Capturing AI’s full value
AI is rapidly changing software development, but research consistently shows that organizations still struggle to translate experimentation into meaningful, sustained impact. The promised land of tenfold to 20-fold productivity from AI still seems far away for most companies.
The challenge goes beyond tool adoption and requires redesigning the product development life cycle (PDLC), including the fundamental ways software teams work. McKinsey research finds that organizations progress across three horizons:
Most organizations are still operating in Horizon 1—experimenting with tools, but not yet rethinking how software gets built end to end, even though the greatest value comes from this broader end-to-end transformation.
But redesigning workflows is only part of the equation.
“The companies getting the most out of agentic development are the ones with the strongest foundations,” says Sonar CEO Tariq Shaukat. “Agents are more cost efficient and effective when they run on well-structured code. Verification, clean architecture, and close attention to technical debt aren’t taxes on speed; they’re what makes speed sustainable.”

The companies getting the most out of agentic development are the ones with the strongest foundations.
Tariq ShaukatSonar CEO
Reimagining the PDLC with AI
To operationalize Horizon 2 in a real-world setting, we partnered with Sonar, a leading AI code verification and governance platform used by more than seven million developers and its AI agents worldwide. Sonar analyzes over 750 billion lines of code daily and has already reached a high level of AI maturity, with strong adoption of coding assistants and foundational enablers in place.
Building on this foundation, Sonar developed its Agent Centric Development Cycle (AC/DC)—an AI-native framework with four continuous steps: guiding agents with the right context, generating code, verifying quality and security, and resolving issues through automated feedback loops.
McKinsey worked with three front-runner teams at Sonar to redesign its PDLC to be AI-native and to pilot the new approach in practice. Together, teams identified where agents could be integrated, which activities could be streamlined, and how roles would evolve as agents took on a greater share of execution.

What distinguished this effort was the focus on the operating model—not just the tools.
Martin HarryssonMcKinsey senior partner
“What distinguished this effort was the focus on the operating model—not just the tools,” says Martin Harrysson, a McKinsey senior partner. “The teams defined how agents would be used, how work would be structured, and how people would interact with AI on a day-to-day basis.”
What changed
The AI-native PDLC drove changes across three dimensions: process, tooling, and people.



