A May 2026 McKinsey survey of 334 product and engineering leaders shows a wide divergence in how AI affects software development. Only 25 percent of respondents at director level and above reported that more than a quarter of their teams achieved twofold or greater productivity gains, while 30 percent said team productivity had actually fallen. The piece builds on McKinsey’s November 2025 baseline research on AI in software development.
The article’s authors — Charlotte Relyea, Gautam Lunawat, Janaki Palaniappan, Martin Harrysson, Matt Linderman and Prakhar Dixit, with contributions from Albert Savitzky and Vasu Tekriwal — argue that technology alone does not explain who wins. Instead, organizations that generate disproportionate value redesign the full product development life cycle (PDLC): end-to-end workflows, roles and responsibilities, control and measurement systems, and change management for an agentic operating model.
Survey scope and potential economic upside
The May 2026 survey included 334 product and engineering respondents, 173 of whom were director-level or higher, across regions (Africa, Asia, Europe, Latin America and North America) and industries (technology, media and telecommunications; finance and insurance; consumer and retail; healthcare and life sciences; and manufacturing, logistics and transportation). Developer-level analysis showed the top 20 percent of engineers reporting an average AI productivity increase of about 55 percent, while the average acceleration was roughly 3 percent.
McKinsey estimates that if the world’s roughly 30 million engineers could be moved toward the productivity of the top quintile — a conservative target given some leading teams already exceed twofold gains — the potential new value would approach the GDP of a midsize economy (about $0.8 trillion, assuming roughly $85,000 average fully loaded cost per engineer). This estimate excludes AI productivity gains in non-engineering PDLC roles.
Four distinguishing practices of outperformers
Across high-performing organizations, four patterns recur: 1) reimagining end-to-end processes rather than merely plugging AI into old workflows; 2) changing team structures, roles and handoffs to focus human effort where judgment and product intent matter; 3) building verification, control and measurement systems that keep up with faster work; and 4) investing in hands-on upskilling and embedded coaching.
Workflow redesign emerged as a top enabler among respondents who reported at least some teams achieving twofold gains. Ninety-three percent of top accelerators — those reporting more than a quarter of teams with twofold or greater productivity increases — embedded AI into workflows. Organizations that redesigned processes before adding AI were more than twice as likely to report over 20 percent productivity gains than those that layered AI onto existing ways of working. By contrast, organizations that adopted AI without holistically reshaping workflows tended to miss meaningful impact despite following many adoption best practices.
Why end-to-end redesign matters
Agents operate independently and continuously, enabling asynchronous progress: agents can advance work during a ‘‘night shift’’ while humans focus on daytime review and high-judgment decisions. This diverges from linear human handoffs and traditional scrum cadences; simply speeding existing cycles rarely yields the same upside as redesigning work for agent-driven execution.
Sonar provides a concrete example: three frontrunner teams embedded agents across research, ideation, backlog definition, coding, testing and remediation. Agents synthesized customer and market signals, converted discovery into Jira tickets, generated tests and code, and advanced bug reports to draft pull requests. As a result, pull-request cycle duration fell 3.4x, pull-request throughput improved 2.2x, and build activities became 50–80 percent more productive.
Role redesign and new human work
Role redesign is often where top accelerators separate themselves. Organizations that simply added AI without changing roles or ways of working (‘‘adopters’’) became top accelerators only 27 percent of the time; ‘‘experimenters’’ (some AI tools, not workflow-integrated) reached that level 17 percent of the time; while ‘‘transformers’’ (AI embedded and roles/work reworked) became top accelerators 40 percent of the time.
As agents take on routine execution work — documentation, test generation, first-draft artifacts — human judgment gains importance. Teams must translate product intent into executable, telemetry-linked requirements: acceptance criteria, definitions of done, guardrails for agent actions, and clear routing to reviewers, tests or approval gates. Vague, tacit narratives that an experienced engineer might interpret through conversation are no longer sufficient; specifications must be explicit enough for an agent to act on, for tests to validate, and for reviewers to judge.
Survey data show shifts in time allocation: product managers at top-accelerating organizations report a 24 percent reduction in time on execution activities and an 18 percent reduction on strategic activities; developers report a 19 percent reduction in execution time but only a 6 percent reduction in strategic work. Squad sizes are shrinking: 79 percent of respondents at top accelerators report smaller squads post-AI, with median squad size dropping from about ten to seven. Nearly 20 percent of director-level respondents in top accelerators are already on teams of one to four people. Role boundaries are blurring as product and design converge and testing splits between automated agents and engineers.
Leading organizations are beginning to define roles for agents as well as people, giving agents scoped mandates, permissions, escalation paths and accountability, so the unit of design becomes a blended people–agent system organized around a redesigned flow of work.
