Where the Winners Get Built
Six Categories, Six Sets of Gates, and the Founders We Back
Part 3 of the AI Defensibility Series
This is Part 3 of my series on AI defensibility. Part 1 — “Judgment May Be All That You Need” — made the case for the judgment layer as the last technical moat. Part 2 — “The Compounding Loop” — explained the mechanism underneath: traces, harnesses, and the compound dynamics that make AI-native products get better with every deployment. This piece puts that framework to work: the six categories, most of them where we believe the loop spins hardest, the specific investment gates we apply to each one, and the kind of founders we back.
In the last piece, I laid out how the compounding loop works — traces feeding harnesses feeding judgment layers, all spinning together in a way SaaS never had. But I also made a promise: that I’d map this onto the specific categories where Bonfire is placing its bets.
This is that piece. And I want to be blunt about what it is and isn’t. It’s not a market map. It’s not a “state of AI” survey. It’s the thesis document behind how we are deploying our $245M Fund IV — the categories where we believe AI-native companies build structural advantages that widen over time, and the specific gates that every opportunity must clear before we’ll invest. We publish this for prospective founders and our upstream pre-seed collaborators to understand whether we are a great fit or not for a particular investment opportunity. Some of these categories are consensus. Some are deeply non-consensus. The question is the same: does the compounding loop spin here, and can this founder make it spin faster than anyone else?
I should also be clear about what we don’t do. We don’t invest in deep tech, cyber, or silicon — foundation models, chip design, robotics hardware, novel architectures. There are amazing opportunities there, but it’s not our background or our bailiwick – there are many other investment firms out there who are better prepared to assist founders on those journeys. These six categories are the areas where we believe our team can best identify exceptional founders, where we see enduring value creation over the next decade, and importantly where we believe we might be the best fit investor for those founders.. We invest at the application and infrastructure layers where domain expertise, go-to-market instinct, and the compounding loop matter more than research breakthroughs.
The Universal Filter
Before the categories, the filter. Every prospective investment we see — regardless of category — has to clear five gates that sit above any category-specific criteria.
First: is this a multiplayer problem? AI democratizes building software but concentrates winning with it. Single-player tools — one user, one task, limited complexity — are the most exposed to being absorbed by the next frontier model update. The companies that endure increasingly solve multiplayer problems: multi-stakeholder workflows with orchestration, governance, and domain logic that no general-purpose model will replicate.
Second: does the product value compound with every customer? This is the compounding loop test from Part 2. If the 500th customer doesn’t make the product measurably better for customer #1, it’s a feature, not a business. We need to see the traces accumulating, the harness learning, and the judgment layer deepening.
Third: when a better foundation model drops, does it help or hurt? If the answer is “hurt,” the company’s value lives in the code — and code is what AI commoditizes fastest. If the answer is “help,” the company owns the domain-specific layers that sit on top of whatever model is best, and every model improvement makes their harness more powerful. We only invest in companies where better models are a tailwind.
Fourth: what does this look like when costs deflate 50%? The cost of intelligence is dropping at a rate (well not recently because that is where people are just token-maxxing as they have yet to establish trusted context layers to tap into) that should terrify any company whose value proposition is primarily “cheaper than the human alternative.” We call this the deflation question: if the customer’s willingness-to-pay for this drops by half over five years, does the company still win? It does if it owns proprietary data and / or sits in the token path and / or has built deep embeddedness. It doesn’t if the margin belongs to the cost curve rather than the company.
Fifth: is the customer buying out of existential need — or nice-to-have improvement? The companies in our portfolio growing fastest are selling into what I call the existential buying moment. Every professional knows AI is coming for undifferentiated work. The ones who master AI-native tools become the people their companies can’t afford to lose. That’s an existential bond, not a convenience preference. We will know the difference after just a few customer calls.
Now, the six categories — and the specific gates inside each one.
1. Vertical AI Software
This has been our highest-conviction category — the one where we concentrate a lot of capital — and it’s where the compounding loop argument becomes most concrete.
Pre-AI, building vertical software for a $500M niche required the same 100+ engineer investment as building for a $10B horizontal market. So most verticals went unserved, or they got generic horizontal tools awkwardly bolted onto industry-specific workflows. AI collapses that cost structure entirely — a 3-person team can now build a deep vertical product that would have taken 30 engineers two years ago. That cost collapse has unlocked hundreds of verticals that were economically unviable in the SaaS era — and the founders moving fastest into them right now will be the incumbents everyone else has to unseat.