Process
Teams redesigned workflows, artifacts, and ceremonies around agents. They introduced dedicated time to build and refine agents, including an “AI experimentation hour,” and created structured sharing forums such as “show and tell” sessions to avoid silos and scale learning across teams.
“People are not only creating more, but also learning from each other, which is driving increased AI usage week over week as skepticism declines and the value becomes clearer,” says an engineering manager at Sonar.
Teams also aligned on standardized templates for key artifacts—for example, requiring every epic to define the target user, the job to be done, the pain point, and the definition of finished. Agents were then built to generate artifacts following these templates, creating a flywheel: Clearer inputs led to better outputs, improving consistency and handoffs across the PDLC.
Tooling
Teams built and deployed agents across the life cycle—from research and ideation to development and testing. In research, agents synthesized large volumes of data—from interview transcripts to public forums and market signals—surfacing patterns in minutes rather than days.
“We can find insights, start on new topics, and test ideas much faster,” says a product manager at Sonar.
In development, teams deployed agents across the workflow—from backlog definition to unit-test generation to code production. Early on, many backlog items relied on tacit knowledge, limiting what agents could execute. Teams addressed this by creating and refining a backlog generation agent to follow standardized templates, flag missing context, and prompt teams to fill gaps—externalizing tacit knowledge into structured, agent-ready inputs.
These agents were complemented by Sonar capabilities such as Sonar Context Augmentation, which improved code understanding, and the SonarQube Remediation Agent, which suggested fixes for quality issues, reducing rework between agent outputs and human review.
“I started using AI to create Jira tickets directly from discovery with proper formatting. It puts things on paper that would usually stay in my head and makes the work much easier to execute,” says a Sonar software engineer.
Teams went further by testing agent-to-agent handoffs and agentic workflows. For example, a sequence of agents could take a bug report from a collaboration channel, create a Jira ticket, clarify requirements, and generate a draft pull request—collapsing a multistep process into a single automated flow.
People
While most companies focus on selecting the right AI tools, the real differentiator is people: teams moving from adopting AI to actively shaping how it works alongside them.
At Sonar, that shift is visible.
As agents became embedded in workflows, role boundaries blurred, allowing team members to contribute beyond their traditional roles. Developers became more involved in research and ideation, while product managers and designers accelerated prototyping. Across the board, team members took on “agent manager” responsibilities—guiding agents, reviewing outputs, and refining how work gets done.
“At the beginning, people were quite skeptical about the agents, but now they’re actively using them,” says an engineering manager at Sonar. “I saw the transformation with my own eyes—not just in our team, but across the company.”
From productivity gains to new ways of working
By the end of the pilot, teams saw improvements across velocity and quality.
Velocity: Redesigning workflows around AI reduced friction across handoffs and enabled more parallel work across product, engineering, and design, while agents supported execution. This helped teams move faster to beta and incorporate user feedback earlier in the development cycle.
While approaches to measuring velocity vary across organizations and teams, at Sonar, teams focused on pull request throughput and cycle time. Front-runner teams improved pull request throughput by up to 2.2 times and reduced pull request cycle time by up to 3.4 times, with self-reported productivity gains in build ranging from 50 to 80 percent across teams.
Quality: Agents connected to company-wide knowledge, helping teams synthesize distributed information at scale—from market research and community signals to customer feedback and internal knowledge—resulting in sharper, more complete artifacts that better informed road map decisions.
“The agent came up with ideas I hadn’t thought of because it had access to so much information,” says a Sonar product manager. “There were still things I needed to refine, but in some areas, it made the feature definition more thorough than what I would’ve done on my own.”
Although not all improvements could be attributed solely to the pilot, the broader signal was clear: Teams were generating better outcomes from similar inputs without sacrificing velocity.

Most organizations are still applying AI to isolated steps in the development process. The real value comes when you rethink the system end to end—how work flows, how decisions are made, and how teams interact with AI.
Prakhar DixitMcKinsey partner
“Most organizations are still applying AI to isolated steps in the development process. The real value comes when you rethink the system end to end—how work flows, how decisions are made, and how teams interact with AI,” says Prakhar Dixit, a McKinsey partner.
From Horizon 2 to Horizon 3
While the effort delivered early results, it also clarified what comes next.
From generation to verification: As AI agents dramatically increase the volume of code and artifacts, the bottleneck shifts from generation to verification—ensuring quality, security, and correctness at scale. As output grows faster than manual review can keep pace, organizations will need stronger, automated guardrails; continuous verification embedded in workflows; and tighter feedback loops to maintain trust.
From AI in pockets to agents at scale: Moving from pockets of impact to organization-wide adoption requires putting the right structures in place—such as bottom-up and top-down change management and shared libraries for accessing and reusing AI agents, among others.
From individual agents to orchestrating agent systems: Looking ahead, the next frontier is end-to-end agentic workflows. Instead of manually coordinating individual agents, teams define outcomes—such as shipping a feature—and agents coordinate execution end to end. This shifts the human role from managing tasks to orchestrating systems, guiding intent, and supervising outcomes while agents handle execution.
What this means for organizations
The shift to AI-native development is not a question of if, but how quickly organizations can adapt. As this case illustrates, AI’s value does not come from isolated use cases, but from rethinking how work gets done end to end—across process, tooling, and people.
Organizations that move beyond experimentation to redesign their PDLC around AI will unlock disproportionate gains not only in productivity, but in how quickly they learn, iterate, and deliver value to users. Those that fail to do so risk optimizing the past, while others redefine the future.
THE TEAM




