The Build vs Buy Head Fake
In January 2026, Anthropic released the massive ecosystem launch of Claude Cowork. That was followed in February 2026 with the launch of a set of industry specific plugins, which triggered one of the sharpest repricings the software sector has ever seen. Nearly a trillion dollars of market value evaporated from software and services companies in a handful of sessions. It was dubbed the “SaaSpocalypse.”
The logic behind the panic is not irrational. If a general purpose model can read your files, write documents, pull data, and orchestrate teams of sub agents, why pay for a seat in a legacy application? If AI coding tools let a product manager or even someone who isn’t technical at all stand up a working internal tool in an afternoon, why buy vendor software at all? The cost of creating software is collapsing toward zero. For software investors and founders, that is a genuinely destabilizing idea.
We think that framing is directionally right about one thing and wrong about the thing that matters most. The narrative mistakes the easy part of software for the whole of it, and in doing so, it accidentally makes the strongest case for why software is and will continue to be a great place to invest.
“Build vs Buy” Was Never About the Cost of the First Build
The most useful way to see through the panic is to actually run the build vs buy decision the way an operator does. Jamin Ball lays out the argument clearly. The “everyone will just build their own” argument isn’t new; it’s the same calculus enterprise operators have run for decades, just dressed up in AI clothing. The answer has always come back the same way: building is rarely the hard part. Companies have always been able to build; they most often choose not to. What’s hard is everything that comes after. Vendors offering specialized software aren’t just selling a product; they’re selling the accumulated R&D, the security and compliance infrastructure, the customer success apparatus, and the operational accountability that a homegrown tool has to replicate in full before it’s a real substitute. AI makes it faster to write the first version, but it doesn’t make everything else that comes with building software cheaper.
The initial build is the smallest part of the commitment. Depending on the complexity of the system, the first working version might represent a fifth or less of the total effort a company will eventually spend; the rest goes to keeping it running, adapting it as requirements change, patching vulnerabilities, and managing the knowledge transfer problem every time someone who understands the system moves on. AI compresses the first chapter of that story, not the rest. And when hundreds of companies each rebuild the same internal tooling independently, the waste is staggering, not just in engineering hours, but in organizational attention, operational risk, and the compounding cost of maintaining something that generates no competitive advantage.
There’s another cost that’s easy to underestimate: inference. Token costs at low volumes feel negligible in a demo; at production scale, across hundreds of employees running dozens of daily workflows, they compound quickly into a meaningful line item. A software vendor amortizes those infrastructure costs across thousands of customers, continuously optimizes for efficiency as new models ship, and absorbs the operational burden of managing rate limits, latency, fallbacks, and model upgrades. A company running its own internal tooling bears all of that alone, and unlike a SaaS seat, the bill scales directly with usage. Building gets cheaper every month; owning the infrastructure behind what you built does not.
The early data backs this up. Retool’s 2026 Build-vs-Buy survey found that 35% of enterprise teams have already replaced at least one SaaS tool with a custom build, and most plan to build more. But the categories being displaced are telling: internal admin tools, BI dashboards, simple workflow automations, the shallow, generic, horizontal layer. Meanwhile, the failure rate of ambitious AI initiatives remains sobering: Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. Industry surveys keep finding that the overwhelming majority of agent pilots never reach production. Most ambitious AI deployments don’t stall because the model or the code wasn’t good enough. They stall because the underlying data was messier than expected, the integrations more fragile, the upkeep more time consuming, the governance questions more fraught, and the domain specific edge cases more numerous than any general purpose tool anticipated. Those are organizational and institutional problems, not technical ones, and they don’t get solved by shipping a better model.
The real risk to incumbent software isn’t the DIY internal build. It’s that when the cost of creating software goes to zero, ten well funded teams can rebuild a category from scratch, AI native, priced for the new world, and commoditize the incumbents. This isn’t to say that all software is safe. Generic, easily replicable software will be exposed, which has always been true. What protects a business is specificity, and that’s a case we’ll make below.
