Jason Shuman

The through-line

The idea that recurs across everything Jason Shuman says is that the best businesses take work, complexity and judgment away from human beings entirely, and that the value accrues to whoever absorbs the most of it. He frames the standard for the current generation of software by pointing at payments: Stripe's real genius was abstracting compliance, fraud and complexity rather than merely processing payments, and that is the bar for vertical AI [1]. Vertical software of a few years ago "required so much interaction by the human being to input things into the system to get outputs," whereas the products he now backs are "doing a lot of the heavy lifting and the work for people and extracting all of the complexity" [1]. The same instinct governs how he thinks about his own firm: Primary is built on the belief that an investor should absorb work from founders rather than make introductions, a distinction he puts as being "Builders not backers," because "backers basically use their Network as currency" while builders "also do the work" [3].

That preoccupation has sharpened rather than changed. Early in his investing career he was making bets he then called "the future of automation," which he now says were really AI companies; of nine such investments, two became or are becoming unicorns and the rest died, "probably because AI just was not ready in for prime time, and I was too early" [1]. The thesis survived; the timing finally caught up.

On what makes vertical AI different from vertical software

The test he applies is whether the product delivers an outcome rather than a workspace. His example is Bobyard, which does AI takeoffs for the trades: previously a landscaping company receiving an architect's drawing set had someone whose full-time job was clicking on the drawing and counting every bush and every tree across perhaps a thousand items, with one-, two- or three-week lead times, or the firm simply declined to bid [1][9]. Since takeoffs and quotes are "the number one bottleneck to revenue" in that industry, compressing them to minutes is not a productivity gain but a revenue unlock [1]. His generalisation: "if you can take the AI products and back into what is a very obviously clear ROI and pain point or bottleneck for the business, I think you have a really good opportunity ahead" [1].

He is equally attentive to the effect of turnaround time on market size. Light Table does AI clash detection and AI peer review on construction drawings, work that historically meant a human hunting for where a structural beam hits an HVAC unit, and which many developers skipped altogether [1]. Where a human review might take twenty weeks and be 50% accurate, the software returns it the same day, which "is unlocking a completely different market" while reducing change orders by 70% and lifting developer returns by roughly two points [1].

On the two entry points, and why he prefers the wedge

He sees two clean shapes for a vertical AI company. The first is a high-velocity wedge with a single-player use case that can be understood in a thirty-second clip, where the customer's reaction is "Oh my god, I want that," and conversion on sales pitches runs north of 50%, which he notes is extraordinarily high [1]. The second is an AI-native system of record, where defensibility is obvious from step one but go-to-market is "a rip and replace motion for a very very heavy, bulky thing," growth is slower, and slower growth means less capital attracted [1].

He has resolved this internally in favour of the wedge, on the bet that wedges can move into the system of record as AI coding shortens migration times and accelerates product roadmaps: "I'd rather be betting on companies that can have the higher velocity wedges right now just because it seems easier to go into the system of record than vice versa" [1]. The risk he names honestly is that a single-player wedge "might not be as defensible when it is in the wedge" [1].

On moats, and why software-only moats are under pressure

He is candid that nobody has the answer yet: "none of us I really think know what are the modes going to be longer term right now. It is a very very ambiguous and uncertain time" [1]. What he does believe survives is switching cost built from multiplayer collaboration, where an AI tool pulls external stakeholders in around a shared unit of work. Light Table is his illustration again: the drawing set is the core unit of work, and developers, general contractors, architects and subcontractors all have to come in on top of it [1]. He also observes that some AI products now capture ACVs as large as or larger than the incumbent system of record off a single wedge before expanding into collaboration [1]. The corresponding problem for incumbents is structural: "they're going to get eaten from the outside and from so many other angles," and rebuilding themselves internally to support a faster product roadmap, go-to-market and pricing model is hard [1].

Hardware is where he thinks the more durable answer sits. "I absolutely love hardware. Hardware can capture things that our eyes can capture, and hardware can also capture things that our eyes can't see" [1]. The argument is about where error and effort actually originate: enormous amounts of time and human error go into capturing things with our eyes and manually keying them into a system, and once that data is in the system many downstream workflows fall out of it [1]. Dandy is the earlier version of this logic in his portfolio, a company that pivoted after noticing dentists given expensive intraoral scanners were ordering from other labs using those scanners, which led to thirty days of research on the dental lab market and a business that gives away the scanner, builds software on top of it and handles the manufacturing [3].

On founders, and the psychology he screens for

The person comes before everything else in diligence. He describes probing resilience, resourcefulness, sense of urgency and speed of learning as "absolutely critical pillars to a successful founder," and then a three-part sales test: can they sell stock, sell their product and sell people [4]. His proxies are behavioural rather than declarative. Urgency shows up in how fast and at what hour a founder answers email during a raise, because prioritisation is the skill the job will demand constantly [4]. Resourcefulness shows up when he asks a founder to walk through exactly how they got to the meeting, what customer development they did, who they talked to and how they reached those people, and it shows up again in whether they can articulate a genuine wealth of learnings between the idea twelve months ago and today [4]. He credits his read on people partly to background: "my mom's a therapist my dad's an entrepreneur" [4].