But the cost collapse isn’t what makes vertical AI defensible — and this is the distinction most investors miss. What makes it defensible is depth. We are specifically looking for verticals with genuinely hard problems — deep regulatory complexity, multi-stakeholder workflows with conflicting incentives, edge cases that multiply rather than converge, domain logic that takes years to encode. A vertical AI company serving dental practices needs to verify insurance eligibility before scheduling, cross-reference CDT codes against each payer’s coverage matrix, manage pre-authorizations that vary by procedure and carrier, and handle the cascade when a patient’s plan changes mid-treatment. A commercial roofing platform needs to understand that a change order triggers a lien waiver, that the waiver requirements differ by state and municipality, and that the insurance implications change based on the sub’s coverage type. A mid-market insurance carrier’s subrogation process has seventeen steps that vary by state, and the judgment layer needs to learn which steps a specific adjuster skips and why.
There’s another dimension to the vertical opportunity that generative AI unlocks, and it’s particularly powerful down-market. Many of the deepest verticals serve customers who have no IT department and are not technology-forward. These are business owners who were never going to hunt and peck through a powerful SaaS application, no matter how many AI-assisted features it had. What AI-native vertical companies can do now is offer a full-stack solution where the complexity is masked entirely by agentic front-ends. The agents use the software on the user’s behalf — the business owner describes what they need, and the system handles it. Especially down-market, these products aren’t headless infrastructure. They become the complete solution the customer uses, eventually evolving into what is essentially an agentic model that runs the business operations for the owner. That’s not a software tool. That’s a fundamentally new relationship between a business and the technology that operates it.
That depth is the test. If the vertical’s workflows are shallow enough that a horizontal AI product could serve them by swapping out a few industry-specific nouns, there’s no compounding intelligence to build. The domain ontology would be thin, the traces would be generic, and a better foundation model could replicate the product in a quarter. Those aren’t vertical AI companies — they’re horizontal products wearing a vertical costume. We pass on them, regardless of how attractive the TAM looks on paper, because there’s no long-term defensible intelligence that compounds.
The verticals we get excited about are the ones where the problem is so structurally deep that the domain ontology itself becomes an asset that takes years to build and can’t be replicated by a team that hasn’t operated in the industry. Every edge case the system learns, every regulatory exception it encodes, every workflow quirk it absorbs — those are traces that feed the compounding loop. And the loop spins harder in deep verticals because the edge cases never stop coming. The 500th customer surfaces a regulatory nuance the first 499 didn’t encounter, and the system gets better for all 500.
The compounding loop goes nuclear in verticals because of two dynamics. The first is what I call the TAM staircase. The best vertical AI companies don’t stay at the point-solution level. They climb: from point solution ($50-200/month per seat) to all-in-one plus payments ($500-2K/month plus processing) to labor replacement ($2-5K/month, outcome-based) to business-in-a-box (percentage of revenue). Each step generates richer traces, builds deeper judgment, and makes the system harder to replace. At the top — business-in-a-box — the AI doesn’t just help run the business. It runs the business. The founder coaches the system; the system operates the company. That’s a relationship no one switches away from.
The second dynamic — and the one we are increasingly convinced separates the great vertical AI companies from the good ones — is cross-node intelligence across supply and demand chains. The deepest verticals don’t serve a single customer type. They serve an ecosystem — the supplier and the buyer, the contractor and the client, the provider and the payer, the manufacturer and the distributor. When a vertical AI platform gets agents operating on both sides of a transaction — and eventually across an entire supply chain — the intelligence it accumulates isn’t about one company. It’s about how the chain itself works. How this supplier interacts with this buyer on this type of order in this region. How pricing moves when lead times change. How quality issues propagate downstream. That cross-node graph is structurally exclusive. No single participant could build it. No horizontal player goes deep enough to earn it. And the switching cost is extraordinary — leaving doesn’t just mean losing your own judgment layer, it means losing access to the collective intelligence about how your counterparties operate and the agent-to-agent coordination that was handling your cross-company workflows.
The verticals where both dynamics are present — the TAM staircase plus cross-node intelligence — are the highest-conviction investments in our portfolio.
