Dec 31, 2025 · 1h 25m · 20vc
Matt Fitzpatrick: Who Wins the Data Labelling Race & Why Al Needs Forward-Deployed Engineers · 20VC with Harry Stebbings
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of the 20VC podcast, host Harry Stebbings interviews Matt Fitzpatrick, CEO of Invisible Technologies, about the realities of enterprise AI deployment, the necessity of forward-deployed engineers, the limitations of synthetic data, and strategic decision-making in the rapidly evolving AI industry.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Harry holds 18.2% of the talking time here. How this is scored →
Harry as informed peer 4.4 Guest teaching 4.8 Guest disagreement 2.2 Harry pushing back 3.5
speaking balance: gold is Harry, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 32:20 Matt rejecting host board member anecdote on price insensitivity
When Harry brings up an anecdote from a board member alleging AI model builders have a complete lack of price sensitivity, Matt bluntly rejects the premise as an exaggeration and asserts standard market price bounds apply.
Hardest push from Harry ▶ 1:05:28 Harry challenging Matt's fun workplace thesis using Revolut's cultureHarry forcefully pushes back against Matt's belief that culture must be enjoyable, explicitly quoting Revolut founder Nik Storonsky's view that culture is bullshit and brutal performance pressure is what actually drives output.
Biggest teaching moment ▶ 47:03 Matt using Jevons Paradox and accounting slide rules to address talent pipeline concernsMatt completely reframes Harry's concern about AI eliminating junior career pathways by introducing Jevons Paradox and illustrating how the shift from manual slide rules to Excel increased accounting work rather than reducing headcount.
Harry holds his own ▶ 50:31 Harry probing AI data labeling revenue vs GMV using Airbnb marketplace mechanicsHarry demonstrates strong financial expertise by questioning whether AI data labeling vendors inflate top-line figures by booking gross pass-through payments to experts rather than net take rates, drawing a direct parallel to Airbnb.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Harry as informed peer | Guest teaching | Guest disagreement | Harry pushing back | Why |
|---|---|---|---|---|---|---|
| Episode Hook: The Reality of Enterprise AI Deployment | 2 | 1 | 1 | 1 | Harry sets up the interview with complimentary remarks and asks Matt to detail his career transition from a 12-year McKinsey partner leading QuantumBlack to CEO of Invisible. The dynamic is conversational and cordial with Matt detailing how he met founder Francis Peraza. | |
| Mentorship and Decision-Making Frameworks | 4 | 5 | 1 | 2 | Matt outlines industry data showing the gap between LLM benchmark jumps and enterprise adoption rates, quoting reports from MIT, Gartner, and KPMG. Harry contributes with an anecdote about a banking CTO dismissing off-the-shelf tools, leading Matt to share an example of a retailer wasting $25M on a flawed return agent. | |
| Operational AI Frameworks and Overcoming the Accenture Paradigm | 3 | 6 | 2 | 1 | Matt educates on how enterprise CFOs should manage AI procurement, critiquing what he labels the 'Accenture paradigm' of buying 50 apps and spending hundreds of millions on IT integration. Harry asks how non-technical CEOs can equip their CFOs for this transition. | |
| Contact Center Playbook and Free Proofs of Concept | 5 | 5 | 2 | 5 | Harry lists major contact center software competitors including Sierra and Decagon. When Matt mentions offering free 8-week proofs of concept, Harry pushes back directly, noting that offering free custom builds is an expensive way to run a business. | |
| The Critical Role of Forward-Deployed Engineers in AI | 4 | 5 | 1 | 4 | Harry probes the economics of forward-deployed engineers (FDEs) and references Palantir's model. Matt explains why Invisible does not charge separately for FDEs and distinguishes workflow changes from simple repository tools. | |
| SaaS is a Lie: Outcome-Based AI Pricing Models | 5 | 6 | 3 | 3 | Harry sets up a comparison as a SaaS VC between traditional SaaS subscription models and outcome-based pricing. Matt makes a provocative claim that out-of-the-box enterprise SaaS has always been a lie wrapped in hidden service costs. | |
