Invest Like the Best · Patrick O'Shaughnessy、Christina Cacioppo · 2026-10-20 · 57 min · 原语言 EN

Christina Cacioppo says Vanta’s 400-integration “trust graph” helps the roughly eight-year-old company build software and agents faster. About 25% of customers use agent flows; she expects raising that share to 75% to improve retention and customer experience. Vanta absorbs token costs rather than passing them through, since markups invite Claude comparisons, leaving commercialization and margin recovery unresolved.

TL;DR

• Cacioppo attributes Vanta’s category win to execution: “you’re really only as good as the last thing you shipped”—someone can vibe-code your thing in a weekend. A structural advantage is the “trust graph”: a security data lake fed by 400 integrations, which she says has let Vanta build new software and agents faster. The company is about eight years old, with ~16,000 global customers and about five products.
• The “SaaSpocalypse” was her scariest moment, and living through the repricing is “cognitive dissonance.” In the fullness of time, she thinks an 80% growth / 80% margin business returns more cash than a 160% growth / 30% margin one, “but today, the latter is valued at 3X, 4X… it’s 10X the multiple of the former.” AI-maximalist investors tell her they’ll fund “the labs, and… basically robot tractors in Iowa, nothing in between.”
• Vanta is eating token costs rather than charging customers for them. Customers notice a markup on a Cloud token and ask whether the application is “4X as good as Claude”; token pricing puts an application in direct competition with the models, and credits do not hide that. Twenty-five percent of customers are in the agent flow; if Vanta can get that to 75%, she expects better retention and experience. “This is a usage phase… nail that, and you can figure out commercialization after.”
• Overhyped: token pricing, low margins for their own sake, post-training “for the marketing value”—“mostly for investors, actually”—and FDEs. She says post-training can make sense for cost in some cases, and a customer-specific model may be reasonable for some enterprise deployments, but doubts that handing customers a model is the right end state. FDEs are often professional services under a more fashionable label; the economics may work for eight-figure contracts but are probably upside down at five- and six-figure ones. Underhyped: company and customer context over functional skill—she’d take a worse salesperson who knows Vanta.
• Total compensation for Vanta’s top person is probably about 10 times what it was five years ago. She says top people’s value is accelerating, that Vanta should have more comp dispersion, and that “you never pay the top people enough.” Tenders are normalized and expected as part of compensation; she calls the situation unhealthy but says her opinion does not change the game on the field.
• Security’s threat picture is shifting: AI models can persistently try vast numbers of things and chain benign-seeming findings together. Patrick framed the future as a constant white-hat penetration-testing agent with a patching “buddy agent”; Cacioppo said that is possible for parts of Vanta, especially for customers who may not pay Wiz’s enterprise-level prices.
• Her broader thesis is that AI has ushered in a “high-excitement-but-low-trust” era, and distrust of tech companies is worsening. She expects proof that software is industrial-grade to matter more to businesses, governments, and regulators. She imagines Vanta could sit at the center of those transactions as it expands from compliance into more security products. A field datapoint: French software companies told her they switched predominantly to Chinese models over business-continuity concerns, not nationalism.

Digest

1. Execution won the category; the trust graph is an advantage

• Vanta at about eight years old: ~16,000 customers globally across about five products, built on the premise of helping companies “build and earn trust”—starting or maturing security and compliance programs, then getting customer credit via audits such as SOC 2, ISO 2701, HIPAA, questionnaires and trust centers. Asked how she beat a crowded field: “Execution… You can be first, and you can start it, but that doesn’t really matter. It’s even more true now ’cause someone can vibe-code your thing in a weekend.”
• The structural answer is integrations: Vanta connects to 400 tools and pulls data into its “trust graph,” which she describes as “really just a security data lake.” She says it is broader than other providers’ data foundations, including some security providers’, and has let Vanta build new software and agents faster.
• Scariest moment: the last 12 months. “SaaSpocalypse. You’re like, ‘Okay, great job building a system of record. You’re now screwed.’” Parts true, parts not—the forcing question being: “Great, what is the agent? You have all this data and a system of record. Fine. What the heck are you doing with it?”

2. A great business, repriced: living with cognitive dissonance

• Patrick’s setup: Vanta is growing fast with very high margins—business traits that four years earlier might have been considered the best model ever—yet the multiple is a third or a quarter of what it was. Her answer: hold two things at once. “An 80% growth business with 80% margins will actually be better” in the fullness of time than 160% growth at 30% margins, “but today, the latter is valued at 3X, 4X… it’s 10X the multiple of the former.”
• The game on the field: financing runs through AI maximalists (“we’ll invest in the labs, and we’ll basically invest in robot tractors in Iowa, nothing in between”), and ambitious employees “if it feels like they’re not in the right boat, they’re gonna change boats.” Her stance: keep doing what makes business sense—“which may or may not be token-pricing models”—and “live for the fullness of time.”

