McKinsey & Company relies on Lilli, a proprietary generative AI platform, as its central cognitive interface. This tool is designed for internal use by its workforce and is built on a model-agnostic architecture that integrates large language models from partners such as OpenAI, Anthropic, and Cohere. It serves as a bridge between the firm’s vast internal knowledge base and real-time strategic analysis.
The role of Lilli in daily consulting
Lilli is not a standalone model but a comprehensive interface that allows consultants to query the firm’s intellectual property. It synthesizes insights from over 100,000 internal documents and past case studies, significantly reducing the time required for initial research. The platform acts as a cognitive aid, enabling staff to access relevant historical data and market intelligence quickly.
The system is strictly internal. Clients benefit from the insights generated through this platform, but they do not have direct access to the Lilli interface. This structure ensures that the firm’s proprietary methodologies and client data remain within a controlled environment. By automating the retrieval and synthesis of information, Lilli allows consultants to focus more on high-level strategic judgment rather than data gathering.
Multi-model architecture and partnerships
McKinsey adopts a multi-provider strategy to avoid dependency on a single technology vendor. While OpenAI is a primary partner, the firm also utilizes Anthropic’s Claude for specific reasoning tasks and Cohere for enterprise-grade deployments. This approach allows the firm to select the most effective model for different analytical needs.
The architecture is designed to be flexible. Depending on the task, Lilli may route queries to different underlying models. This diversity ensures that the firm can leverage the specific strengths of each partner, such as long-form reasoning or structured data processing. It also helps mitigate the risks associated with any single model’s limitations or biases.
Integration with QuantumBlack
For tasks that go beyond generative text, McKinsey leverages its QuantumBlack division. This team focuses on advanced analytics, predictive modeling, and custom software development. They often use Python-based environments and specialized libraries to build bespoke solutions for complex problems.
QuantumBlack handles the heavy lifting of machine learning, such as forecasting supply chain disruptions or optimizing industrial processes. Their work complements the generative AI capabilities of Lilli by providing rigorous, math-driven insights. This combination ensures that the firm can address both qualitative strategy questions and quantitative operational challenges.
Data privacy and security measures
Protecting client confidentiality is a core priority for McKinsey. The firm uses heavily siloed and secure instances of large language models to ensure that proprietary data does not leak into public training pools. This segregation is critical for maintaining the trust of clients who share sensitive financial and strategic information.
The technical setup includes robust firewalls and access controls. Data used for inference is processed in a way that prevents it from being used to retrain or improve the base models of the partners. This “air-gapped” approach treats the AI environment as a fortified digital vault, safeguarding the intellectual property of both McKinsey and its clients.
Accuracy and human oversight
AI outputs are treated as draft-level intelligence rather than final answers. McKinsey employs a human-in-the-loop framework where every insight generated by Lilli must be verified by a consultant. This process involves cross-referencing AI-generated data with primary research and internal benchmarks.
The firm also invests in technical audits and de-biasing algorithms to minimize errors and systemic biases. Consultants are trained to evaluate the context and relevance of AI suggestions. This hybrid approach ensures that the final strategic recommendations are grounded in verified facts and professional judgment, rather than relying solely on automated outputs.
Limited public availability
Lilli is not available for public or external client use. It is an internal tool designed to enhance the productivity of McKinsey’s employees. While the firm may discuss its AI capabilities in general terms, the specific platform and its underlying data are not accessible outside the organization. This exclusivity helps maintain the competitive advantage derived from the firm’s accumulated expertise and internal knowledge base.