The Twenty Minute VCMatt Fitzpatrick: Who Wins the Data Labelling Race & Why Al Needs Forward-Deployed Engineers
CHAPTERS
- 0:00 – 0:54
Why enterprise GenAI deployments fail despite rapid model progress
Matt opens with sobering stats on enterprise GenAI: strong model and consumer adoption, but very low production success in enterprises. He argues the gap is not model capability but deployment realities—data infrastructure, workflow redesign, accountability, and trust/observability.
- •MIT report: only a small share of GenAI deployments are working in any form
- •Gartner projection of widespread project cancellations
- •Enterprise adoption depends on data readiness, workflow redesign, and ownership
- •Trust, observability, and validation are prerequisites for production
- •Enterprise rollout is a multi-year journey, not a 1–2 year wave
- 0:54 – 4:30
Matt’s path from McKinsey QuantumBlack to CEO of Invisible
Matt explains his non-traditional McKinsey career leading QuantumBlack Labs and scaling engineering inside the firm. He describes meeting Invisible’s founder Francis Peraza and why the company needed a US-based CEO to reach the next stage.
- •Led McKinsey’s tech development org (QuantumBlack Labs) and major AI builds
- •McKinsey engineering scale-up from ~100 to ~7,000 engineers
- •Met founder Francis via a non-work forum; relationship built over time
- •Decision to leave a stable, intellectually rich role for a CEO seat
- •Motivation: build in AI during a uniquely dynamic era
- 4:30 – 8:36
Decision-making principles: mentors, regret minimization, and values
Harry asks about decision frameworks; Matt emphasizes enjoying the work, the people, and building something durable. He shares how mentor and partner advice quickly clarified the CEO decision as avoiding future regret.
- •Optimizes for meaningful work and strong collaborators over material outcomes
- •Uses reading and a small trusted circle for calibration
- •Mentor advice: the biggest risk is not taking the shot and regretting it
- •Leaving stability can be hard even when the opportunity is compelling
- •Long-term builder mindset: “next two decades” orientation
- 8:36 – 15:29
The biggest barriers to enterprise AI: internal builds, weak discipline, and “science projects”
They dig into why enterprises struggle: internal teams often lack the discipline vendors impose (milestones, ROI, accountability), and talent is scarce. Matt shares an example of a $25M internally built returns agent that failed due to flawed evaluation metrics and hallucination risk.
- •Externally driven builds can outperform internal builds due to accountability and discipline
- •Enterprises often lack elite AI engineering depth; talent pool is limited
- •Example: $25M returns agent shut down after poor evaluation design
- •CFO guardrails will increasingly demand ROI, metrics, and stage gates
- •Advice: focus on 3–4 high-impact initiatives led by business owners (not IT)
- 15:29 – 17:34
How to buy in a noisy market: vendor saturation, agent underperformance, and proof-of-concept sprints
Harry presses on choosing among many contact-center vendors; Matt describes the reality of hundreds of similar pitches and many non-working agent products. Invisible’s approach is to start with no-cost, time-boxed “solution sprints” that prove outcomes before payment.
- •CTOs/COOs face extreme vendor volume with similar-sounding claims
- •Out-of-the-box agent accuracy can be weak, especially on multi-turn workflows
- •Enterprises fear picking a vendor and getting stranded with failure
- •Invisible’s “sell”: free 8-week sprint to prove capability before spending
- •Selection should be outcomes-based: test, measure, fire quickly if it fails
- 17:34 – 20:14
Invisible’s modular enterprise platform: data, agents, workflows, experts, and evals
Matt explains Invisible’s five-component platform and how it is configured per customer. He illustrates with a healthcare concierge example: unify fragmented data, build query/reporting intelligence, fine-tune models, and deploy task-specific agents like scheduling.
- •Five modules: data platform, agent builder, process/workflow builder, expert marketplace, evaluation platform
- •Same modular stack applied across many industries (public sector, oil & gas, agriculture, sports, etc.)
- •Healthcare case: unify EHR/CRM/ERP/notes; enable patient-journey intelligence
- •Conversational interrogation + fine-tuning for complex cohort questions
- •Thesis: shift from generic SaaS to hyper-personalized software per customer
- 20:14 – 27:13
Why forward-deployed engineers (FDEs) are mandatory for enterprise GenAI—and how pricing changes
Matt argues enterprise AI cannot be delivered as pure out-of-the-box SaaS; it requires forward-deployed build teams and continuous tuning. He distinguishes real FDE work (shipping workflow builds) from solutions engineering, and explains a pay-on-success model aligned to acceptance testing.
- •FDEs are essential when AI changes workflows and needs deep embedding
- •“FDE” varies: best version is building hyper-specific workflows on a platform
- •Typical Invisible deployment: ~3 months, then ongoing tuning
- •Invisible doesn’t charge separately for FDEs; customers pay when it works (UAT/validation)
- •GenAI delivery resembles classic enterprise ML engagements more than SaaS
- 27:13 – 30:40
Are AI talent marketplaces dead? Invisible’s model and the mechanics of high-quality AI training
The conversation shifts to the ‘talent marketplace’ label and why it’s incomplete. Matt positions Invisible as an AI training platform: rapidly sourcing niche experts, orchestrating them in a “digital assembly line,” and producing statistically validated data under tight timelines.