The vertical AI gates we apply at IC:
The core question: is the problem deep enough that the compounding intelligence built by serving this vertical is structurally defensible — or is this a horizontal AI product with a vertical label? If a well-funded horizontal player could encode the domain ontology in six months, it’s a horizontal product. If the domain requires years of accumulated edge cases, regulatory knowledge, and workflow intelligence that only deepens with every deployment — that’s a vertical we want to own.
Beyond that, we walk through five gates. Does the founder have the domain depth to build the ontology, and the builder instinct and more importantly taste to ship it as product? Domain without builder DNA means a consulting company. Builder without domain means a thin wrapper. Is the TAM staircase visible from seed — can we see the path from point solution to business-in-a-box, even if the founder is landing on step one today? Is there a cross-node opportunity across the supply or demand chain — who else in the chain would benefit from being on this platform, and what intelligence would you accumulate by connecting them? Can we see winner-take-all dynamics within the fund’s life — because vertical categories consolidate fast in AI, and if we can’t see the gap opening within 3–4 years, the timing is wrong. And finally: in mid-2026, is the product they demo to us already agentic — and is the founder furiously working to get their harness and reinforcement loops running? If the product is still a traditional SaaS interface with AI bolted on, the window is closing. The founders who win in verticals are the ones whose products already feel like agents, not applications, and who are building the compounding infrastructure to make those agents get better with every deployment.
A great example of our investment thesis here is Jerry Zhou and the team at Supio - the agentic solution for personal injury law and mass tort. Supio’s rapid growth has been fueled by the fact their customers, many of whom never used software before for this work, can both handle far higher case loads and secure far higher settlements for their clients. In recently watching their customers in person use their fully agentic offering that integrates to all the law firms practice information data, their case information, and Thompson Reuter’s Westlaw research, you hear and feel the power of a smart agent with all of the business context - i.e. “I have been running this firm for 10+ years. Today, with Supio agents, it feels like this is the first day of a brand new firm and how we operate.”
2. AI + Proprietary Data
This category is about companies that sit on data assets that are genuinely proprietary — not because of legal protection or exclusive contracts, but because the data is generated by the company’s operations and can’t be replicated by training a model on the public internet. Sensor data from industrial deployments. Claims history from insurance operations. Transaction patterns from financial infrastructure. Operational intelligence earned from thousands of real-world agent deployments.
The compounding loop works here because the proprietary data feeds the harness, the harness makes the agents better, which generates more proprietary data. The intelligence network effects are particularly powerful when multiple nodes in a vertical supply chain — supplier, operator, buyer, regulator — all run agents through a shared platform. The platform accumulates intelligence across the chain, not just within one customer. That cross-node graph is structurally exclusive: no single participant could build it, no horizontal player goes deep enough to earn it, and switching means losing access to collective intelligence, not just your own data.
I described the cross-node intelligence dynamic in the Compounding Loop piece — the GC’s agent negotiating with the sub’s agent, the supplier’s agent coordinating with the operator’s agent — and this is the category where it lives. The companies building these cross-node data platforms aren’t just collecting intelligence within a customer. They’re building the intelligence graph of an entire vertical supply chain. The switching cost isn’t your data walking out the door. It’s everyone else’s data walking out the door.
The proprietary data gates we apply at IC:
The core question: is the data structurally exclusive — earned through deployments, not easily aggregated by AI? Could an AI replicate this dataset in weeks by crawling the internet? If yes, it’s not proprietary. We apply this test ruthlessly. Structured public data that’s been reformatted is not a moat. Data that can only be generated by operating in the domain over time — that’s structurally exclusive.
Beyond that: does the data get more valuable with every deployment, or is it a static asset? A data moat that doesn’t compound is an appreciating asset, not a flywheel. We need the harness on top — the mechanism that converts proprietary data into better agent performance, which drives more usage, which generates more data. Is the data in the token path? Proprietary data that sits in a warehouse and never touches the AI at inference time is not a compounding asset. And can we see cross-node network effects? The strongest versions of this category are platforms where multiple participants in a supply chain or ecosystem all generate data through the same system — the deepest moat we’ve found.