3. The agent is now the product—and she says nobody is further along

• Benchmarking the Vanta agent against the names you’d imagine as “AI applications,” “we have yet to find someone who is further along”—while acknowledging, “we’re doing a bad job looking” is the obvious critique. Behind the product-level “Vanta agent” sit hundreds of task-specific agents with tools and computers; the front end has long-running agents that auto-generate UI on the fly, “‘cause sometimes you want chat and sometimes you want UI.”
• The grounding example: an employee leaves and isn’t removed from every account. The agent: “I just wrote a custom agent for you that will go do that. Do you want me to turn this on? … Do you also want me to remove access from these three people?”—borrowing lab UI patterns: “You want me to do this once? You want me to do this always?”
• Twenty-five percent of customers are in that kind of agent flow today. If Vanta can get that to 75%, she expects better retention and a better customer experience; pricing can be worked out later.
• The trade-off she’s living: building agent-first rather than building a lot of SaaS UI (“we’re not gonna put Tableau inside of Vanta,” which Vanta might have done in 2021 or 2022) against enterprise reality. A U.S. enterprise asked whether Vanta could turn its AI features off because non-AI tools clear procurement faster. Patrick framed the diffusion risk as being “right, but early and therefore wrong.” She says they are not going to build an off switch.

4. Pricing: don’t put yourself up against Claude

• Why token pricing fails: customers are “really smart” and may say, “You’ve just marked up a Cloud token 4X. Are you 4X as good as Claude?” It puts the application in competition with the models—“no application really wants to” be there—and credit-token exchange rates do not fool users. “We’re eating the token cost right now” and taking lower margins.
• Everyone wants outcome pricing—the former microeconomics professor in her loves its incentive alignment—but she distinguishes outcome from output: a support ticket closed is an output; “the outcome is actually my Sonos now works.” Agreeable units of work are hard to settle on: “I just named three things. I could name another 47.” That creates cognitive overhead for users comparing agent work with doing the task themselves.
• Model selection runs on task- or agent-specific evals written by subject-matter experts, including people who ran security and compliance programs. The experts evaluate their evals; model swapping is “mostly about quality, not cost.” Vanta’s decision-and-action work uses fewer tokens than coding agents; legal and document drafting may use more.

5. Building with AI inside an eight-year-old company

• Asked how much AI changes company-building, one to ten: “Eight or nine. The principles are still the same… all the tropes are still the tropes, but the way you do stuff should be quite different.”
• The legacy-codebase problem: agents lack the context window to understand eight years of code, “will do silly things,” and need “a high-context engineer to babysit it,” which can undermine the point. The fix—where she says Notion is about nine months ahead—is modularizing so agents can fully understand smaller parts. Companies always had this as new-engineer onboarding friction; “agents have made it, ‘Well, great, now we want to onboard 9,000 new engineers tomorrow.’” She gives Vanta a lower grade here.
• Adoption 101: use AI yourself and build something for your company; check at least every two weeks because things change quickly. That gives you a visceral sense of what works and helps you question claims like “we’re gonna fire all of our SDRs ’cause I built one agent.” She has become more empathetic toward strong people who have not dived in: find what is blocking them and help them get into the work.
• The onboarding-wizard parable: a talented, motivated team kept bringing SaaS onboarding wizards instead of an LLM flow. The real blocker was: “We don’t know if we can get the drag-your-files thing to work… We basically don’t think we’re allowed to fail on this project.” Her response: bet on the LLM flow and “burn the boats.” The Claude-style big canvas—drag in messy files and screenshots, let the system make sense of them and ask questions—shipped and works well.

6. Talent wars: trillion-dollar pitches and 10x compensation

• The pitch her people get elsewhere: “We’re $6 billion today, but Sequoia’s gonna do a $10 billion check in a month, and then somebody’s gonna write it up to $25 billion next year, but we can get you in at six today.” She’s considered counter-escalating (“the real question… can you take us on a journey to, like, $250 billion?”)—“Probably not, but the answer is not definitely not.” Her actual playbook is “do what it says on the tin,” show candidates the metrics, and “counter the vibes with data and trustworthy vibes versus YOLO vibes.”
• Total compensation for Vanta’s top person is “probably a factor of 10” compared with five years ago. She says Vanta should have more comp dispersion and that the value of top people is accelerating. What earns it: “High agency, get shit done. That sounds so milquetoast, and that’s worth millions of dollars.” She says you never pay top people enough, though there are also people paid much more than their contribution. The counterforce would be investors reviving the dilution/SBC conversation—“now we’re definitely not talking about it.”
• One retention case: a strong leader entertained an offer promising “full decision-making authority.” Her response was to unblock them: “What do you need to feel like you can make these decisions? Just go make these decisions.” On tenders: Vanta had not historically done them, but “the game on the field is that you’re doing a tender every time.” They are normalized and expected as part of compensation; without them, companies need to offer other things. She agrees they feel unhealthy but says her opinion does not change that expectation.
• Exit: an IPO is a means to an end, not her personal bucket-list item. The goal is a “big, independent, standalone company.”

