Episode Summary
Executive Summary: Harry Stebbings interviews Matt Fitzpatrick, CEO of Invisible Technologies, about why enterprise AI adoption lags model capability. Fitzpatrick argues that successful AI deployments require clean data, workflow redesign, business ownership, and human-in-the-loop validation—not just out-of-the-box agents. He also explains Invisible’s modular platform, forward-deployed engineering model, enterprise growth, and why he’s investing aggressively instead of optimizing for near-term profitability.
Main Topics: Matt Fitzpatrick’s path from McKinsey to Invisible CEO: Fitzpatrick explains how 12 years at McKinsey, leading QuantumBlack Labs and large-scale AI builds, prepared him to run Invisible. The move was driven by a desire to build in AI at a time of rapid market change and growth. The gap between model performance and enterprise adoption: He argues that while models and consumer usage have improved rapidly, enterprise adoption remains limited because companies need more than models: data infrastructure, process redesign, trust, observability, and validation. Why enterprise AI needs forward-deployed engineering: Fitzpatrick says workflows that change how a business operates cannot be solved by generic SaaS. Invisible uses small, embedded engineering teams to configure customer-specific solutions and prove value before charging. Invisible’s modular platform and business model: He describes Invisible’s stack—Neuron, Axon, Atomic, Meridian Expert Marketplace, and Synapse—as a configurable platform spanning data integration, agents, workflow building, expert sourcing, and evaluation, serving both AI training and enterprise deployments. Data, human feedback, and the future of training: Fitzpatrick rejects the idea that synthetic data will replace humans. He believes human feedback, niche expertise, and statistically validated labeling will remain essential for years, especially for specialized reasoning and enterprise use cases. Pricing, revenue, and competitive dynamics: He discusses revenue concentration, pricing discipline, and why data quality commands payment. He argues the market will likely support multiple winners rather than a single dominant player, with specialization across tasks and sectors. Leadership, hiring, and operating culture in AI: Fitzpatrick emphasizes recruiting exceptional people, flattening hierarchy, co-location, and letting teams at the edge make decisions. He also says AI’s fast-changing landscape makes rigid long-term strategy less useful than continuous iteration.
Key Arguments: Enterprise AI adoption is lagging model progress because implementation depends on data, workflow redesign, trust, and governance, not just model quality. Internal enterprise builds are often less effective than external/vendor-led deployments because teams lack discipline, talent density, and clear ROI accountability. Forward-deployed engineering is essential for workflow-changing products; out-of-the-box SaaS works for information repositories, not for transforming operations. Invisible’s competitive edge comes from modular software plus human expertise, enabling hyper-specific enterprise deployments in weeks to months rather than years. Human feedback and specialized experts will remain critical because many enterprise and multimodal tasks cannot be validated with synthetic data alone. AI training is becoming more specialized and niche; public benchmarks matter, but enterprise success depends on specific task-level precision and validation. Pricing power in this space comes from quality and performance, not scarcity alone; customers will pay for data that improves model outcomes and reduces compute waste. The market likely supports several players rather than a single winner, with specialization by task, vertical, and modality. The best AI businesses may be AI-native services or businesses using AI to transform physical-world operations, not just software sold as an agent layer. Management in this environment requires empowering local teams, reducing hierarchy, and accepting that strategy and product direction must evolve quickly.
Data Points: McKinsey engineers at start of Fitzpatrick’s tenure: about 100 - He said McKinsey had roughly 100 engineers total when it first began hiring engineers. McKinsey engineers by the time he left: 7,000 - He oversaw the growth of the engineering group over about a decade. Share of McKinsey engineering group he oversaw: about 1/5 - He led approximately one-fifth of the engineering organization. Model performance improvement: 40% to 60% - He cited public benchmarks showing model performance gains over the last two years. Weekly consumer GenAI usage: 60% - He referenced a KPMG report saying 60% of consumers use GenAI weekly. Enterprise GenAI deployments working: 5% - He cited a MIT report stating only 5% of enterprise GenAI deployments are working in any form. Enterprise projects likely canceled by 2027: 40% - He referenced Gartner saying 40% of enterprise projects may be canceled by 2027. Out-of-the-box agent accuracy: 58% single-turn / 33% multi-turn - He cited Salesforce AI Research results for enterprise agents on workflow tests. Enterprise deals confirmed recently: 12 deals in 45 days - He said Invisible had confirmed 12 enterprise deals in the last 45 days. Invisible funding: $130 million - He said the company had announced $100 million initially and had now raised $130 million total. Prior primary capital raised before this round: $7 million - He said Invisible had only raised $7 million of primary capital in its first nine years. Company size: 450 people - He said Invisible now has eight offices in eight cities and 450 employees. Number of offices/cities: 8 offices, 8 cities - He described the company’s distributed physical presence. Expert marketplace scale: 1.3 million experts a year - He said Invisible sources experts through a large marketplace and selects from them weekly. Weekly active selected experts: 26,000 - He said 26,000 experts are selected each week to start within 24 hours and produce data. Typical forward-deployed engineering deployment time: 2 to 3 months - He contrasted Invisible’s deployment speed with traditional consulting timelines. Typical FDE motion duration at Invisible: about 3 months - He said their FDE engagements are usually around three months. Agriculture example deployment: Starlink terminals deployed to farms - He cited herd safety work requiring physical-world infrastructure. US healthcare spend per capita: $14,000 per year - He used this to argue AI can materially improve healthcare economics. Avoidable deaths from errors: 250,000 per year - He cited Johns Hopkins on avoidable healthcare errors. Data center electricity usage: about 1% of total global electricity - Used to argue AI’s environmental footprint is currently smaller than perceived. AI data center electricity share: 0.25% to 0.5% of total global electricity - He separated AI-specific usage from broader data center consumption. Cooling/air conditioning electricity usage: 14% to 20% of global electricity - He compared AI load to existing energy usage. Bank tech maintenance spend: 93% maintenance / 7% new builds - He used banks as an example of legacy technology constraints. Bank KYC workforce example: 6,500 people - He cited a bank with 6,500 people in KYC alone. Office travel cadence: 60% of the time on the road - He described the demands of leading a global enterprise AI business. Revenue concentration in data providers: two players make up more than 50% of revenue for many providers - He discussed concentration dynamics in the AI data market.
Pivotal Quotes: "the gap between model performance and adoption" — Harry Stebbings: Introduces the core thesis of the episode and the main tension discussed throughout. "I think the simple advice I give, and by the way, this is how we sell, is start with proof of concepts, start with what we call solution sprints. Don't pay a dollar until you prove the tech works." — Matt Fitzpatrick: Explains Invisible’s sales philosophy and why customer validation comes before payment. "I think the biggest misconception now: I think many people were hoping the adoption curve will be: I buy something and I just push it in my business, and it takes a whole process." — Matt Fitzpatrick: Summarizes his view that enterprise AI requires redesign, not plug-and-play adoption.
Implications: Enterprise AI winners will be those that combine great models with data, workflow integration, validation, and human expertise. Buyers should demand proof, clear KPIs, and business ownership. For builders, hyper-specific, vertically embedded solutions may outperform generic agents.