20VC: Enterprises Will Not Adopt AI without Forward-Deployed Engineers | Who Wins the Data Labelling Race: How Does it Shake Out? | Lessons Learned Hitting $200M ARR with Matt Fitzpatrick, CEO of Invisible Technologies
Timestamps are as accurate as they can be but may be slightly off. We encourage you to listen to the full context.
Podcast Summary
In this episode, Matt Fitzpatrick, CEO of Invisible Technologies, discusses the massive gap between AI model performance and enterprise adoption. (04:40) While AI models have improved 40-60% in performance over the past two years and consumer adoption has reached 60% weekly usage, only 5% of enterprise AI deployments are actually working according to MIT research. Fitzpatrick, who transitioned from leading McKinsey's QuantumBlack Labs to running Invisible, explains why enterprise AI adoption requires forward-deployed engineers, custom implementations, and fundamental workflow redesign rather than out-of-the-box solutions. (09:35) The conversation explores the future of AI training, data specialization, and why human feedback will remain critical for the next decade despite synthetic data advances.
• The episode centers on bridging the enterprise AI adoption gap through custom implementation, the evolution of data training markets, and building sustainable AI businesses beyond the hype cycle
Speakers
Matt Fitzpatrick
Matt Fitzpatrick is the CEO of Invisible Technologies, leading the company's mission to make AI work across enterprise contexts. Since joining as CEO in January 2025, he has raised over $100M for the company and achieved the milestone of over $200M in annual recurring revenue. Previously, Matt was a Senior Partner at McKinsey for twelve years, where he led QuantumBlack Labs, the firm's AI R&D and software development arm, overseeing about 1,400 of the firm's 7,000 engineers and building enterprise AI solutions across Fortune 1,000 companies.
Key Takeaways
Enterprise AI Requires Custom Implementation, Not Out-of-Box Solutions
Fitzpatrick emphasizes that successful enterprise AI deployment cannot rely on generic, out-of-the-box solutions. (15:26) Unlike consumer adoption which has reached 60% weekly usage, enterprise AI faces fundamental challenges around data infrastructure, workflow redesign, and trust validation. The MIT research showing only 5% of enterprise AI deployments working demonstrates that enterprises need hyper-specific customization using their own data and workflows. This requires forward-deployed engineers who can configure modular platforms to specific enterprise contexts, similar to how credit models in banking require extensive model risk management and validation processes.
Data Specialization is Creating Sustainable Competitive Advantages
The AI training market has evolved from simple "cat-dog" labeling to requiring incredibly specialized expertise. (38:01) Fitzpatrick explains that modern AI training now requires sourcing experts like "a PhD in seventeenth century French architecture who speaks French" on 24-hour notice. This specialization creates institutional memory advantages similar to Toyota's production system - knowledge that competitors cannot easily replicate despite knowing the processes. The ability to manage 1.3 million experts and consistently produce statistically validated data creates a moat that extends beyond simple talent marketplaces.
Forward-Deployed Engineering is Essential for Enterprise AI Success
Traditional SaaS models fail in enterprise AI contexts because AI implementations require ongoing customization and fine-tuning. (18:28) Fitzpatrick advocates for forward-deployed engineering teams that work on-site with clients for 2-3 month implementations, proving technology works before charging. This approach differs from the Accenture model of multi-year, expensive integrations. Instead of charging for consulting time, Invisible absorbs FTE costs upfront and only charges when software passes user acceptance testing, creating aligned incentives and proving value before payment.
Human Feedback Will Remain Critical Despite Synthetic Data Advances
Contrary to predictions that synthetic data will replace human input, Fitzpatrick argues human feedback will be essential for the next decade. (49:56) Synthetic data works well for mathematical problems with clear right/wrong answers, but complex reasoning tasks requiring PhD-level expertise across multiple languages and contexts cannot be effectively trained synthetically. Legal services, for example, rely on proprietary data within law firms that isn't publicly available. Enterprise fine-tuning requires human expertise for statistical validation, especially for multimodal, multilingual, and highly specialized applications.
Pricing Models Must Shift from SaaS to Outcome-Based Payment
The traditional SaaS pricing model doesn't work for enterprise AI because implementations are inherently custom and nondeterministic. (25:48) Fitzpatrick advocates for outcome-based pricing where customers "pay as it works" rather than paying upfront for uncertain results. This requires proving technology works through solution sprints and proof-of-concepts before any payment. The model acknowledges that unlike traditional software where "if I deliver a SaaS box, I know it will work," AI agents and workflows require validation that they actually perform as intended, as demonstrated by the retailer who spent $25M building a returns agent that ultimately failed because it lacked proper evaluation frameworks.
Statistics & Facts
- Only 5% of enterprise Gen AI deployments are working according to MIT research, while 60% of consumers use Gen AI weekly (KPMG report). (10:13) This massive gap demonstrates the fundamental difference between consumer and enterprise AI adoption challenges.
- Gartner predicts 40% of enterprise AI projects will likely be canceled by 2027. (10:31) This statistic highlights the widespread struggle enterprises face in successfully implementing AI solutions.
- The average bank spends 93% of its technology budget on maintenance initiatives, with only 7% going to building new capabilities. (58:39) This explains why internal AI development efforts often fail and why banks struggle with innovation adoption.