7. Overhyped: post-training theater and FDEs; underhyped: context

• Overhyped: “token-based pricing, low margins for the sake of having low margins, and custom training models for the marketing value.” She says inference providers will often say the reason to post-train is cost, not quality; some businesses have a cost need, and a customer-specific model for an enterprise use case can be reasonable. Her broader skepticism is that customers may not want to be handed a model: “reasonable iteration, but not the right solution.”
• FDEs (forward-deployed engineers): she calls them “a great talent arbitrage, the same way a chief of staff is,” but says the role is often professional services under a different label. She doubts a high-agency Harvard CS graduate will sign up for a conventional pro-services role. Palantir is the company she thinks does this well, with eight-figure contracts; at five- and six-figure contracts, the economics are probably upside down.
• Underhyped: company and customer context over functional skill. She would “take a worse salesperson on sales skill who knows the Vanta pieces,” and thinks labs building go-to-market teams by hiring good salespeople may be undervaluing context.
• Field note from France: Cacioppo says many Paris software companies she met use Chinese models predominantly. She initially guessed nationalism or anti-American sentiment, then was corrected: customers told her they switched after what they described as a U.S.-model access restriction, citing business continuity rather than nationalism. The specific model/service name in the transcript is unclear.

8. Security at machine speed—and where she’d invest

• The models “are just so much more persistent than people. They’ll try nine quadrillion things, where a person would give up after 9,000.” The shift is not only zero-days: she says models can chain “17 benign-ish” things together. She cites an OpenAI RSA talk about Hugging Face and the difficulty of dealing with findings that might previously have been considered low severity and fixed later. “If you’re fixing things at machine speed, you don’t have to prioritize”—though “it’ll be rocky as we get there.”
• Patrick pictured a constant white-hat penetration-testing agent with a “buddy agent” that patches what it finds. Cacioppo said that is possible for some parts of Vanta, especially for its customer base; she expects CrowdStrike to do it better for enterprises. For cloud misconfigurations, she contrasts Wiz’s multi-thousand-person-company customer base and roughly $150,000 starting price with Vanta’s focus on already-embedded customers who may pay “40 or 50” for a solution suited to them.
• Vanta sees itself as a security company that began with compliance: “The best version of compliance is proving the security you have.” It plans to build more security-focused products on top of its existing data and agent intelligence. As companies face growing pressure to prove they are industrial-grade to enterprises, governments, regulators and others, Vanta aims to connect both sides of that marketplace.
• Wearing her old investor hat (“do not envy the job at all”), she says certainty around open-source models feels disconnected from how underbuilt the ecosystem around them is; “a bunch of the lab playbooks seem like they should apply to the open-source models.” Green shoots: cybersecurity founders, rev tech (some incumbents are choosing PE-bent, sponsor-style playbooks instead of building), robotics—which she sees as underweighted despite being somewhat hot—and government modernization. AI-related risks are helping CISOs make the case for enterprise-risk budgets.

9. The billboard saga, and rebuilding trust

• The billboard war story: a billboard skeptic (“founder ego is why someone would do a billboard”) was persuaded by PMM Sarah, who went around her boss with “Compliance that doesn’t talk too much.” Cacioppo laughed, checked with investor Andrew Reid, and approved it. The billboard went up near Burning Man; the tweet went viral and the line circulated on r/accounting. She says it worked because it expressed Vanta’s wink: taking the work seriously without taking itself too seriously.
• In 2022, the market crash made billboard inventory available, and Vanta got a good spot at the 101/280 intersection. Despite her saying on podcasts that the contract was locked up for years, it was just a placement order. A startup took the spot, and her offer to pay $3 million—double the roughly $1.5 million the other company was paying—failed when its founder said, “We’re actually cash flow positive.” “I am an extremely inelastic buyer here” turned into “literally, I’m an inelastic seller.” She half-jokes that losing the spot cost Vanta market cap.
• The closing thesis: “We just don’t trust software companies, tech companies the way we used to, and that is getting worse year over year.” She agrees with Patrick’s framing of a “high-excitement-but-low-trust” era. Vanta’s job, in her view, is to rebuild trust “legitimately so, not just on vibes.” In B2B, she favors specific data, improvement curves and possibly contract-backed promises; consumer trust is more about vibes and “seems very hard to turn.” Asked about big companies using personal agents, she says no: she would rather have Google, Apple or another company she chooses hold her email or chats. She imagines Vanta could sit at the center of transactions that demonstrate software is trustworthy and industrial-grade; the $250 billion figure is part of her aspirational framing, not a forecast.
• Her motivation has shifted from “You don’t think I can get off the starting line? Let me show you”—“still a little dark edge, but less of it”—to “let’s see how far we can go with this.” One word for Vanta’s next five years: “Expansive.” The company has an established base of customers and data, and she sees room to build on it.

来源：Podcast

Read on BidClub: https://bidclub.ai/e/christina-cacioppo-building-enduring-software-co