What Actually Stays Defensible
Strip away the code and ask what a software business is really made of. The value of software was never in the lines of code themselves; it was in the accumulated understanding of how a specific set of people do a specific kind of work, translated into something durable and repeatable. The things that make software hard to displace; deep customer relationships, institutional trust, distribution earned over years, and a product shaped by thousands of hours of customer feedback, have nothing to do with how the code was written, and don’t get cheaper to replicate just because the next model shipped.
That reframing points directly at the moats that survive.
Proprietary, compounding data. Not “we have customer data,” but data that feeds back into the product and improves it in ways a new entrant starting from zero cannot replicate. The key isn’t just accumulating a data lake; it’s inserting your product at the point where data is first created. When you automate a previously manual workflow, you generate an entirely new data exhaust that never existed before: you build the dataset by doing the work. Abridge’s proprietary dataset is derived from more than 80 million medical conversations; Harvey works across legal documents in dozens of countries; EvenUp sits on hundreds of thousands of injury cases. Every time a denied claim is analyzed, every time our portfolio company Freight Hero handles a freight booking call, every time our portfolio company Superpanel processes a legal intake, the product learns. That learning compounds. Over time, the product that has processed 100,000 intake calls in a specific vertical has a structural advantage that cannot be replicated by a new entrant, not because the features can’t be cloned, but because the data cannot.
Systems of record and workflow embedding. The stickiest software in the SaaS era was the system of record, the place where every other workflow ran through. Replacing it meant touching everything, and that switching cost compounded quietly for years. The equivalent dynamic in the AI era belongs to whoever becomes the governance layer for agents: the system that decides what an agent is allowed to know, access, and do. That’s a fundamentally different kind of lock in than storing data; it’s about controlling the rules of engagement for autonomous action. The broader shift underway is from software that records what happened to software that decides what happens next, actively reasoning over accumulated data and executing under defined guardrails. The companies best positioned for that shift are the ones that inserted themselves early, at the point where work begins, where the first form is filled out, where the first call is logged, where the first document is created. Own that entry point, and every downstream step in the workflow becomes a natural expansion surface.
Speed and distribution. No moat lasts forever; they have always been fleeting, that has always been true, and AI hasn’t changed the rule, only the pace. What used to be a durable advantage for 12 to 18 months can now be approximated by a capable team in weeks. The implication is that defensibility isn’t a destination you arrive at; it’s something you have to keep earning. The companies with staying power aren’t necessarily the ones with the best technology; they’re the ones that have already embedded themselves in how customers buy, procure, and operate. Enterprise sales cycles in complex verticals routinely stretch from six months to a year. Legal, security, and procurement reviews don’t compress just because a better product exists. A customer who signed two years ago and has since trained their team, integrated their data, and built workflows around your product isn’t switching because a competitor shipped a sharper demo. That accumulated friction, earned through real deployment, not feature development, is one of the few advantages that genuinely doesn’t clone.
Notice that none of these are “we have a better model.” As Harvey’s own trajectory shows, the company orchestrates multiple models under a legal specific layer; the model is increasingly the commodity. The moat is everything wrapped around it.
Why Vertical AI, Specifically, Is Built for This World
Every one of those durable moats is easier to build and deeper to hold in a vertical market than in a horizontal one. That’s not a coincidence; it’s structural, and it’s why the “AI eats software” story is, for vertical AI, closer to a tailwind than a threat.
Depth is the defense against commoditization. Horizontal tools optimize generic workflows that a good model can approximate. Vertical products encode the messy, specific, regulated reality of one industry: the multi step workflows, the payer specific denial rules, the code compliance edge cases, the union specific payroll logic that a general model gets 80% right and 100% wrong in the ways that matter. Vertical AI companies compete on proprietary data, workflow depth, and compliance, precisely where “good enough” DIY software is at its worst.
The prize is labor, not software budgets. This is the point we’ve hammered at Field Ventures, and it’s now the consensus framing: traditional SaaS captures a sliver of an employee’s value; vertical AI can capture a large share of the labor itself. This reframes the commercial conversation entirely. When your customer’s alternative is a full time employee or an offshore BPO contract, the calculation is not abstract: each of my intake coordinators costs $65,000 per year and makes mistakes. Your software does the same work with a lower error rate, at half the cost, and is available 24/7. That is not a productivity pitch. That is a P&L pitch. The ROI is unusually clear, and it unlocks budgets that were never available to SaaS vendors before.
