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

Fitzpatrick explains the massive disconnect between exponential model performance improvements and minimal enterprise adoption (only 5% of Gen AI deployments work). He argues that enterprise deployment requires data infrastructure, workflow redesign, trust frameworks, and rigorous testing - not just better models. This adoption curve will take a decade, not two years.

• •MIT report shows only 5% of Gen AI enterprise deployments are working, with 40% likely to be canceled by 2027
• •Consumer adoption hit 60% weekly usage while enterprise lags due to trust, observability, and validation requirements
• •Enterprise needs same model risk management processes banks use for credit models - testing, training, validation
• •External vendor builds are 2x as effective as internal team builds according to MIT research
• •$25M agent project at major retailer failed because they built custom eval tools that missed hallucinations giving away millions

Fitzpatrick makes the case that enterprise AI cannot work without forward-deployed engineers (FDEs), contrasting with the failed 'out-of-the-box SaaS' approach. He explains Invisible's model of not charging for FDEs, instead proving value through 2-3 month solution sprints before any payment, and why this approach wins against the traditional Accenture paradigm.

• •Invisible operates 450 FDEs across 8 offices globally, focusing on hyper-personalized software builds
• •Does not charge for FDE time - only charges when software is proven working after user acceptance testing
• •Typical deployment takes 2-3 months vs. Accenture's 2-3 years for similar scope
• •Forward-deployed engineering done well means executing specific workflow builds, not just solutions engineering
• •Economics work by spending less on sellers and more on FDEs who prove value before any contract

Discussion of how traditional SaaS business models don't work for Gen AI due to the nondeterministic nature of AI systems. Fitzpatrick explains why 'pay when it works' models are replacing subscription pricing, and how this mirrors the machine learning deployment paradigm from a decade ago rather than traditional software.

• •Out-of-the-box SaaS was always partially a lie - major enterprise software companies derive significant revenue from services
• •Gen AI requires hyper-specific customization per customer, making it fundamentally different from SaaS
• •Payment happens after 2-3 months when passing user acceptance testing, not upfront subscriptions
• •Nondeterministic AI systems create more risk in 'fake it till you make it' approaches than deterministic software
• •AWS report shows 70% of 'AI agents' are actually just traditional script writing and automation

Fitzpatrick addresses the talent marketplace business model, explaining how Invisible manages 1.3M experts annually and why human-in-the-loop data will remain critical for the next decade. He discusses revenue concentration with major model builders, pricing dynamics, and why data is not commoditized despite common perception.

• •Manages 1.3M experts with 26K active weekly, requiring ability to source PhD in astrophysics from Oxford in 24 hours
• •Revenue was materially concentrated in AI training in 2024, but rapidly diversifying into enterprise (12 deals in 45 days)
• •Customers willing to pay premium for quality data - one week of bad data burns enormous compute costs
• •Human feedback essential for multi-step reasoning, multimodal tasks, and 45+ language contexts for next decade
• •Institutional memory in digital assembly lines creates competitive moat similar to Toyota production system

Evolution of data labeling from commodity 'cat-dog' work five years ago to highly specialized expertise today. Fitzpatrick explains how the market has moved to incredibly niche requirements and why this specialization creates sustainable competitive advantages through institutional knowledge.

• •Market evolved from Google Sheets and cat-dog labeling to digital assembly lines requiring specific expertise
• •Example: Must validate architectural expert on 17th century French architecture who speaks French on 24-hour notice
• •Five years of data on expert performance creates institutional memory for selection and assessment
• •Pricing works like Uber - dynamic based on market context, location, and specific expertise required
• •Expertise needed varies so much month-to-month that bottling up finite supply doesn't create lasting advantage

Concrete examples of Invisible's enterprise deployments, including fine-tuning models for US Navy underwater drone swarms and building complete tech backbones for LifespanMD's concierge medicine business. Demonstrates the breadth of applications and technical complexity involved.

• •US Navy project: Fine-tuning models for underwater drone swarms to process sensor data and make engagement decisions
• •LifespanMD: Building entire tech backbone integrating fragmented EHR, ERP, genomics, wearables data in 2-3 months
• •Five modular platforms: Neuron (data), Axon (agents), Atomic (process builder), Meridian (expert marketplace), Synapse (evaluation)
• •Serving food & beverage, public sector, asset management, agriculture, sports, oil & gas with same architecture
• •Agricultural use case: Monitoring herd safety for largest ag conglomerates using forward-deployed engineers and Starlink

Fitzpatrick explains the strategic decision to move from profitability to aggressive investment after raising $130M (initially announced $150M). He discusses the tension between harvesting capital versus investing for 10-20 year enterprise value creation in the greatest growth environment that has ever existed.

• •Company was profitable having raised only $7M in nine years, now investing heavily after $130M raise
• •Decision driven by belief this is 'greatest environment for growth that has ever existed'
• •Not investing in physical world interactions yet - example: sending FDEs to farms with Starlink for computer vision
• •Spending 70% of time on road building brand after company had only one public article when he joined
• •Focus on building for 10-20 years, not harvesting - 'hope we never get to harvest stage, definitely not now'

Transition from fully remote for nine years to eight offices globally. Fitzpatrick shares data on exponential productivity increases, particularly for engineering teams, and why younger engineers especially wanted colocation. Discusses the balance between office culture and work flexibility.

• •Moved from fully remote to offices in NYC, SF, London, Paris, Poland, DC, Austin after CEO transition
• •Engineering productivity increased exponentially - tripled team size with most wanting in-person work
• •London office has 40 people who voluntarily come in even on Fridays due to peer collaboration benefits
• •Does not mandate office attendance but sees huge voluntary adoption for culture and problem-solving
• •Separates physical colocation from work hours - flexible on weekends/evenings but values in-person collaboration

Counterintuitive take on strategic planning in AI, arguing that the entire world changes every three months, making traditional long-term strategy less useful. Fitzpatrick advocates for core beliefs with 30-40% constant iteration based on new technology releases.

• •Contrasts with biotech where seven-year capital cycles make strategy critical
• •Every three months entire AI landscape changes - must be comfortable with constant iteration
• •Build interoperable frameworks that can integrate new tech as it becomes obsolete
• •Five-year planning useful for culture and institutional memory, not specific tech roadmaps
• •New voice agents and tools require constant incorporation - can't survive without quick reaction

Fitzpatrick's contrarian investment view: rather than betting on AI agent software companies, invest in AI-native businesses that serve customer needs directly. He highlights Y Combinator's recent class having 2x the revenue of any prior class, driven by businesses using AI rather than selling AI tools.

• •Most interesting opportunities are AI-native services businesses, not AI agent software companies
• •Question is whether new AI-native companies get distribution faster than incumbents adopt AI
• •Examples: AI-native tax accountants, loan servicing, banking alternatives like Revolut
• •Y Combinator's recent class has 2x revenue of any prior class - businesses serving needs, not selling software
• •Application layer is tricky - many AI apps are 'nice to have' without full workflow embedding