Market comes second, on the reasoning that "the product can oftentimes change the market changes far less," so even a pivot usually happens inside the same market [4]. He wants the "gotta believes" for the market to support a billion-dollar outcome, macro tailwinds that could double or triple it, and a 10x product capable of building barriers to entry [4].

On product-market fit, hiring and the shape of the early company

He treats a startup as an overarching hypothesis with a stack of sub-hypotheses beneath it, and each one proved out unlocks both more capital and the right to test the next [3]. He is emphatic that PMF is the hard part: "finding product Market fit that is by far the hardest part of a startup," and that founders who move fast, iterate and listen to customers eventually get there through sheer volume of attempts [3]. That belief drives his hiring advice, which cuts against the instinct to staff up. Founders themselves have to find fit with "a very very small very very Scrappy team of generalists," and only afterwards should specialists arrive, because a salesperson trained to execute someone else's playbook has never had to invent one [3]. Waiting, he argues, is more advantageous than being early [3].

On capital, milestones and the current gold rush

He pushes back on the idea that bootstrapping is surging: "the narrative of like people are bootstrapping at a much much higher rate is a little off" [2]. What he actually sees is a generation of founders fluent in venture language before they have raised anything, treating the next three years as a gold rush, and raising much larger early rounds "not only because they want to, but because they can," because there is a great deal of money looking for a home [2]. He also warns that a lot of capital early "might not give you the constraints needed to find true product market fit" [2].

Against that he sets a disciplined milestone framework. A seed check should buy 18 to 24 months of runway, preferably more, with the raise starting at least six months before cash-out [3]. From a $15 million entry he wants a 3x markup at the next round, and works backwards from whatever metrics the next investor needs to see, typically $1 to $2 million ARR for SaaS with a clear path to tripling, and something closer to $5 to $10 million for recurring-revenue consumer businesses [3]. Between rounds, Primary sits down with founders to write the "future equity story," agreeing what the Series A pitch will look like in eight to fourteen months and the path to get there [2]. On exits he is deliberately uninterested in acquirers: "we're not thinking about like who is the business going to exit to we're thinking about what does this business look like from a financial profile perspective in seven to ten years" [3]. The market correction taught the obvious lesson, that "growth isn't free and capital is not free and you can't just scale a business with zero percent gross margins and expect that the public markets are going to like you," so he wants capital-efficient businesses with sticky recurring revenue that can eventually be very profitable [3].

He applies the same first-principles logic to pricing, using buy-now-pay-later as the worked example: if a company's conversion problem is that customers cannot pay upfront and do not want a 28% APR credit card balance, you can quantify how many additional buyers a zero-percent-APR option produces, translate that into incremental revenue, and price the product off that value [3].

On non-consensus bets and the limits of pattern matching

He notes that Anthropic's Series A was passed on by every tier-one VC as a glaring case of the best companies looking wrong early [1]. His conclusion is not that pattern matching is useless: "Yes, I do think there's still pattern matching, especially when it comes to founders," and that the best companies of any generation "appear non-consensus at the very early stages" before becoming consensus over time [1]. What has changed is the price of being wrong, since "we are writing checks at much, much higher valuations than we did before" [1]. He has also discussed being non-consensus and right alongside Slow Ventures partner Will Quist, whose firm approaches investing very differently from Primary [6].

On how a seed firm earns the right to be chosen first

Primary's model rests on four choices he returns to repeatedly: concentration, with each partner doing only three or four deals a year; specialisation, with partners building multi-thousand-person expert and customer networks that inform insight, diligence, deal-winning and portfolio support; a large forward-deployed impact team, staffed at 30 to 35 people and growing toward 60, embedded in companies around go-to-market, engineering, product, outbound, recruiting, finance and corp dev; and incubation through Primary Labs [1][2][3]. The word "platform" is rejected internally as a dirty word with a weak connotation, where "impact" implies a north star around driving real value [2][4].

The economic defence of the impact team is blunt: "if it was just marketing why would we be putting money into this program? Why would we have 35 people into this program?" Those dollars come out of partner pockets, so the ROI has to be real [2]. He is equally blunt about investor humility, saying he would never try to out-advise Cassie Young on go-to-market or Rebecca Price on recruiting, and that "we are fooling ourselves as venture capitalists, especially if we have not operated and scaled the company in a big big way to believe that we can give better advice than the operators" [2][3]. The design constraint is ratios, because a sales resource covering five or ten companies a month works and one covering forty does not [2]. Operationally it runs on one-to-one founder relationships with impact partners, real-time channels for hires and customer wins, and a monthly line-by-line review of every portfolio company where the team identifies the "glass balls," the one or two things that cannot be dropped without putting the company in trouble [2].