Under digitized industries leapfrog straight to AI. The antiquated verticals we love, construction, logistics, healthcare administration, the trades, and many others, run on slim IT budgets and enormous labor budgets, with workflows still living in spreadsheets, faxes, and phone calls. They were never going to build their own software; they barely bought the last generation of software. And importantly, the incumbent in these markets is often a spreadsheet, a phone call, a manual process, or software built in 2003 that has barely been updated since. There is no sophisticated incumbent to displace. There is only operational drag that the buyer experiences every single day, and they would gladly pay to eliminate it.
Agents accumulate context; that context becomes a moat. An agent, like an employee, gets more valuable the longer it’s deployed and the more institutional context it absorbs, a genuinely new switching cost that compounds over time. Unlike traditional software, which does the same thing on day one as it does on day one thousand, an AI agent deployed in a real production environment gets better over time, absorbing institutional quirks, edge cases, and workflow preferences that no new entrant can replicate from a standing start. The longer it runs, the more it knows about how this billing team handles denials, how this logistics operation manages exceptions, and how this legal practice structures intake. That accumulated context is a switching cost unlike anything the SaaS era produced, because replacing the software means losing the institutional memory embedded in it. That is a moat that accumulates quietly and compounds over time, one that didn’t exist in the SaaS era and can’t be bootstrapped by a new entrant overnight.
Domain experts are now the builders. Domain expertise has become more important, not less. AI has also changed who can build. A solo founder with Cursor, Replit, and Claude can ship in days that would have taken a team of four six months to build two years ago. That democratization matters; it means domain experts (the paralegal who has spent fifteen years in mass tort litigation, the billing specialist who has processed ten thousand Medicare claims) can now build the tools they wished had existed, without needing a technical co-founder to translate their vision into code. But the consequence is that a compelling demo is no longer a signal. When anyone can build an impressive product in a week, it tells you almost nothing about whether you have a compelling business. Defensibility moves upstream: to workflow depth, to trust, to distribution. The founders who win will be the ones who got to the right problem before anyone else, because they were living inside it.
The Counterarguments and Why They Don’t Break the Thesis
“The foundation models will just enter the biggest verticals.” They’re already doing it. OpenAI and Anthropic both launched healthcare offerings in early 2026, and OpenAI has signaled a legal product. This is real, and it caps the upside in the largest, most horizontal adjacent use cases. But it also validates the verticals, and the labs consistently take the low hanging fruit while struggling with exactly the things vertical players own: messy data integration, industry specific compliance, editorially maintained authoritative data, and multi party workflows. The durable answer isn’t “out model the labs”; it’s staying model agnostic and building the data, workflow, and trust layer they won’t.
“The real threat is the next AI native startup, not the incumbent.” This is the version of the risk we take most seriously. When the marginal cost of shipping software collapses, a small, focused team with access to the same frontier models can close what used to be a meaningful feature gap in months rather than years. First mover advantage is a much weaker defense than it used to be, and at the earliest stages, moats are non-existent today. This is precisely why structural advantages are crucial: position over product leads, proprietary data that accumulates and compounds with usage, regulatory credibility that takes years to earn, distribution embedded in how customers already buy, and workflow depth that a new entrant would have to replicate from scratch. Features are a snapshot. Structural advantages compound.
“Foundation models will commoditize every application built on top of them.” This critique is accurate about one category of company and largely irrelevant to another. If your product is a thin interface sitting on top of a foundation model with no proprietary data, no workflow depth, and no switching costs, then yes, the model provider will eventually absorb your use case. The market has already priced that in; capital has moved away from pure interface plays. But that description doesn’t fit companies that own a critical workflow, carry irreplaceable customer data, and have embedded themselves in how an industry actually operates. Those businesses aren’t wrappers; they’re the layer the model runs inside of.
“Incumbents will catch up.” Some will; many are trying. But most legacy application vendors have, so far, shipped little meaningful AI despite having every structural advantage, which is itself an opening for AI native vertical entrants who build the system of work, not just the system of record.