| Human-in-the-Loop AI Training and Specialized Workflows | 6 | 6 | 4 | 6 | Harry presses Matt on exact revenue percentages coming from talent marketplaces, customer concentration, and pricing power using Hamilton Helmer's 7 Powers framework. Matt directly pushes back on an anecdote Harry shares regarding customer price insensitivity, calling it an exaggeration. | |
| Beyond the Hype: LLM Benchmarks vs. Enterprise Precision | 4 | 5 | 2 | 2 | Harry cites recent model drops like Gemini 3 and Opus 4.5 and questions whether short-lived benchmark wins matter. Matt reframes benchmark rankings as societal gauges that are largely orthogonal to enterprise task-specific precision. | |
| AI and the Future of Junior Roles: Jevons Paradox | 4 | 7 | 2 | 3 | Harry asks if automating junior work creates a long-term talent pipeline crisis. Matt reframes the thesis by bringing up Jevons Paradox and citing the historical transition in accounting from slide rules to Excel. | |
| Market Composition and Competitor Landscape | 5 | 4 | 2 | 3 | Harry asks about competitive market structure and expresses surprise when Matt names Palantir as Invisible's primary respected peer rather than RLHF players like Surge or Mercor. Matt explains Palantir's ten-year headstart on FDE culture. | |
| Revenue vs. GMV and Business Dynamics | 6 | 5 | 3 | 6 | Harry directly probes whether reported revenue figures in the AI labeling space are inflated GMV pass-throughs, drawing a direct parallel to Airbnb's take-rate accounting. Matt clarifies why variable pricing constitutes legitimate revenue recognition. | |
| The Decision to Invest in Growth vs. Profitability | 4 | 4 | 2 | 4 | Harry asks point-blank if Invisible is profitable. Matt reveals they recently raised $130M and are intentionally burning capital on expansion, prompting Harry to explore the strategic trade-offs of shifting away from bootstrap profitability. | |
| Building Invisible's Brand: Truth, Trust, and Public Narratives | 5 | 4 | 3 | 6 | Matt references Marc Andreessen's concept of divergence between public and private narratives. Harry pushes back strongly, arguing that the fundamental nature of the tech industry is selling vision before product exists ('fake it till you make it'). | |
| Fake It Till You Make It & Non-Deterministic Systems | 4 | 5 | 2 | 3 | Matt explains why non-deterministic AI systems make 'faking it' dangerous for vendors, citing a new AWS report showing 70% of claimed AI agents are merely traditional scripts. Harry references a viral video of an ineffective household robot. | |
| Lessons in Early Career & First Principles | 5 | 5 | 1 | 2 | Matt discusses his early days building AI offerings at McKinsey. Harry connects Matt's point on bank infrastructure to a recent interview he conducted with a vibe coding CEO regarding legacy software maintenance. | |
| Talent Acquisition and Company Culture | 6 | 4 | 3 | 7 | Harry challenges Matt's view on building enjoyable work cultures by quoting Revolut CEO Nik Storonsky's philosophy that 'culture is bullshit' and intense pressure drives high performance. Matt responds by differentiating research environments from pure execution cultures. | |
| Difficult Decisions: Capital, Hyperscale, and Strategy in AI | 5 | 6 | 2 | 4 | Harry brings up competitors raising massive rounds ($2B) and questions Matt on dropping remote work in favor of physical offices. Matt explains why central control in AI is a fallacy, citing military strategy to argue for edge-empowered teams. | |
| Personal Life and Constant Travel | 2 | 2 | 4 | 3 | Harry transitions to personal travel balance questions before introducing a quickfire 'discomfort round'. Matt explicitly refuses to answer a valuation choice question between OpenAI and Anthropic. | |
| Where to Invest $400M in the AI Era | 6 | 6 | 3 | 5 | Harry asks Matt to act as a co-GP on a $400M fund and questions whether software gross margins will compress permanently. Matt challenges the premise, arguing that historical 80% software margins rarely existed below the net operating line. | |
| AI Optimism: Energy, Healthcare, and Education | 3 | 6 | 1 | 1 | Harry invites an optimistic closing perspective, referencing his mother's health condition. Matt provides detailed statistics across energy optimization, U.S. healthcare waste ($14k per capita, 250k error deaths), and education reform. |
Statements from this episode (41)
Fitzpatrick: MIT report shows only 5% of Gen AI deployments work
“MIT just released this report that five percent of Gen AI deployments are working in any form.”
