Despite adopting AI tools for software development, many companies are failing to achieve expected results because they've simply added tools while keeping their existing Agile methodologies and operating models intact. McKinsey emphasizes that dramatic productivity gains require shrinking team sizes from 10-person units to smaller 3-5 person "One-pizza pods," along with redefining AI-native workflows and roles. Ultimately, successful AI adoption is not just about technology -- it demands people-centered innovation that fundamentally transforms workforce operations and organizational culture, accompanied by systematic performance measurement.
1. The Reality of AI Adoption and Its Bottlenecks: Why Is Productivity Stagnating?
Today we're talking about the future of software development from McKinsey's Software X organization. Looking back over the past few decades, every major technology shift has brought significant changes to how we develop software. Just as we transitioned to Agile about 20 years ago with kanban boards and standup meetings, today's AI revolution heralds yet another massive paradigm shift.
As of 2025, many developers are using AI agents to accomplish in minutes what used to take days. Individual productivity tools are clearly delivering impressive results. But ironically, at the enterprise level, the impact often falls short of expectations.
"We surveyed about 300 companies. When we asked how much improvement companies adopting AI are seeing on average, most reported only 5%, 10%, 15% improvement company-wide. There is a clear gap between AI's enormous potential and the actual reality."
Why does this gap exist? The problem is that we've installed a new engine called AI, but our ways of working remain stuck in the past. Code generation speed has become incredibly fast, but collaboration methods and review processes between people can't keep up.
"We've started generating far more code, but many companies are still reviewing code in a very manual way. (...) As a result, while AI-automated portions have increased, new bottlenecks like manual reviews have actually emerged."
2. The Secret of the Top 1%: Transitioning to AI-Native
According to McKinsey's research, the top-performing companies didn't just adopt tools. They rewired the entire product development lifecycle (PDLC). These companies were 7 times more likely to have AI-native workflows and 6 times more likely to have defined AI-native roles.
So what specifically makes them different?
First, they flexibly adjust their operating model based on the nature of the work. For example, legacy code modernization uses a "Factory of agents" model, minimizing human involvement since deliverables are clearly defined. In contrast, new feature development uses an "Iterative loop" model where agents and people continuously exchange feedback as co-creators.
Significant changes are also happening in planning and roles.
"Companies with AI-native workflows are moving from quarterly planning to continuous planning. The unit of work is also shifting from 'story'-centric to 'spec'-centric development. Product managers are iteratively refining specs with agents instead of writing lengthy PRDs."
From Two-Pizza Teams to One-Pizza Pods
The most noticeable change is team size. Even Amazon's Jeff Bezos's "two-pizza team" (about 8-10 people) is now too large.
"On the talent side, AI-native roles mean moving away from two-pizza team structures to pods of 3-5 people -- 'one-pizza' size. Instead of having separate QA, frontend, and backend engineers, 'product builders' who understand the entire architecture and can direct agents take on integrated roles."
3. Real-World Examples: Changes at Banks and Startups
Can these changes happen in large enterprises too? A recent project with a major global bank confirms it's possible. The bank experimented with reordering traditional Agile ceremonies and redefining the roles of people and agents.
Key change points included:
- Data-driven task assignment: Team leaders used agents to allocate sprint stories based on team velocity and deployment history data.
- Collaborative prototype generation: Before starting development, teams co-created multiple prototypes with agents, including security and observability standards, to prevent downstream rework.
- Workflow-based restructuring: Simple bug-fix teams and new-development teams were separated, while agents analyzed cross-repository impact in the background to reduce debugging time.
"The results of these interventions were remarkable. Agent usage increased over 60x, and deployment speed aligned with business priorities also accelerated. Code merges increased by 51%, and overall efficiency improved significantly."
Engineers also evolved from simple coding (execution) into orchestrators who delegate tasks to agents, while PMs prototyped directly in code without waiting for developers and modified backlogs in real-time based on customer feedback.
4. Scaling Across the Organization: The Key to Change Management
Changing one team and changing an organization of thousands are fundamentally different challenges. Simply announcing "use the new tools" is a recipe for failure. In fact, there were cases where AI tools were deployed early on, but usage spiked briefly before reverting to old ways.
Successful scaling requires change management.
"I usually say change management is about 'getting a million small things right.' The core challenge of scaling is getting 20-30 elements -- communication, incentive structures, upskilling, and more -- to work correctly at the same time."
Successful companies took the following measures:
- Hands-on training: Not just observation-based training, but code labs where people bring 'their own code' to practice with.
- Certification programs: Certification systems that prove proficiency in new methods, providing motivation.
- Intensive early support: Coaches providing close support during the first few sprints before new habits are formed.
5. What Should We Measure?
The final critical piece is measurement. Lower-performing companies weren't even measuring velocity, or at best, only about 10% were measuring productivity. But achieving results requires a comprehensive measurement system that goes beyond simple 'tool adoption rates.'
McKinsey proposes systematic measurement following the chain of Inputs -> Outputs -> Economic Outcomes.
- Inputs: Not just tool costs, but time and resources invested in training and change management.
- Outputs: Beyond speed and capacity increases, developer satisfaction (NPS) and code security and resiliency.
- Economic Outcomes: Faster time to revenue targets, price differentiation for high-quality features, labor cost savings, etc.
"Many organizations simply think that increased AI tool adoption will lead to faster speed. But it's also important to understand whether developers are enjoying their work more (NPS) or feeling frustrated. You must also verify whether code has become more secure and whether bug resolution time (MTTR) has decreased, improving resiliency."
Conclusion: Start Now
We can't perfectly predict the future, but within the next five years, the software development model will move toward 'shorter sprints, smaller teams, but far more teams.' Even as technology advances and agents become smarter, the importance of this new human-led collaboration model will remain unchanged.
This isn't merely a technology adoption -- it's a massive human change. It will be a journey that takes time, so you need to start right now.
"The key is this: Start now. As we always tell our clients, this is a people problem. It takes time and requires significant change. Find the right model for your organization and set bold ambitions."