Verification, observability and AI Ops
Faster generation can outpace the ability to test and govern output. The survey shows time savings (average 11.8 percent) improving faster than rework reduction (average 6.2 percent); in development activities, time savings average 11.2 percent and rework reduction 6.8 percent. If leaders prioritize velocity without visibility, technical debt can accumulate faster than teams can address it.
External studies align with this risk: DORA’s 2025 research describes AI as an amplifier of existing delivery dynamics; Veracode’s 2025 report found 45 percent of AI-generated code samples failing security checks; GitClear saw increases in duplicated code in repos with AI-assistant activity; and Stack Overflow’s 2025 survey reported low developer trust in AI outputs.
Top performers design verification into AI-supported workflows: agents generate tests and security scans, compare outputs to requirements, surface missing acceptance criteria, identify likely defects and propose fixes. Checks feed into the normal release path with auditability and traceability rather than being a post hoc layer. Organizations triage by risk and value, automating low-risk recurring fixes while routing higher-risk issues — customer-facing behavior, regulated data, security-sensitive code, architecture changes — to accountable human review.
A shared AI Ops layer is essential: identity, permissions, tool access, routing, evaluation, logging, escalation, policy enforcement and cost visibility must be centralized. The survey suggests many organizations have AI Ops capabilities but staff them lightly and allow teams to improvise; outperformers invest more and treat orchestration and evaluation as foundational. Organizations with five or more AI Ops mandates (governance standards, agent/platform performance monitoring, etc.) report greater positive impact on quality and time savings.
Example: a convenience retailer embedded verification into the delivery flow, standardizing CI/CD and DevSecOps so coverage, standards and release-readiness checks are enforced automatically and agents generate tests. Front-runner teams achieved velocity increases up to 30 percent and reduced defect-resolution effort by 50 percent.
Change management, coaching and measurement
Hands-on training and embedded coaching matter. Embedded coaching is about 30 percent more common in top accelerators than in low accelerators. Successful adoption typically places experienced AI engineers and coaches within teams long enough to shift habits, help rewire workflows, configure agents and review outputs. The model is moving from one-off training to continuous, embedded coaching at scale.
Measurement should move from adoption metrics (licenses, daily active users, prompt volume) to outcome metrics (quality, time to market, security, reliability, cost, employee experience, customer impact). Eighty-six percent of top accelerators track outcome metrics, whereas low accelerators rely disproportionately on tool adoption as their North Star.
Leaders must also plan for AI-related costs: tokens, compute, infrastructure and AI Ops can increase to as much as 20 percent of existing labor costs in some cases. Organizations that lead are treating AI adoption as something to fund, practice, measure and govern.
Practical steps and progress markers
The authors suggest concrete steps and illustrative progress markers (not survey-derived benchmarks):
-
Target highest-value workflows and redesign end-to-end. Progress marker: the redesign funds itself by automating whole categories of manual work (first-pass research, test writing, reporting), freeing capacity to pay for tokens and coaching.
-
Redesign roles for people and agents. Progress marker: at least 70 percent of job descriptions change and squads shrink to two–four people working with agents.
-
Raise the standard for requirements to be testable and telemetry-linked. Progress marker: a requirement handed to an agent returns a near-shippable artifact.
-
Build verification and a shared AI Ops control plane. Progress marker: agents operate across teams, tools and knowledge sources without bespoke plumbing while leaders see activity, cost and risk on a single control plane.
-
Run adoption as a change program with embedded coaching, show-and-tell, office hours and incentives. Progress marker: 90 percent of human/agent teams running day-and-night shifts, two-day sprints and 100 percent outcome metric tracking.
Conclusion
AI is reshaping product and software development by shortening the path from customer signal to product decision, requirement to code, defect to fix and release to measured outcome. The upside is substantial and some organizations are already capturing disproportionate value by redesigning the entire system for an agentic PDLC rather than merely adding tools to old routines. For others, the priority is to scale AI across processes, roles, verification and change programs quickly enough to prevent the performance gap from becoming entrenched.
Authors: Charlotte Relyea (senior partner, McKinsey New York), Martin Harrysson (senior partner, Bay Area), Gautam Lunawat and Janaki Palaniappan (partners), Albert Savitzky (associate partner), Vasu Tekriwal (consultant), Matt Linderman (partner, Connecticut) and Prakhar Dixit (partner, Seattle). The authors thank Abdul Wahab Shaikh, Adam Thelwall, Ling Lau, Nandita Bothra and Nicolas de la Flor. Edited by Daniel Eisenberg (executive editor, New York).
