In April, the Tech-Software ETF (IGV) had fallen 35% from ts peak, which we called it, “The SaaSpocalypse”. Investors stopped assuming that a good product with high switching costs is an annuity. That assumption underpinned two decades of software valuations, and it is not coming back in its old form
Foundational models are replacing / disintermediating applications (e.g., coding)
AI makes switching cost significantly lower
New entrants / challengers can compete with incubments with less resources
End-markets are building their own software
Software margins are compressing due to inference cost
The prize is services, not software
For the past two decades, software was built to help people do work. It never did the work itself, so it could only charge for the tool. AI changes that. When software can apply judgement into workflow contexts, it can start delivering work beyond tools
In 2024, software had grown to about $650 billion, with cloud at roughly $400 billion. Meanwhile the services market, the money companies spend paying people to do work, sits above $10 trillion. Software has always been a sub-$1 trillion business living next to a market more than ten times its size.
Today, we are already seeing meaningful vertical applications across verticals:
EliseAI runs an 24/7 leasing and maintenance agent, handling lead qualification, resident inquiries, and after-hours requests so property managers can focus on resident experience, renewals, and property performance
OpenEvidence acts as a research assistant, synthesizing the latest clinical research and peer-reviewed literature so physicians can spend less time searching for evidence and more time making clinical decisions
Hebbia handles the scaffolding across deal research and document creation, pulling insights from filings, transcripts, and thousands of deal documents so bankers can iterate faster and spend more time on client delivery
Trunk Tools gives construction teams instant answers from drawings, submittals, and project documents in the field, helping project managers resolve questions faster and catch compliance issues earlier
Vertical breakdown analysis
Value accrues at the application layer
If AI unlocks the services budget, who captures it? My answer is the application layer, not the models or the infrastructure beneath them.
Picture the AI stack as five layers: (i) energy, (ii) data centers and GPUs, (iii) compute, (iv) tokens, and (v) applications. The first four are priced based on unit, per unit of power, rack, chip or token. They are priced on supply and demand, and supply is growing fast. History says that kind of layer earns a return on capital, not a monopoly rent. It also matters that most workflows don't need frontier capability, which keeps pressure on model pricing.
Applications are structurally different for four reasons:
They sit closest to the work. They own the workflow, the customer relationship and the context needed to deliver a measurable outcome.
They turn intelligence into work. A model produces tokens; an application produces a closed ticket, a filed claim or a signed lease.
They have a natural path to usage and outcome pricing. Once the app owns the process, it can charge for the result rather than the seat.
Their context compounds. First-party context accumulates with use, until the application becomes a ledger of the institution's knowledge that is hard to replace
Application software is a stored process. It encodes opinions about how a function should run, and those opinions calcify over years into something inseparable from the organization. The deeper workflows penetrate, the more process engineering matters and the harder it is for a challenger to cop
Four Enduring Models
When everyone has access to the same intelligence, a feature is worth less. The companies I find most durable are building defensibility somewhere outside the software itself. I group them into four models
Service to Differentiate Software
Data to Differentiation Software
Marketplace to Differentiate Software
Software to Differentiate Services
Model 1: Services to differentiate software
A lot of software are packaged like off-the-shelf products sold at zero marginal cost with little support after the sale for customers to realize the value. For example, a CRM works roughly the same on day one as on day ninety. AI applications are different. It needs configuration, evaluation, workflow redesign and real organizational change before it produces value
Consultative implementation is the new CAC
AI application companies spend far more on implementation than traditional software companies did. They deploy forward-deployed engineers (FDEs) who embed themselves with the customer and work through the technical and organizational changes required for adoption. In old SaaS terms, this looks like vulnerability and growing pains. In practice, it is customer acquisition and retention spend, moved downstream to where it actually converts.
The FDE sits between two worlds. On one side is the commercial world: workflows, cost, incentives, risk, adoption and business value. On the other are the technical capabilities: models, systems, APIs, data, code and reliability. The FDE's job is the integration layer that connects the two, usually through three processes:
Audit: map how processes are done today, identify opportunities where AI is well positioned to automate (verifiable workflows that are repetitive and costly)
Evals: define evaluation criteria and verification loops, then measure and iterate on pilot results
Deployment: embed the system into daily work, train the team, and monitor as volume scales
EliseAI’s sample engagement roadmap
Model 2: Data to differentiate software
Most enterprises already spend heavily on data, even if it may not be a clear line of expenses in their financial statements. Analysts spend hours sourcing, cleaning and enriching information before they can use it. A better data product, plus the means to capture data that is currently lost, is an easier sale than another workflow tool because the budget and the pain already exist
Data integrated into the workflow compounds
When data and workflow live in the same system, the software turns into an intelligence layer enriched with first-party context such as deal history, call notes and outcomes.
There are three types of data that live in these applications:
Third-party data (scraped or licensed)
Uncaptured first-party context (transcripts, notes, recordings)
Systematically captured first-party context (events, data feeds, connections)
New entrants are mostly going after the first two buckets. Third-party data and uncaptured context are exactly the parts incumbents tend to leave fragmented, which gives a newcomer the opening to pull them into one system.