Rwazi is a compelling example of AI + Proprietary Data in action. The company, led by Joseph Rutakakngwa, has built a network of over 3 million consumers across 160+ countries — primarily in emerging markets like Africa, South Asia, and Latin America — who voluntarily collect and share billions of zero-party data points about what they buy, where, when, how much they pay, and why (including information about retail locations using computer vision). This is ground-truth consumer and retail intelligence that simply cannot be scraped from the internet or synthesized by an AI model. The compounding loop is textbook — every new consumer log and gig completion enriches the dataset, which feeds Rwazi’s AI engine (including their autonomous agent, Sena), which delivers better insights to Fortune 500 customers like Unilever, P&G, Red Bull, and Mastercard, which drives expansion and usage, which generates even more data. A cross-node dynamic is emerging as well: as brands, retailers, and distributors across the CPG supply chain all query and act on the same underlying consumer intelligence graph, Rwazi becomes the shared data fabric for an entire vertical ecosystem in markets where no alternative exists. With 148% NRR, zero logo churn, 80% gross margins, and 25x ARR growth since Bonfire’s initial investment, Rwazi demonstrates what happens when structurally exclusive data meets a true compounding harness.
3. Services-to-Software
For every dollar spent on software, six dollars are spent on services. AI-enabled services is now an entire thesis category — firms like Emergence are all in, and for good reason. AI unlocks that $5T+ market at software margins.
Services-to-software companies look like services on the outside but run AI economics on the inside: 75%+ gross margins on what used to require armies of billable humans. They compress the services pyramid — outcome-based pricing replaces hourly billing, AI handles the volume work, humans handle the exceptions. The result is software-like margins on services-sized TAMs.
The critical distinction we make: dispatchers versus builders. A dispatcher has borrowed margin — they’re arbitraging the gap between what AI costs and what humans cost. When inference costs drop further (and they will), the dispatcher’s margin compresses. A builder is accumulating durable assets with every engagement: proprietary data, process knowledge, customer relationships, and most importantly, traces from full task lifecycles including failures and corrections. Those traces are uniquely rich — services-to-software companies that own the outcome generate the most complete behavioral intelligence in the market, because they see the entire task lifecycle, not just one step.
Seed investors trained on pure SaaS pattern-match past these deals. The gross margins look wrong in Year 1. The business model doesn’t fit the traditional SaaS template. But the companies that get this right are addressing a TAM that’s 10x larger than anything SaaS could touch — and the compounding loop gives them a moat that deepens with every engagement.
The services-to-software gates we apply at IC:
The core question: is this company accumulating durable assets with every engagement, or just dispatching cheaper labor? This is gate one, and it kills most deals. We draw the line hard: if the business model is primarily cost arbitrage, the margin belongs to the inference cost curve, not the company. We’re looking for builders — companies that get structurally better with every engagement because the traces from full lifecycle ownership feed back into the harness.
Beyond that: what does Year 1 versus Year 3 look like? In Year 1, gross margins might be sub 50% as the AI handles the easy cases and humans handle everything else. By Year 3, the compounding loop should have pushed margins above 75% as the AI handles progressively more complex work and the human intervention rate drops. If the margin trajectory isn’t visible in the unit economics model, the loop isn’t spinning. Can we underwrite this without SaaS pattern-matching — because we’ve built internal models specifically for this category, and we’re comfortable with the Year 1 optics if we can see the Year 3 economics. And is the TAM genuinely services spend that software never could touch? The best deals in this category are replacing work that was previously unbiddable by software — not because the software companies didn’t want the market, but because the work required too much judgment, too many edge cases, too much contextual understanding. AI makes that work biddable for the first time. If the TAM is just existing software spend at a lower price, that’s a pricing play, not a category creation.
Juno Tax is a textbook illustration of the Services-to-Software thesis. Tax preparation has historically been a pure labor business — practitioners spending hours manually entering thousands of data points, with economics dictated by billable hours. Juno compresses that entirely: its platform automates roughly 90% of data entry across 90+ document types, cutting prep time from 2–3 hours to about 7–10 minutes, with disruptive per-return pricing replacing the hourly model. Founded by David Haase, a CPA who ran his own firm and used early Juno prototypes inside it, the company is a builder, not a dispatcher — accumulating proprietary process knowledge, deep integration with legacy tax software, and traces from complete return lifecycles including edge cases, corrections, and human reviewer overrides that generic AI tools will never see. Juno keeps a human in the loop, meaning it owns the full behavioral signal of how expert CPAs handle exceptions. The result is software economics on a massive services TAM, with a compounding moat that deepens with every return processed.
