McKinsey’s Journey in Developing Its Generative AI Platform
McKinsey recently developed a generative AI platform called “Lilli” to utilize its extensive knowledge repository more effectively. This platform is part of McKinsey’s broader attempt to boost internal efficiency and serve clients in new ways. In a recent episode of the “At the Edge” podcast, Erik Roth, a senior partner, along with Lareina Yee, another senior partner, discussed the creation of Lilli, its development process, and the impact it has had internally at McKinsey. While the platform has certainly shown potential, the process of creating and adopting Lilli also provides insights into the challenges of generative AI adoption in professional services.
The Birth of Lilli
Lilli originated from a basic need to help McKinsey consultants access internal knowledge more easily and quickly. The initial idea was straightforward: use generative AI to extract and synthesize McKinsey’s proprietary insights. As AI tools like ChatGPT gained attention, McKinsey decided to train an internal tool on its data. Lilli started as a relatively simple knowledge extraction tool but evolved into a more complex system that integrates multiple types of technologies. This evolution transformed Lilli into what Erik Roth describes as an “orchestration layer” for McKinsey’s knowledge resources.
How Lilli Differs from Other AI Platforms
Unlike some generative AI tools, Lilli was designed to specifically cater to McKinsey’s workflows and the nature of its consulting business. Roth mentions that the platform utilizes both large and small models to understand context-specific questions and provide answers that are more directly relevant to McKinsey’s needs. This customization aims to differentiate Lilli from generic AI tools by making it “fit for purpose.” However, this also means Lilli is highly tailored to a specific context, limiting its utility beyond McKinsey’s internal ecosystem.
Additionally, Lilli is presented as an “orchestration layer” rather than a typical retrieval-augmented generation instance, blending several technologies to meet McKinsey’s internal requirements. This complexity makes it both a unique tool for a specific use case and one that requires significant ongoing resources to maintain and adapt.
User-Centric Development with Limitations
The development of Lilli was user-centric from the start. It began with a small team and expanded as McKinsey learned more from internal feedback. Initial development involved ethnographic studies to understand what users needed most. The team focused on four main domains: team building, enhancing client development with technology, improving client service, and maintaining connections after project completion. Although these domains have provided a clear framework for development, it is clear that user feedback has shaped Lilli in an ongoing iterative process, with successes often accompanied by limitations and challenges that needed addressing.
McKinsey’s approach emphasized keeping users at the center of the development process, ensuring that each new feature or tool addressed a specific internal pain point. While this approach helped create a tool that is highly relevant to McKinsey consultants, the narrow focus may limit broader applicability and scalability. The iterative nature of Lilli’s development also highlights the slow, resource-intensive process of building effective internal AI systems.
Adoption Challenges and Leadership Strategies
Driving Lilli’s adoption across McKinsey has not been without its challenges. Roth acknowledged that encouraging widespread use of Lilli required proactive effort, including role modeling by senior leaders. Roth himself frequently asked team members, “Have you used Lilli today?” to reinforce its importance. This sort of top-down encouragement was supplemented with structured training and local user groups to help build a community of practice around Lilli.
Despite these efforts, challenges remain. The platform’s integration into daily workflows necessitated ongoing user education, and even with consistent advocacy from leadership, not everyone adapted at the same pace. Roth’s emphasis on asking colleagues about their usage of Lilli underscores a significant cultural barrier to technology adoption—having an AI platform is not enough; there must be a concerted effort to change habits and workflows.
Unexpected Features and Their Reception
One notable feature that emerged during Lilli’s development was the “McKinsey Tone of Voice” agent, which helps users craft written content in a style consistent with the firm’s standards. According to Roth, this feature was particularly popular among non-native English speakers and within the beta testing group. While this feature has been well-received, it also highlights some limitations inherent to Lilli’s use—namely, that a significant use case for the platform involves cleaning up internal communications, which may not reflect transformative client outcomes.
This points to a broader challenge with generative AI adoption in professional services: the technology is still evolving, and many of the most immediately practical uses involve augmenting basic administrative functions rather than fundamentally transforming consulting work.
The Changing Role of Consultants
Roth also provided his perspective on how Lilli could change the nature of consulting at McKinsey in the future. He suggests that consultants may increasingly rely on AI to handle analytical tasks, freeing them to focus more on activating insights rather than generating them. This shift could make consulting more accessible to individuals with diverse backgrounds, as the emphasis may shift from purely analytical skills to other capabilities like empathy and creativity.
However, it is important to note that this vision remains speculative. While tools like Lilli might eventually change the skillsets required for consulting, the practical impact of these changes has yet to be fully realized. As of now, much of Lilli’s impact appears to be incremental rather than revolutionary.
Lessons for Other Organizations
One key insight from McKinsey’s development of Lilli is the critical importance of data quality and architecture. As Roth pointed out, generative AI tools are only as good as the data they are trained on. For companies considering similar tools, a clear data strategy—including tagging, labeling, and curation—is essential. Furthermore, while the benefits of AI tools are often touted, the challenges of building, adopting, and maintaining these systems are considerable. McKinsey’s experience suggests that organizations should not underestimate the level of effort required to make such tools effective.
Roth also emphasized curiosity and education as necessary components for organizations trying to adopt AI tools. Leaders need to be willing to learn and experiment, despite the steep learning curve and potential setbacks. The journey of adopting generative AI is not just about implementing technology but also about fostering a culture that embraces change and continuous learning.
A Cautious Path Forward
The development of Lilli has been a balancing act between speed and careful learning. McKinsey opted for a gradual rollout, starting with a small group and scaling slowly to allow for refinement based on feedback. This cautious approach reflects the inherent uncertainty in generative AI technologies. While Lilli shows promise, its real-world impact is still emerging, and its effectiveness will depend on continued iteration and adaptation.
Roth concluded with a cautiously optimistic outlook for the future of AI in consulting. He suggested that AI has the potential to change the way consulting firms operate, provided it is used responsibly and with an awareness of its limitations, including issues around bias and data privacy. However, the full transformative potential of AI tools like Lilli remains to be seen, and their success will hinge on how well they are integrated into existing workflows without disrupting the core value that human consultants bring.
Conclusion
Lilli represents McKinsey’s attempt to modernize its internal processes through generative AI, and while it offers an interesting case study, it also reveals the complexities involved in implementing such technologies within a large organization. The platform’s development and adoption were shaped by the need to balance technological capabilities with practical user needs. Lilli’s journey underscores that while generative AI holds promise, its real impact lies in careful integration, realistic expectations, and an ongoing commitment to user-centric development. McKinsey’s experience serves as a reminder that the path to effective AI adoption is a long one, requiring not just technical innovation but also cultural and organizational change.
Source: McKinsey Digital
