McKinsey's AI boss jumped ship to run a $2B AI's workforce - Matthew Fitzpatrick [Invisible]

Matthew Fitzpatrick, CEO of Invisible Technologies, discusses the challenges and opportunities in enterprise AI adoption. He highlights that data is often the biggest impediment, and the next generation of AI will be anchored in coherent platforms for production deployment. Fitzpatrick transitioned from McKinsey, where he led the firm's AI R&D arm, Quantum Black Labs, to lead Invisible Technologies, a company specializing in training large language models and building modular AI software platforms. Invisible Technologies offers a suite of four platforms: Neuron for data integration, Atomic for process building, Axon for agent orchestration, and Meridian, an expert marketplace with 1.1 million experts used for reinforcement learning. These platforms are used to build hyper-personalized enterprise software across sectors like food and beverage, healthcare, asset management, and consumer retail. A key example is their work with Swiss Army luggage (SwissGear) to improve inventory forecasting, doubling the number of SKUs they could effectively forecast by rapidly mapping 1750 data tables. The business model is a managed platform approach, not a large upfront services fee. This is crucial because AI models require ongoing maintenance and adaptation, especially as regulations or business processes change. Their pricing is a mix of platform fees, consumption-based fees (e.g., per call in a contact center), and outcome-based fees tied to shared customer success. This contrasts with traditional SaaS models, as it makes the ROI clearer for clients. Fitzpatrick addresses the AI funding market, noting a concentration of investment in a small number of companies, creating a scarcity dynamic. He argues this is not a bubble like the dot-com era, as these companies are generating substantial revenue rapidly, supported by unique scaling laws. The macro-economic environment, with fewer high-return areas, also drives capital into AI. Comparing AI to the internet's transformative impact, Fitzpatrick references Mary Meeker's predictions. While she underestimated the internet's potential, her misjudgment was in predicting established players like Blockbuster would capture the growth, when in reality, new entrants often thrived. AI is expected to follow a similar paradigm, taking over a decade for true enterprise adoption. Currently, only about 5% of GenAI projects reach production in enterprises, while consumer adoption is much higher (around 70% weekly use). The gap between enterprise and consumer AI adoption stems from the need for extreme precision in enterprise use cases like credit underwriting or medical coding, where probabilistic models with potential hallucinations are problematic. This requires a new approach to designing, quality controlling, and testing AI systems, a capability gap that enterprises are still developing. The analogy of mortgage underwriting, which took 15-20 years to automate, illustrates this lengthy process of redesigning workflows and gaining regulatory trust. Fitzpatrick anticipates uneven AI proliferation across sectors. Contact centers and media (especially video creation) are likely to see rapid adoption, while regulated industries like financial services will take longer. Unlike the internet, which is a broad network, AI adoption is more individual to institutions and sectors, leading to varied adoption speeds. He cites a study showing individual productivity gains of 15-20% with AI, but only 1-3% organizational productivity gains, indicating that the fundamental redesign of work is still underway, akin to the automobile requiring factory redesign for mass adoption. Digitally native companies founded recently are accelerating faster than older companies re-engineering themselves. Regarding the impact of AI on jobs, Fitzpatrick dismisses the idea of widespread job losses as "AI washing." He invokes Jevons paradox, suggesting that new technologies often create more usage and thus more jobs, citing the example of bank tellers and lawyers. He believes AI will make it easier to scale technology and software, leading to more new businesses and offerings, shifting work towards creative and entrepreneurial roles rather than manual administrative tasks. He envisions a society with more small, fast-growing companies and a reduction in administrative costs, particularly in sectors like healthcare, leading to higher-value work. Fitzpatrick pushes back on the synthetic data thesis, arguing that while it works for commodity data, the vast matrix of language, culture, and expertise makes human feedback crucial for hyper-specific tasks and complex reasoning. He believes humans will remain in the loop for validation, testing, and fine-tuning AI models, especially in regulated industries. Invisible Technologies recently acquired VCP, a talent intelligence platform with extensive interview and assessment data, to enhance its ability to assess expert capabilities for AI training. For companies with proprietary data, Fitzpatrick advises focusing on business initiatives like cross-selling or pricing intelligence, rather than data monetization. The key is strong business leadership to drive AI initiatives, with clear KPIs and a focused roadmap of 3-4 prioritized initiatives, not an overwhelming number. He concludes by emphasizing that the biggest impediment to AI in enterprises is often not technology but the human element of leadership and organizational change. He recommends following him on LinkedIn and Invisible Technologies' website for updates.