4. Fintech Hard Rails and Deterministic Infrastructure
Not everything in AI is probabilistic. Some outputs must be exactly right — and the companies that provide deterministic guarantees in a probabilistic world occupy a structural position that only gets more valuable as AI adoption accelerates.
Payment processing, compliance verification, identity resolution, audit trails, regulatory reporting — these are the hard rails every AI system needs to interact with but can’t afford to get wrong. A legal AI agent can draft a contract with probabilistic reasoning, but the payment terms need to be deterministic. An AI-native accounting system can categorize expenses intelligently, but the tax calculations need to be exact.
The compounding loop works differently here than in the other categories. It’s less about the product getting smarter through traces (although that happens) and more about volume: more AI adoption across the economy means more transactions, more compliance checks, more identity verifications, more payments flowing through deterministic pipes. The TAM expands automatically as AI adoption grows. Every new AI-native company that launches needs hard rails to operate on. The companies that provide those rails don’t compete with AI — they benefit from every new AI deployment.
The fintech hard rails gates we apply at IC:
The core question: does AI adoption expand this company’s TAM automatically — or does growth depend entirely on the company’s own sales effort? The best deals in this category have a TAM that grows as a function of AI deployment across the economy. More agents operating means more transactions through your rails, more compliance checks, more identity verifications. We want to see the structural linkage between overall AI adoption and this company’s revenue — not just a correlation, but a causal mechanism. If the company’s growth is decoupled from the broader AI adoption curve, the category tailwind isn’t real.
Beyond that, we walk through three gates. Is the regulatory or compliance moat real and growing — because if the barrier is purely technical, it’ll get competed away, but if the company is earning certifications, navigating compliance frameworks, and building trust infrastructure that takes years to establish, the moat compounds; regulatory moats are the only moats that actually get wider when incumbents try to cross them, because the regulatory load increases with scale. Is the company in the token path for high-frequency, high-value transactions — a hard rail that processes $10,000 payments for enterprise AI workflows is structurally different from one handling $0.01 micro-transactions, and we’re looking for companies sitting in the payment or compliance flow for transactions large enough to sustain meaningful economics per event. And can the company build a judgment layer on top of the deterministic rails — because the best outcomes here are companies that start with deterministic infrastructure and then build an AI-native intelligence layer on top, using transaction data to detect fraud, optimize compliance workflows, or predict regulatory changes; that combination of deterministic foundation plus probabilistic intelligence is extremely hard to replicate.
Straddle is our conviction bet in Fintech Hard Rails. ACH moves $86 trillion a year across the U.S. economy but runs on infrastructure built in the 1970s — slow, opaque, and vulnerable to fraud. Straddle is rebuilding it: a programmable payments platform that makes bank-to-bank transfers move with debit-card speed and reliability across ACH, RTP, and FedNow through a single API. The wedge is the identity layer — KYB, KYC, AML, and fraud prevention built in from the ground up, not bolted on, collapsing what used to require five or six separate vendors into one cohesive platform and settling payments in under 24 hours. CEO Keith Raphael spent years deep in the payments and compliance ecosystem — he and co-founder Chad Willard previously built and scaled a legacy processor from $25K to $350K MRR before founding Straddle to rebuild the stack with modern primitives. As AI agents increasingly initiate autonomous transactions, every one needs deterministic rails underneath — Straddle’s TAM expands automatically with every new AI deployment that moves money.
5. Agent-Native Infrastructure
I’ll be honest about this category: it’s the most consensus-heavy in our portfolio, and I have mixed feelings about that.
This is the picks-and-shovels layer — identity, payments, orchestration, trust, observability for AI agents. The thesis is straightforward: as the world moves to agentic systems, you need infrastructure specifically designed for agents. The compounding loop exists here but operates at the infrastructure level: more agents in the ecosystem means more interactions through your pipes, more edge cases in your orchestration layer, more signal in your trust framework.
My concern — and I’ll flag it openly because intellectual honesty is the whole point of this series — is that this is the category where every mega-fund is already playing. The valuations reflect it. At seed, we need conviction that our specific company will be crowned before the market realizes it’s obvious. We’re highly selective here.