What We Look For
After assessing the founding team, where we believe deep industry insights are more crucial than ever, we look for three things: a convincing wedge, durable differentiation, and operational reliability.
The wedge is the entry point, the first, specific workflow the product solves, narrow enough to demonstrate value clearly, valuable enough to command budget, and positioned close enough to revenue generation or cost elimination that the buyer can see the ROI immediately. Wedges are easier than ever to create. The wedge alone means nothing.
Durable differentiation comes from domain depth and data positioning. The best vertical AI companies are not just automating tasks; they are inserting themselves into the source of data creation. The companies that build and maintain defensibility are the ones that sit at the data creation layer, where the first input is created, where everything downstream flows. Win that entry point, and you earn the right to expand into every subsequent step in the process. We are looking for founders who understand the human workflow deeply enough to automate it appropriately, not just what tasks can be handed to software, but which tasks must remain human in the loop, and why. The best founders in this category can draw a clear line between what AI can handle today and what requires human judgment, and they can articulate exactly where that line moves as the technology improves.
Operational reliability is perhaps the least celebrated but most important evaluation criteria. A demo in front of a sophisticated buyer is table stakes. What separates a pilot from a platform is whether the product performs reliably under actual production conditions, with all the edge cases, regulatory requirements, and operational messiness that never appear in a controlled demo. In verticals like healthcare, legal, and insurance, buyers are not asking whether the AI is capable. They are asking whether they can trust it. Switching costs are not an accident. They are a product design decision.
The Opportunity That History Keeps Confirming
The February 2026 selloff will, we suspect, be remembered the way the dot com and cloud transitions are remembered: a moment when the market correctly sensed disruption and wildly mispriced its shape and timing. The market has already started grading that mispricing. Situational Awareness LP, the AI focused fund built around the thesis that foundation models would gut software incumbents, had constructed exactly the trade this narrative implies: long AI infrastructure, short software names like Adobe. Last week, that position unraveled fast enough that the fund had to offload roughly $16 billion of its long and short book to Citadel in a single block trade just to meet margin calls. Assets that stood near $45 billion at the start of the month finished around $10 billion. The fund that bet hardest on the “software is dead” thesis was also the fund forced into a fire sale by it, a costly real time reminder that being early to the right narrative and being right about the trade are not the same thing.
Every prior platform shift ended with more software, not less. The PC didn’t kill the mainframe; it expanded the total compute market by orders of magnitude. eCommerce didn’t kill retail; it restructured it, creating trillion dollar outcomes on both sides. Streaming didn’t kill media; it created more content, more consumption, and more business models than anyone predicted. In each case, what died was the shallow, generic, undifferentiated middle. What thrived was whatever owned the critical data, orchestrated the critical workflows, and earned the trust to act.
That is a near perfect description of vertical AI.
AI making software cheap to build doesn’t erode the vertical thesis; it turbocharges it. It collapses the cost of automating a workflow while leaving everything hard about a specific industry entirely intact: the data, the regulations, the integrations, the trust, and the domain knowledge locked in experts’ heads. The build vs buy calculation has not been repealed by AI coding agents. It has been sharpened. What gets commoditized is generic, shallow, horizontal software. What gets more valuable is deep, industry specific software that encodes regulatory knowledge, owns proprietary workflow data, sits at the data creation layer, embeds in systems of record, and carries the trust to act autonomously.
And what gets truly valuable, in a way the SaaS era never allowed, is software that stops competing for a line on the software budget and starts competing for a line on the labor or operations budgets, priced against headcount, not seats.
The markets that were too unglamorous to attract the last generation of software builders are, for that exact reason, the most fertile ground for this one. The industries furthest behind are the ones with the most to gain and the fewest entrenched software incumbents to dislodge. The build vs buy head fake is real. But for founders and investors who know where to look, what it reveals is not a threat. It’s a map.
At Field Ventures, we’ve backed domain experts selling outcomes into labor budgets in under digitized verticals since before “generative AI” was a headline. We believe it more today than when we started. If you are building along this thesis, please reach out!