Primary Labs, run by Brian Schechter, rests on a deflating premise: "VCS really aren't smarter than anybody else they just see a lot of stuff," so the job is connecting dots, taking a business model that worked in one industry and finding another industry that resembles it [3]. The output is a vetted idea handed to an excellent operator who lacks the time to generate one, plus a check on the day they resign, which raises conviction enough to make them leave [3]. He reports the studio spinning out four or more companies a year, with incubation seed rounds averaging north of $5 million on north of a $20 million valuation even in a bear market [3]. He has also worked on incubations directly with founders, noting AI has made that far more accessible "for a non-technical person like me" [1].

On how the venture landscape is shifting

He describes a decade in which the market resets roughly every five years [2]. The dominant force now is that the biggest firms are getting bigger, absorbing LP capital, investing across the full stack from seed to Series E, and organising into specialised pods, which lets their investors bring a prepared mind, move faster with more conviction, see everything in their space and build a sourcing brand through tweets, posts and events [2]. The danger for founders is that a seed check from such a firm is an option bet, which he dramatises with a question: if you own a $35 million house and a $2 million house and both are on fire, which do you run to first? The implication is to work with someone whose livelihood is seed investing [2]. Alongside this he notes the rise of solo funds and emerging managers with proprietary networks out of places like scale AI and Uber, who reach talent earlier than most [2].

Sourcing has changed just as sharply. Six years ago it was "night and day": a herd of investors writing $100k to $250k checks, with 30 to 50% of deals arriving through friendly investors and founders, a channel now down to roughly 10% [4]. The replacement is proprietary technology for finding deals, content marketing, purpose-built micro sites like the New York City Founders Guide, the equity-free Future Founders program, and thesis-first outreach into a market followed by surrounding themselves with its best operators and angels [4].

On marketplaces and democratizing access

His marketplace framework centres on promiscuity across both sides, since in Uber, Lyft or dash the demand side does not order from the same supplier and the supply side does not serve the same customer twice, which is what suppresses leakage of transactions off the platform [4]. He points to Bill Gurley's scorecard and the NFX team as the reference material [4]. For consumer businesses generally, repeat rate and the velocity of repeat matter most, and the goal is for a business to look like a subscription even when it is not [4].

Underneath the sector work sits a stated personal mission: "one thing that really like underlies almost all the investments i make these days is all about democratizing access," extending tools and services historically reserved for the wealthy or for large corporations [4]. In digital health that means improving outcomes, increasing access and cutting cost, especially bringing city-grade specialist care to what he calls care deserts as the population ages [4]. In fintech he is watchful about companies "loading up even more debt on americans" and prefers access plays, citing a company giving freelancers the accounting, lending and tax tooling big businesses already have [4]. He says he did not join venture to make money, and that the alignment with his mission of empowering others is the reason he is still in the game [4].

On the arbitrage of timing, and learning without permission

He reads his own beginning as a lesson about market conditions rather than passion. Launching a direct-to-consumer footwear company in 2011 worked because "Facebook was very cheap" and Shopify "was not good at all," creating an arbitrage in acquiring customers directly, and he is explicit that "by no means was I a good operator" [1][2]. He draws the parallel to now deliberately, treating the present as the same kind of window. The other lesson he draws from a period of high-pressure, self-directed work is that capability is not conferred: "you don't need to be told by anybody what you are capable of. go out there and learn and figure things out on your own." Information was already democratized then and is more so with AI, so "people that have drive and are just curious are going to thrive right now" [1].

Takeaways

• The bar for vertical AI is Stripe's: abstract the compliance, fraud and complexity, not just the transaction, because old vertical software "required so much interaction by the human being to input things into the system to get outputs" [1].
• Back the high-velocity wedge over the AI-native system of record, since a wedge demoable in thirty seconds converts above 50% and can move into the system of record as AI coding shortens migration times, while rip-and-replace grows too slowly to attract capital [1].
• Underwrite AI products against a named revenue bottleneck: takeoffs and quotes gate revenue in the trades, which is why compressing weeks of manual counting into minutes sells [1][9].
• Hardware is the durable data moat, because it captures what eyes can capture and what eyes cannot, removing the manual input step that feeds every downstream workflow [1].
• Screen founders on resilience, resourcefulness, sense of urgency and speed of learning, then on whether they can sell stock, sell product and sell people; diligence the market second, because products pivot and markets rarely do [4].
• Do not hire specialists before product-market fit; founders find fit with a small scrappy team of generalists, and a salesperson brought in early has never had to write the playbook [3].
• Work backwards from a 3x markup: from a $15 million entry, target the metrics the next investor needs, roughly $1 to $2 million ARR for SaaS with a path to tripling, on 18 to 24 months of runway with the raise starting six months before cash-out [3].
• Choose an investor whose livelihood is seed: for a multistage firm your seed check is an option bet, and when the $35 million house and the $2 million house are both on fire, they run to the big one [2].
• Investor value-add has to be staffed, not claimed; Primary runs 30 to 60 impact people against a concentrated portfolio because a sales resource covering ten companies works and one covering forty does not [2][3].