The agent-native infrastructure gates we apply at IC:
The core question: is this genuinely a new primitive, or is it an existing infrastructure pattern rebranded for agents? This gate kills most deals. Payment rails for agents aren’t fundamentally different from payment rails for humans unless the company is solving a structurally new problem — like agent-to-agent settlement, autonomous authorization, or machine-speed compliance verification. If the product is essentially Stripe with an “AI” badge, it’s not agent-native infrastructure. The bar is high: we need to see a technical problem that didn’t exist before agents, not an existing problem with “agent” appended to the marketing copy.
Beyond that, we apply two gates that reflect the honest tension in this category. Can a seed-stage company win this against well-funded incumbents — because if Stripe, Datadog, or Okta can bolt on an agent layer in six months, the seed-stage company is toast; we invest here only when we believe the company has found a wedge the incumbents can’t reach quickly, either because the technical problem is genuinely novel, the go-to-market motion requires being AI-native from day one, or the network effects compound in a way that rewards being first rather than being biggest. And do we see winner-take-all dynamics in this specific sub-layer — because infrastructure tends toward natural monopoly when there are strong network effects, and agent infrastructure has them, but the category also tends to fragment; we need to believe our company is competing in a sub-layer that will consolidate, not one that will splinter into ten equivalent providers.
A great example here is Enoch Zhu and the team at Orbifold AI — building the training data infrastructure that physical AI needs to actually work. The premise: the LLM era trained on free internet data, but physical AI has no such luxury. Every hour of usable robot training data must be physically captured and pose-recovered to sub-15mm accuracy — and recent research shows that world action models collapse below 30% success when trained on weakly-aligned video. Orbifold produces sub-5mm pose-accurate human-centric data at 5,000 hours per week — 3x tighter than the industry standard — with an eval-to-curation flywheel where every delivery makes the next one better. The validation: DeepMind, NVIDIA, and World Labs — the very labs whose published research defines this data type — all chose to buy from Orbifold rather than build it themselves. Physical AI won’t deliver on its promise without this layer.
6. Structural Reshaping
This is our non-consensus allocation.
The macro premise: AI is going to structurally reshape how entire sectors of the economy operate. Not in a decade — in the next three to five years. Industries that only exist because of friction will compress. Roles that exist because humans were the only option will disappear. New industries that nobody has imagined yet will emerge around the capabilities AI creates.
This is where the “AI replaces the seat” thesis lives in its most aggressive form. Not AI that helps a person do their job better — AI that eliminates the need for the job entirely and creates something new in its place.
But structural reshaping isn’t just about AI making existing jobs more efficient — it’s about entire market structures changing shape. The middlemen that existed because of information asymmetry get disintermediated. The bundled services that existed because unbundling was too expensive get disaggregated. The industries that were structurally small because the friction kept participants out suddenly become large because AI removes the barrier to entry. The reshaping creates new categories, new workflows, new market participants, and new value chains that bear little resemblance to what they replace.
The structural reshaping gates we apply at IC:
The core question: is the macro shift a tailwind that will make this category inevitable within the fund’s life — and can we see structural evidence that it’s already happening? Timing risk is the dominant risk in this category. We’re comfortable being early — but not so early that the market doesn’t arrive before the fund needs returns. A plausible narrative about how AI might reshape an industry isn’t enough. We need leading indicators: incumbents losing pricing power, talent flowing toward the new model, early customers choosing the restructured version over the legacy option. If the shift requires three more platform cycles before it’s real, the timing is wrong for a seed fund.
Beyond that, we apply three gates. Does the founder have the domain depth to see around corners — because structural reshaping requires founders who understand the industry well enough to know which frictions are load-bearing and which are artifacts; a founder who’s spent 15 years in an industry and can articulate exactly which $50B services category is about to collapse, and why the replacement looks nothing like the incumbent, that’s the founder for this category; outsiders with a thesis and no operating experience need not apply. Can the company survive the transition period — because these businesses often look weird in Year 1, with economics that don’t match any existing category, unconventional customer bases, and products that look like services from the outside; we need to see a credible bridge from “strange early-stage company” to “dominant platform in a category that didn’t exist three years ago,” and this is where the seed investor’s advantage lives, in underwriting ambiguity that later stage investors might not touch. And is the restructured market large enough to justify venture returns — because non-consensus bets need outsized outcomes to compensate for the timing risk; we’re not interested in structural reshaping of a $100M market, we’re looking for categories where the reshaping creates a multi-billion-dollar opportunity because the friction being removed was the thing that kept the market small.