Fitzpatrick: External AI builds are twice as effective as internal builds
“You've seen Gartner saying 40% of enterprise projects will likely be canceled by 2027, and I think the reason for that is externally driven builds are two X as effective as internal team builds.”
Fitzpatrick: Belief that synthetic data replaces human feedback is wrong
“Look, I think the biggest one is just the view that synthetic data will take over, and you just will not need human feedback.”
Fitzpatrick: Business strategy is an overrated concept in AI
“In the AI world, at least, strategy is a somewhat overrated concept, and what I mean by that is...”
McKinsey grew engineering headcount from 100 to 7,000 during Fitzpatrick's tenure
“When I started, we had about a hundred engineers total in firm. By the time I left, we had 7000.”
Fitzpatrick: AI performance grew 40% to 60% over two years
“If you look at all the public benchmarks, models have increased 40 to 60% in performance over the last two years.”
Fitzpatrick: Enterprise AI deployment will take a decade, not two years
“I think that whole process is in the first inning in the enterprise. I think it's going to take a decade, not two years.”
Fitzpatrick: E-commerce retailer scrapped $25M custom AI agent after failures
“I was talking to an e-commerce retailer that had built an agent to handle their returns process. And they spent twenty-five million bucks building this agent. And at the end of it... What ended up happening was a couple months later they shut it down and moved…”
Fitzpatrick: CFOs will enforce strict ROI guardrails on AI within two years
“I think over the next two years you're going to see the CFO function put different guardrails on how this stuff is built and say, what is the ROI?”
Fitzpatrick: Leaders do not need technical background to evaluate AI tools
“One misconception is that that, that leader has to be highly technical to make that decision, and I would actually argue they don't at all.”
Fitzpatrick: Gen AI projects should be led by business, not tech function
“Figure out the list of three to four things that move the needle for your business. Focus on those three to four. Don't spend money on a thousand science projects. Take your best four operational leaders and put them on those four things. Don't locate it in th…”
Fitzpatrick: Out-of-the-box AI agents achieve only 33% accuracy on multi-turn workflows
“Salesforce AI research released this report that if you take if you test a lot of the out of the box agents on single-term and multi-term workflows, they're about 58% accurate on single-term and 33% accurate on multi-term workflows, which means they don't real…”
Fitzpatrick: Software industry will shift from standardized SaaS to hyper-personalized AI
“That is my view on where, where this whole industry goes, is you move from SaaS, out-of-the-box SaaS, to much more hyper-personalization using the specific data of an individual customer”
Fitzpatrick: Enterprise AI requiring workflow adaptation demands forward-deployed engineers
“If you're building something where the hardest part is getting adoption and workflow embedding, and you need to actually change the way a company works, then yes, forward deployed engineers are the only way to do it.”
Fitzpatrick: Out-of-the-box enterprise software has always been a lie
“I think you could kind of argue that out of the box software has always been a lie to some degree. Ah, it's a weird thing to say, but they always had a ton of configuration and they just dressed it up to some degree.”
Fitzpatrick: Enterprise generative AI adoption mirrors machine learning, not SaaS
“So what's happening now is we're starting to realize that the Gen AI adoption paradigm in the enterprise works the same way that ML does.”
Fitzpatrick: AI training will move into banking and healthcare next
“I actually think AI training will be used next in banking and healthcare, and then after that in, in many other different enterprise contexts.”
Fitzpatrick: Human feedback in AI training will remain essential for 10 years
“For a multi-stage reasoning test that requires a PhD in multi-different languages, and, like, human feedback is going to be important in that for the next decade.”
