Transforming the software development cycle with generative AI
Incubation and R&D from QuantumBlack, AI by McKinsey
Over the years, QuantumBlack, AI by McKinsey has helped organizations reinvent themselves to achieve accelerated, sustainable, and inclusive growth with AI. The QuantumBlack Labs Incubation, Research, and Development (R&D) team explores emerging technologies. The team works in close collaboration with clients, McKinsey’s practices, and a network of strategic alliances, including hyperscalers, model and tool providers, and AI platforms.
Generative AI (gen AI) is reshaping software development. It promises faster delivery and greater efficiency, but realizing these benefits requires more than just dropping new tools into existing workflows. Success depends on a well-rounded strategy, team preparation, risk management, and process optimization. This article describes a structured approach to the effective integration of gen AI that blends automation with human oversight.
Software developers, engineers, and designers need a comprehensive strategy to incorporate AI assistants into their development cycles yet maintain code quality and security standards. Teams need to set guidelines for AI tool usage, implement verification processes, and upskill developers to use these new capabilities effectively.
The transformation may reshape the way teams work by automating certain tasks and could render some roles obsolete. Managing the transition and addressing the impact on employees is a necessary part of the process of AI transformation.
Insights from traditional development workflows
The R&D team within QuantumBlack Labs listens to our clients and our in-house technical teams. By diving deep into real-world challenges, they discovered that our data scientists were spending considerable time on tasks outside their core expertise, such as UI development. It was like watching Formula One drivers stuck in city traffic, a clear misallocation of specialized talent.
The pain points of traditional development workflows include:
- Limited availability of design and front-end (FE) talent, resulting in long delays.
- Tedious manual workflows for tailoring UI elements to different use cases.
- Delays across all parts of the development stack due to a lack of streamlining for rapid prototyping and testing.
- Time-consuming backend integration requiring extensive manual configuration and debugging.
- Complex data integration that required extensive effort to handle inconsistent formats, ensure data integrity, and maintain scalability.
- Troubleshooting inefficiencies and performance bottlenecks, especially when integrating multiple systems, causing delays and scalability issues.
The reality of gen AI-assisted software development
The landscape of AI-assisted development has evolved rapidly. What began with contextual auto-completion in your favorite IDE, has progressed to tools that enhance full-stack developer productivity. Those auto-complete capabilities have evolved beyond simple syntax completion. New tools now understand complex context and offer meaningful code suggestions that span multiple lines or even entire functions.
Take, for instance, Cursor’s Agentic mode, which represents a significant advancement in contextual coding assistance. The agents actively analyze an entire codebase to offer suggestions and solutions aligned with its existing architecture and coding patterns. Unlike earlier iterations of AI coding tools that worked best with greenfield projects, current solutions can analyze existing code patterns and propose extensions that maintain consistency with established patterns.
While AI tools offer significant potential, they come with important limitations that need to be considered from the start. Technical limitations include:
- Code hallucinations that require careful validation.
- Misalignment of generated code with design specifications.
- Tech stack restrictions due to limited support for specific frameworks only.
- Limited understanding of custom components if they are not covered in training data.
- Generated code is optimized for new development but is less suited to existing projects.
There are also practical challenges. As one of our senior developers pointedly observed, “There are plenty of cool tools out there, but are we just kicking the can down the road?” This perspective highlights a real concern: AI tools can automate many tasks, but often require human intervention at critical moments. If developers are not involved from the start, understanding and fixing issues becomes harder when issues arise, such as:
- The risk of “vibe coding” producing large volumes of code that could be hard to understand or modify if intervention is required.
- The AI code generator struggles with unclear prompts or complex requirements, leading to incorrect code generation.
- Integration challenges with existing infrastructure, such as microservices. This includes compatibility problems between services, API mismatches, data inconsistencies, etc.
- Complications and deployment delays in CI/CD pipelines when handling microservices and distributed systems.
It’s also imperative to perform diligent monitoring and code validation. LLM-generated code may include potential malware or bugs that pose security risks. Vulnerabilities may arise from the unwitting inclusion of malicious code, data leaks, bias, and unpredictable behavior.
Building a seamless AI development process
Successful implementation requires human expertise to integrate AI capabilities and use well-defined checkpoints and review processes. Without a structure, teams risk creating technical debt that eventually outweighs initial time savings. Downstream complications can be prevented by establishing specific handoff points between AI automation and human oversight.
We developed a systematic approach to an AI-assisted development framework through experimentation and real-world application. It emphasizes human oversight and structured processes in software development. Our framework comprises five crucial phases: discovery, backend development, frontend implementation, testing/validation, and deployment/iteration.
Each phase includes well-defined handoff points where human expertise is essential, alongside automated processes where AI can safely accelerate development. It focuses on workflow optimization and ensures clarity when human guidance is needed, particularly during architecture decisions, security reviews, and integration points.
The success of AI-assisted development begins with discovery and planning. During this phase, we define project goals and identify core functionality needs. This preparation is essential to ensure that subsequent technical decisions are aligned.
Next comes the crucial development stage, where we identified two effective paths for AI-assisted development, each suited to different scenarios:
a) Frontend-first approach
For projects where the business and functional logic is still being refined and the final vision is evolving, starting with the frontend often proves more effective. As seen in the diagram below, this process typically involves:
- Creating the frontend with mock data using structured JSON
- Iterating on the interface while refining the data model
- Developing backend services to match the established patterns
This approach helps teams validate concepts quickly and adapt to changing requirements efficiently.
b) Backend-first approach
When requirements are well-defined, starting with backend development provides a solid foundation. As seen below, this approach involves building robust APIs and establishing clear business logic before moving on to interface development.