A great example of a non-consensus bet in this category is Jim Milton and the team at CMS.ai — building the infrastructure layer that allows B2B brands to be discovered, dialogued with, and trusted by AI agents. Most CMOs still think “AI strategy” means optimizing their website to show up in ChatGPT search results. Jim’s insight is that’s a dead end — 94% of B2B buyers already use AI to research purchases and 75% of final vendor selections come from AI’s initial shortlist, yet only 4.3% of B2B companies even appear in AI answers. The other 95.7% are invisible. CMS.ai is building the owned infrastructure to fix that — Brand Registry, MCP-native content servers, and trusted answer layers — with $500K ARR at pre-seed. Almost no one in the market sees this yet. The entire B2B marketing world is retrofitting SEO tactics for AI while Jim is building the agent web’s brand layer from scratch.
The Founders We Back
I’ve laid out where we invest. Let me close with who we invest in — because in a world where code costs approach zero, the team IS the differentiator.
We have a hard requirement that we have never had before in 20+ years of investing: the founders must be builders themselves. Not just of their product, but of how they operate the company. If the CEO isn’t personally building with AI — using Claude Code, Cursor, whatever the tools are — to ship product and run operations, we cannot invest. Hard stop.
This is not a soft preference. It’s a gate. And I’ll explain why.
The founders who are personally building with AI tools develop a fundamentally different intuition about what’s possible. They feel the capabilities expanding in real time. They discover product ideas in the act of building that no planning session would surface. They move at a velocity that non-builders cannot match — compressing what used to take quarters into weeks, because they’re not waiting for their engineering team to context-switch onto their request. They’re shipping it themselves.
More importantly, they make better architectural decisions. A CEO who has personally experienced the difference between a well-structured harness and a bolted-on AI feature makes different choices about what to build, how to build it, and where the compounding loop will generate the most value. They understand trace capture intuitively because they see how the AI improves when they feed it context. They understand the judgment layer because they’ve experienced it sharpening in their own workflow.
Founders who delegate AI to their engineering team while they “focus on strategy” are building a company for the last era. They’re making the same mistake the CEOs made in 2008 who thought mobile was a feature their dev team could bolt on while they focused on the desktop product. Mobile wasn’t a feature. It was a new way of building. AI is the same — and the founders who treat it as something the engineering team handles are going to lose to the founders who treat it as the way everything gets done.
There’s another reason the builder requirement is non-negotiable, and it’s one that doesn’t get talked about enough: hiring. A founder who isn’t actively building and running their company on AI will struggle to identify who the key people are to add to their team — and more importantly, what those people actually do. The roles in an AI-native company don’t map to the old SaaS operating model. You don’t need a VP of Engineering who manages 40 developers. You need someone who can architect harnesses and design the compounding loop. You don’t need a traditional CS team. You need people who can train the judgment layer and curate the traces that make the product smarter. The org chart of an AI-native company looks nothing like the org chart of a SaaS company — and a CEO who isn’t living inside the tools can’t see the new roles, can’t evaluate candidates for them, and ends up hiring the team they would have hired in 2019. That’s a company that’s structurally disadvantaged before it ships a line of code.
I’ll go further. We look at how the founder runs the company itself as a signal. Are they using AI agents to handle their own operations — customer success, content, internal processes? Or are they running a traditional org chart with traditional headcount assumptions? The AI-native founder runs a lean team where every person is amplified by AI, and the compound output per employee looks nothing like what traditional headcount models would predict. That operational philosophy — not just using AI in the product, but using AI in how you build and run the company — is the clearest signal we’ve found for whether a founder truly understands what era they’re building in.
I’ve started watching for a specific tell in founder meetings: when I ask “show me what you shipped last week,” the builder-CEO pulls up their commit history, their Claude Code session, their Cursor workspace. They show me the thing they built. The non-builder CEO pulls up a roadmap slide and talks about what the team is “working on.” That’s the gap. And in a world where the best founders are personally shipping products at 10x the speed of a traditional CEO, the non-builder isn’t just at a disadvantage. They’re in a different league.
Beyond the builder requirement, we look for five things in the founders we back. None of them are soft preferences. Every one of them is load-bearing.
