McKinsey's AI leader moved to head a $2B AI workforce - Matthew Fitzpatrick [Invisible]

The AI platform bet behind Invisible's $2bn valuation

Invisible Technologies trained large language models for the frontier labs before Matthew Fitzpatrick left McKinsey to run it as a platform for enterprise AI work. Model companies take the headlines, but the durable last-mile work stays with whoever holds a company's unique context and data. Funding and compute are concentrated in a few hands, and the contest is whether new entrants reach distribution before incumbents build the technology.

Takeaways

• Fitzpatrick left McKinsey for an AI platform bet. After twelve years at McKinsey, where he led 1,000 engineers at QuantumBlack Labs, he took over Invisible Technologies in January 2025 and raised $100m at a $2bn valuation. He bet the next AI work would sit on platforms rather than bespoke deployments.
• Invisible sells a modular AI platform to enterprises. The suite includes Neuron for data, Atomic for process building, Axon for agentic work, Meridian as an expert marketplace of 1.1m experts, and Synapse for observability. SwissGear mapped about 750 data tables and doubled the SKUs it could forecast on a six-month lead time.
• Three AI companies absorb 30% of funding. Fitzpatrick says about 100 companies take 70%, a scarcity dynamic, and argues these are not dot-com eyeball businesses because they carry hundreds of millions or billions in revenue. He cites Mary Meeker's 1999 predictions, which underestimated the internet but wrongly expected incumbents to capture it.
• Most announced AI restructures are AI washing. Fitzpatrick invokes the Jevons paradox and cites bank tellers and lawyers, professions that grew despite automation. An NBER study of 60,000 employees found individuals gain 15 to 20% in productivity while companies gain only 1 to 3%.
• A third of agents are basic scripts. Few companies implement multi-agent orchestration well, Fitzpatrick says. Contact centers are the cleanest case for agents because IVR routing is very bad, and other routing problems are better handled by traditional software. Future solutions will mix deterministic software, machine learning, agents and LLMs.
• Data is the biggest impediment in large enterprises. Fitzpatrick says enterprise models are probabilistic and hallucinate, and each company's context will never be in an LLM. He advises picking three or four data initiatives, each with a business owner and a KPI, and using the data for cross-selling and pricing.

What Fitzpatrick predicts

• Enterprise adoption will take north of a decade, with material change in how much work AI shapes still five to ten years away.
• Contact centers and media move fastest, regulated industries slowest. Mortgage underwriting took 15 to 20 years for intake forms, quality control, model risk management and regulatory review to be redesigned.