Fitzpatrick: Invisible Technologies partnered with US Navy, SAIC, and Vantor on AI drone swarms
“We worked with SAIC, Vantor, and the US Navy on fine tuning a model for underwater drone swarms”
KPMG report found 60% of consumers use AI weekly
“KBMG had this report that 60% of consumers use this on a weekly basis”
Fitzpatrick: Enterprise AI adoption requires 99% precision, not benchmark generalizability
“I think an enterprise uptake depends on trust and precision on specific tasks at 99% accuracy, not generalizability.”
Fitzpatrick: Recent college graduates are the highest adopters of AI tools
“I actually find a lot of the people coming out of college right now are some of the highest adopters of this and the most useful for these kind of tools, and so we're hiring more and more people of that profile, not less.”
Fitzpatrick: Palantir uniquely foresaw the necessity of forward-deployed engineering
“I think I call out Palantir because I think they realized 10 years before the rest of the kind of tech market that forward-deployed engineering and customization would be important, and I think that was a very countercultural leap at the time. You know, I, cau…”
Fitzpatrick: AI data labeling figures are actual revenue, not GMV
“I think it is revenue. I think that your, the rate you get on every project is different. The margin you make on every project is different. So I do think it is revenue”
Fitzpatrick: RLHF is the only way to accurately fine-tune context-specific AI
“And so the only way to actually do the fine-tuning process consistently And to get it accurate for any specific context is RLHF.”
Invisible Technologies raised only $7M in primary capital over nine years
“Historically Invisible had only raised seven million of primary capital in its entire nine-year journey.”
Invisible Technologies raised $130M and will forgo profitability to invest
“We've now, we initially announced a hundred, actually right now raised a hundred and thirty million, and so I'm investing very heavily in technology, so we will not be profitable this year, no.”
Invisible deploys engineers and Starlink terminals for agricultural AI computer vision
“We we're serving one of the largest agricultural conglomerates in the US on herd safety. So actually like monitoring risk factors. When should you send a vet for their herd of cows basically? And that whole process relies us on us actually sending forward depl…”
Fitzpatrick: Physical world interactions and robotics are top AI growth vectors
“And so I think physical world interaction patterns are the, some of the most interesting growth vectors for this. But they do take time and money to invest in. Robotics being another big part of that.”
Fitzpatrick: AWS report shows 70% of AI agents are traditional automation scripts
“There was actually a report AWS came out with today, it's interesting, that like, 70% of agents are actually not even, ah, AI agents, as you think of it. Like, most of the agent, agentic processes today are actually traditional script writing and just traditio…”
Fitzpatrick: Robotics will take longer and require task-specific rather than broad-based systems
“I think robotics is another one that will take longer, but will be really interesting when it works. But by the way, I think even in that case, you'll need more task specific robotics, not just broad based.”
Fitzpatrick: 70% of US software is over 20 years old
“70% of the software in America is over 20 years old.”
Fitzpatrick: Average bank spends 93% of tech budget on maintenance
“The average bank spends 93% of its cost, of its tech cost, on maintain initiatives. Seven percent go into building new things.”
Fitzpatrick: Top hires should not be recruited for specific roles
“When you recruit a great person, I don't think about role most of the time. Meaning I think people are very role focused of like, I will hire this person and they will only do oil and gas as an example, right? But the reality is like really good people will ru…”
Fitzpatrick: Five-year strategic planning is useless in AI today
“Five-year strategic planning is not a useful exercise right now.”
Fitzpatrick: Databricks is the most useful foundation for AI
“Look, I think their tech is great. And I think that it's interesting in a lot of ways, the most useful foundation for AI is really good Databricks infrastructure.”
Fitzpatrick: AI-native physical-world service companies will be highly disruptive
“Some of the most interesting new businesses are actual businesses using AI in the physical world that are AI native and that will be highly disruptive.”
Fitzpatrick: Recent Y Combinator class generated double the revenue of prior batches
“If you look at Y Combinator's recent class, I think it's like the largest, it's two X the revenue of any prior class.”
Fitzpatrick: Public software multiples fell from 20x to 10x over two years
“In the last two years, you've seen public software multiples go from 20 X to 10 X, partly because of growth changes, and partly because they've tried, as they move profitable, their growth slows materially”