We use structured API specifications with JSON payloads to enable systematic UI generation. This method works particularly well for teams with clear technical requirements and established data models.
From concept to reality: A one-week success story
We used our AI-assisted framework alongside the selected gen AI-powered coding tool on a project that would typically take 2–3 weeks. It was transformed into a successful one-week accomplishment:
- Days 1–2: Set up APIs and UI specifications
- Day 3: Generated UI components and integrated context engine
- Day 4: Gathered and incorporated feedback
- Day 5: Polished and deployed working prototype
Understanding the constraints of gen AI-supported tools has been a guiding principle in our success. While AI tools can automate many tasks, they often require human intervention at critical moments. Our structured approach uses a robust code review process, where experienced developers assess AI-generated code to ensure quality while also learning the AI’s problem-solving methods.
Detailed security measures must be in place, including dedicated audit protocols for AI-assisted code, as well as clear guidelines for tool usage, covering prompt engineering and documentation. Regular knowledge-sharing sessions encourage the team to discuss effective prompts and learn from each other. Enhanced testing protocols specifically target common AI-generated code issues.
The outcome of AI integration
The integration of AI has transformed our development process in remarkable ways. We’ve seen a dramatic reduction in time to prototype, with days-long tasks now completed in hours. This acceleration has led to improved resource allocation, and enabled team members to focus on high-value activities. Team satisfaction has notably increased, particularly among senior developers who can now concentrate on complex architectural challenges rather than routine tasks.
By shifting standard implementation coding tasks to gen AI tools, our specialists can dedicate their expertise to developing advanced features and innovative solutions that directly enhance our client’s value and competitive advantage.
Design work continues to require significant human input due to the need for nuanced user understanding and aesthetic judgment. By creating the visuals early, it serves to motivate the team and guide their focus, which ensures that they stay aligned with strategic goals throughout the process.
The limitations of the supported underlying programming languages and frameworks can affect the usability of AI tools. This may lead to reduced effectiveness when dealing with custom components or less-common technologies. Complex integrations, especially in microservices architectures, often require more manual intervention than anticipated. Security considerations necessitate strict review protocols, particularly for code handling sensitive data or user information. This overhead is a necessary trade-off for the benefits of AI-assisted development.
We’ve developed a balanced approach through experimentation and real-world application. It maximizes the advantages of AI tools while maintaining the high standards our teams and clients expect. This pragmatic perspective has been key to the successful integration of AI into our development workflow.
Navigating the gen AI-powered development tools landscape
The landscape of AI-assisted software development is evolving rapidly. New tools are continuously pushing the boundaries of coding, automation, and productivity. Below is a non-exhaustive, alphabetical list of some emerging AI-powered development tools to keep an eye on:
Bolt, Cursor, GitHub Copilot, Visual Studio IntelliCode, Lov able, Replit, Tempolabs, v0, and Windsurf.
Each of these tools brings a unique approach to AI-driven development, bringing distinct advantages and trade-offs.
So, developers, dive in! Experiment, explore, push their boundaries, break things, and rebuild them even better. The future of software development belongs to those who embrace innovation, iterate fearlessly, and shape what’s next.
Looking ahead
It’s clear that AI tools aren’t simply adding a new layer to development: they’re fundamentally reshaping how we build software. For experienced developers, gen AI accelerates coding far beyond human speed, and boosts productivity. Still, challenges remain for those without full-stack expertise. Complete automation appears unlikely soon, as certain aspects of development still require human oversight and specialized knowledge.
Crucial development areas benefit from human judgment because AI tools currently augment but don’t fully replace us. This includes architectural decision-making, security vulnerability assessment, complex system integration, performance optimization, and custom business logic implementation. Domain-specific expertise and translation of client requirements also remain areas where human capabilities provide essential value.
AI tools are designed to complement, not replace, human expertise. Although these tools can greatly enhance our abilities, success still depends on a solid foundation in development principles. Teams must continue to provide guidance and ensure quality control. Additionally, as technology continues to evolve, staying adaptable and ready to embrace new developments will be essential.
Our recommendations for development teams considering gen AI support:
- Start with clear processes and guidelines
- Maintain strong security practices
- Keep experienced developers involved in review processes
- Use AI tools strategically, not as a complete solution
- Invest in proper testing and validation infrastructure
Successful AI integration accelerates developer productivity by amplifying human creativity and expertise. You can unlock the potential and deliver high-quality software faster than ever before with a solid foundation and strategically applied AI tools. A fundamental element for success is to find the balance between assistance from gen AI and human input.
As AI models continue to advance in reasoning capabilities and code understanding, we anticipate these tools evolving from code completion assistants to true development partners capable of handling increasingly complex architectural decisions and system design tasks. This rapid evolution will likely redefine developer roles and team structures in the next period of time. We’ll explore these emerging capabilities, organizational impacts, and strategic adoption frameworks in our upcoming articles in this series.
QuantumBlack Labs Research & Development is dedicated to curating and rapidly incubating an innovation pipeline that drives the next wave of technology and AI-enabled solutions. We source candidate projects from our extensive work across enterprise organizations, industry trends, and McKinsey’s specialist Tech Council. A rigorous, battle-tested methodology validates which ideas should graduate into mature assets. We also determine whether to integrate these assets into existing solutions or create new product teams.
To learn more about what QuantumBlack Labs Research and Development can do for you, please email rickard_strom@mckinsey.com.
Thanks to all who contributed to this article: Eva Cigic, Rickard Ström, Jo Stichbury, Rory Walsh, Carlo Giovine, Alberto Mario Pirovano, Lukas Olson & Joanna Sych.