The first is an almost pathological drive to win. I don’t mean ambition in the polished, LinkedIn-bio sense. I mean the kind of relentless, consuming intensity where losing isn’t an acceptable outcome — where the founder wakes up at 3am thinking about the deal they didn’t close and is back on it by 6. The AI era is compressing competitive cycles so severely that the founders who treat this as a lifestyle business, who optimize for balance before they’ve earned it, will be lapped by the ones who don’t. This has always been true in venture-backed startups, but the velocity of AI makes the gap between intense and moderate founders wider and more visible than it’s ever been. The founders we back aren’t seeking work-life balance — they’re seeking to win, and they’ll sort out the rest once they’ve built something that matters.
The second is what I’d call mental plasticity — the ability to operate and thrive despite a pace of change that is genuinely unprecedented. The playbook that works in January might be obsolete by June. The product architecture you committed to in Q1 might need to be reinvented in Q3 because a new model capability made a different approach possible. The go-to-market that was working might stop working because the landscape shifted underneath you. The founders who succeed in this environment aren’t the ones who avoid setbacks — they’re the ones who are undaunted by them. If you can’t handle the idea of re-inventing major parts of the business every year — or more often — this isn’t the era for you. We look for founders who don’t just tolerate that chaos but draw energy from it, who see each disruption as a chance to pull further ahead while their competitors are still processing what happened.
The third is magic with customers. The best founders we back are their company’s best seller — not because they can’t hire salespeople, but because they have an instinct for what customers need that no hire can replicate. They close the first 20 deals themselves. They hear what the customer isn’t saying. They turn a pilot into a platform relationship because they’re in the room, reading the signals, and making commitments that land. In AI, where the product is evolving weekly and the use cases are still being discovered, the founder who is personally in front of customers is learning things that no dashboard or sales report can capture. That customer intimacy feeds directly back into the product — and into the compounding loop. We can’t teach this. Either the founder has it or they don’t.
The fourth is going for big wins. We back founders who are building to dominate a category, not to build a comfortable business. Big vision paired with a sharp wedge — they know exactly where to start, but they also see where this goes when it works. The ruthless taste for saying no to good opportunities in order to preserve focus on the great one. The willingness to make architectural bets that only pay off if the company becomes the category leader, because they’re not building a backup plan. In a world where AI compresses competitive timelines and winner-take-most dynamics are more extreme than ever, the founders playing for second place are playing to lose
The fifth is domain depth. The founders we back bring context that can’t be automated — earned through years of operating in the industry they’re serving. They know which workflows are load-bearing and which are artifacts. They know which stakeholders have real power and which have performative power. They know the regulatory landmines, the cultural resistance points, the incumbent relationships that look unbreakable but aren’t. That domain expertise is the raw material the compounding loop needs to spin — without it, the AI is building on shallow foundations, and the first competitor with real domain knowledge will outrun you.
The gap between an outlier team and an average team doesn’t narrow in AI — it widens spectacularly. A founder with the compounding loop spinning and the domain depth to keep feeding it will pull away from the competition faster than anything we’ve seen in SaaS. The winners in each of these six categories won’t just be good companies. They’ll be companies that the market can’t catch once the flywheel is turning.
One final thought on the founder bar. I spent six years at Siebel, twelve years at Salesforce, and another run as COO/CRO at SmartRecruiters. In every one of those eras, the best founders were the ones who understood the technology deeply enough to make architectural bets that their competitors wouldn’t make for years. Tom Siebel understood client-server before the market did. Marc understood multi-tenant cloud when everyone else was still shipping on-premise. The AI era is no different — except that the velocity of change means the gap between “gets it” and “delegates it” opens in months, not years. The builder-CEO isn’t a nice-to-have. It’s the single strongest predictor of whether a company will earn the compounding loop or just talk about it.
The speed at which all of this is happening still catches me off guard. Categories that had no name 18 months ago now have multiple funded companies competing for the #1 position. The compounding loop isn’t theoretical — it’s separating winners from losers in real time, in every category we invest in. And the founders who see it, who build for it, who are personally immersed in the tools that make it possible — they’re the ones pulling away.
The context era taught us what information matters. The judgment era taught us who wins. The compounding loop told us how they will win. And now — with these six categories and the specific gates inside each one — we believe we know where our winners will get built, and what it takes to back the right ones.