern McKinsey found when it analyzed nearly 300 public companies: handing developers AI tools doesn't move the needle. The top quintile - companies seeing 16-30% improvements in productivity, time to market, and customer experience, plus 31-45% gains in software quality - did three things differently. They invested in hands-on upskilling. They aligned incentives, with about 80% tying gen AI goals to developer and PM evaluations. And critically, they institutionalized tracking outcomes - release frequency, defect rates, customer experience - not just adoption metrics.
Read that last one again. The winners aren't counting Copilot seats or measuring "percent of engineers using AI." Those are adoption metrics, and they tell you nothing. A 100% adoption rate is compatible with engineers shipping more bugs faster.
IBM learned this the expensive way
The article's IBM case study is a warning label for every AI rollout happening right now. IBM's first attempt at an AI-powered development model looked good on paper - 200 people onboarded every other week. But adoption was patchy, tooling didn't stick, and when the AI misbehaved, engineers quietly reverted to their old habits.
Leadership had no way to see this happening until the results didn't show up. It took a second, far more expensive push - coaches embedded with every team, office hours, a champions community, more than 8,000 developers over six months - to make it work.
The gap between "we bought the tools" and "the tools are creating value" is exactly where most AI investments die. And it's invisible unless you're measuring actual output.
The 20x world makes measurement more urgent, not less
McKinsey's night-shift factory model - humans set direction by day, agent fleets code, test, and document by night - comes with new failure modes the article is refreshingly honest about:
Quality drift. Without well-scoped, agent-ready tasks, agents stall or wander off course. Someone has to catch that early, not at the sprint review.
Runaway cost. Agents that spawn subagents can make token consumption grow exponentially, producing serious cost overruns. McKinsey's prescription: build FinOps for agent activity before scaling it.
Guardrail violations. The factory model depends on an orchestrator halting workflows when policies break - which assumes you can see the violations in the first place.
In other words: the higher you climb the productivity curve, the more your leverage depends on observability. At 1x, a bad week costs you a bad week. At 20x, an unmonitored agent fleet can ship a month of problems overnight.
Where you actually are on the curve
The honest answer for most teams reading this: you're at level two, you're paying for tools across three vendors, and if the CEO asked "is this working?" the answer would be a vibe, not a number.
The 20x frontier is coming. The companies that get there first won't be the ones that bought the most tools. They'll be the ones that could see, week by week, whether they were actually moving up the curve - and corrected course when they weren't.
Know your number before your competitors know theirs.
_Source: "The AI revolution in software development" by Charlotte Relyea and Martin Harrysson, McKinsey & Company, April 2026, excerpted from Rewired (2nd edition, Wiley)._