- •AI training has multiple models; “marketplace” is an oversimplification
- •Core capability: source any expert quickly and produce validated data at scale
- •Digital assembly line metaphor: speed, quality control, and repeatability matter
- •Thesis: training/fine-tuning expertise translates directly to enterprise deployments
- •Business mix is evolving from training-heavy toward more enterprise growth
- 30:40 – 42:42
Data labeling market outcomes: pricing power, specialization, and why human feedback persists
Matt addresses market concentration, negotiations, and whether data is commoditizing. He argues demand for human feedback expands with task complexity (multimodal, multilingual, domain reasoning) and that institutional memory—not finite supply cornering—drives durable advantage.
- •Market is concentrated because few model builders exist, but enterprise expands customer diversity
- •Good data is worth paying for because bad data wastes expensive compute
- •Specialization trend: from ‘catdog labeling’ to niche expert validation
- •Synthetic data won’t replace human feedback for many contextual tasks (e.g., legal, enterprise-specific)
- •Moat comes from operational institutional memory and quality systems, not monopolizing supply
- 42:42 – 52:56
Benchmarks vs enterprise reality: 99% task precision, model risk management, and long adoption curves
They discuss rapid-fire model releases and public benchmarks. Matt says benchmarks are useful for gauging progress, but enterprise adoption hinges on hyper-specific task performance, rigorous evaluation, and trust frameworks akin to bank model risk management.
- •Public benchmarks show progress, but don’t predict enterprise readiness
- •Enterprises need high precision on bespoke tasks, not generic generalization
- •Firms will build their own task benchmarks (e.g., LBO modeling, IC memos)
- •Validation and governance (model risk management) will become standard
- •Longer adoption horizon reduces fears of immediate talent pipeline collapse
- 52:56 – 55:22
Capital allocation and scaling Invisible: choosing heavy investment over profitability
Matt explains Invisible’s shift from capital-light growth (only $7M raised historically) to raising ~$130M and investing aggressively. He frames it as a return-on-capital decision in a uniquely large growth environment, with more focus on enterprise platform expansion and physical-world data.
- •Invisible historically raised minimal primary capital; now investing heavily
- •Profitability traded for product/platform and go-to-market acceleration
- •Investment thesis: exceptional growth environment + strong positioning
- •Where he wants to invest more: physical-world data acquisition and deployments (farms, oil rigs, robotics)
- •Scaling tension: steady profitable growth vs hyper-scale ambition
- 55:22 – 1:03:34
Brand and trust in enterprise AI: aligning public narrative with real delivery
Matt describes Invisible’s historically low public profile and why brand now matters for trust and awareness. He emphasizes credibility—saying only what can be delivered—especially important with non-deterministic systems where “fake it till you make it” can backfire.
- •Brand matters in enterprise for trust, awareness, and deal velocity
- •Invisible historically relied on customer work over public storytelling
- •Andreessen idea: risk/opportunity when public and private narratives diverge
- •Non-determinism makes overpromising especially dangerous in GenAI
- •Positioning: agents are one tool; real value is end-to-end systems that work
- 1:03:34 – 1:09:24
Recruiting, culture, and execution: hiring ‘athletes,’ flattening hierarchy, and operating at the edge
Matt argues great teams drive outcomes; he prioritizes hiring versatile people, making work enjoyable, and giving teams autonomy. He also shares management learnings: centralized control doesn’t scale in complex, fast-changing environments, and AI strategy must be flexible with rapid iteration cycles.
- •Recruit/retain/develop great people as the primary CEO job
- •Hire for adaptability over rigid roles—great people span multiple functions
- •Culture should be challenging and enjoyable; different functions may run ‘war mode’
- •Management shift: reduce hierarchy, empower decision-making at the edge
- •In AI, long-range strategy is less useful than core beliefs + continuous iteration
- 1:09:24 – 1:17:06
Remote vs in-person: offices, customer proximity, and productivity trade-offs
Matt explains Invisible’s transition from fully remote to largely in-person with multiple global offices. He argues co-location strengthens culture, accelerates thorny problem-solving, and improves customer intimacy, while still allowing flexibility outside core collaboration time.
- •Invisible operated remote for years; moved to multi-office footprint
- •In-person improves culture, collaboration, and complex problem resolution
- •Customer closeness in key cities (e.g., London/Paris) matters
- •Many newer hires prefer co-location even without strict mandates
- •Balanced model: frequent office collaboration without sacrificing flexibility
- 1:17:06 – 1:25:56
Quickfire: infra picks, investing beyond agents, margins, and an optimistic AI future
In the closing round, Matt avoids commenting on model-builder bets, calls out Databricks as foundational, and argues investors should consider AI-native service businesses—not just SaaS-like agents. He ends with optimism: AI’s net impact on energy, healthcare, and especially education could be transformative.
- •Most underrated infra pick: Databricks as a strong AI foundation layer
- •Investing angle: AI-native businesses delivering services, not just selling tools
- •Margins: challenges the myth of universally ‘80% margin’ software; focus on integrated profitability
- •Optimism on energy: AI can drive grid/cooling optimization and support clean energy investment
- •Optimism on healthcare (cost, errors) and education (access, skills, credential disruption)