Path 1: Third-party data integration to core workflows
Clay and Rox are the poster children for horizontal use cases where they act as a shared data spine for go-to-market teams, combining CRM records, lead and industry data platforms, web traffic and social engagement, then pushing tailored outreach and ad campaigns at scale
Vertical versions are also being built for specific vertical needs:
Hebbia (bankers and investors) pulls in SEC filings, earnings materials, press releases, podcasts, credit ratings and market data, then runs agents for modeling, visualization and presentation creation.
NationGraph (companies selling into government) aggregates meeting minutes, budgets, RFPs, contracts, grants and SLED directories, with agents for targeted outbound, contract and FOIA requests, and ranked opportunities.
Prophetic (land buyers in infrastructure and CRE) layers geospatial, ownership, zoning, comparable sales, utilities and nearby development, with agents for parcel screening, site assessment and yield studies.
Path 2: Uncaptured first-party context providing enriched insights
These companies are building a new category of data infrastructure: information that was always being generated, but was never systematically captured.
The clearest wedge is voice AI. Granola and Gong capture conversations that historically disappeared after the meeting, then enrich the CRM and other systems of record. The important insight is that these products are not just automating note-taking. They are creating a proprietary layer of context that does not exist anywhere else. As that context compounds, so does the data gravity around the platform, creating opportunities to build increasingly valuable products on top of it.
Junior starts with expert-call transcription on a per-project basis, grows into a firm-wide research repository, and ends up generating deliverables from that compounding asset
Abridge follows the same staircase in healthcare: per-visit conversation capture, then coding and revenue-cycle documentation, then personalized clinical decision support. First-party data moats make each of these an N-of-1 company
Rilla records field sales visits, turning every in-home conversation into searchable data that coaches reps
Valence captures individual and team feedback conversations across VPs, managers and ICs, creating a channel for documentation that used to live in people's heads
Model 3: Marketplace to differentiate software
The cost of customer acquisition is increasing in relative to the cost of building software, which increasingly favor marketplaces that act the central and transactional layer for specific products and verticals. That structural advantage as the source of demand creates customer gravity is far harder to displace than any feature.
The playbook is to capture demand, take a cut of the transaction, then move upstream into the supplier's workflow. Non-customers become marketplace participants, then participants who also buy ads, then product customers, then full-suite customers who run their business on the platform, and so on….
The examples are familiar:
DoorDash extends from delivery orders into digital storefronts, point-of-sale, reservations and customer support for restaurants.
Airbnb layers listing management, smart pricing, co-host tools and host protection insurance on top of bookings.
Instacart sells retailers storefront solutions, a digital checkout assistant, shelf labels and procurement tools.
Faire adds retail purchasing management, shipping and invoicing for independent retailers.
Dutchie pairs its cannabis marketplace with compliance-native point-of-sale, digital storefronts, kiosks and analytics.
Winner-take-most dynamics
The biggest marketplaces tend to get bigger. More liquidity means transactions clear faster and at better prices, which pulls more volume onto the platform.
Model 4: Software to differentiate service (full-stack AI)
The fundamental issue facing AI-native software is a growing misalignment between the immense value it creates and the limited value it captures, which is brewing a new category of businesses that are built to sell outcomes, not software.
Instead of selling research tools to law firms, founders are building the AI-native law firms. The ceiling for these companies are no longer the software budget for a particular industry, they are designed to compete with incubments for the service spend category, which is far more than the software budget.
The beauty of traditional service firms is that it requires more headcount for more revenue, so margin stays flat. An AI-native service firm lets the product absorb more of the work over time, so margin compounds with scale.
The natural starting point is where labor is the biggest cost and the work repeats: claims, bookkeeping, diligence, back office, property management. The more labor inside a workflow, the more room to price around the outcome.
BPO verticals as the prime targets for disruption
Two shapes emerge, sorted by frequency and task standardization.
BPO (high frequency, highly standardized): This is work already outsourced under defined procedures and evaluation criteria. Most
There is a visible, contracted budget line to substitute
There are existing benchmarks and SLA attainment
These services are already outside the building
The best BPO targets share a profile: high existing spend, high volume, high repetition, mature SOPs, digital inputs (email, PDFs, portals), measurable accuracy, dissatisfaction with the current provider, manageable switching complexity and adjacent work to expand into
The single most important variable is verifiability. When the outcome is highly verifiable, you can create clean evals, compare actual vs. expected output, identify failure modes, and potentially use reinforcement learning or fine-tuning to improve performance systematically.
The time between doing the work and knowing it was right is the grading interval. It determines your feedback loop. Models became superhuman at code first because code carries its own verifier. Every compile is a grade and every unit test is the referee. Where verification is free, capability compounds. Everywhere else, it stalls
Buy and transform
AI roll-ups have become a prominent thesis for capital allocators and entrepreneurs. The boldest version doesn't win customers one at a time. It buys the service business outright, transforms it with platform technology and talent, fuels it with bolt-on acquisitions to expand market share and realize economies of scale
The construct of these companies revolve around three major functional roles:
Engineering builds model-agnostic agentic systems
Transformation team embeds them in live workflows and captures usage traces
Corporate development underwrites the margin gain and recycles the cash into the next acquisition
Two deals show the scale:
Metropolis started in 2017 as a computer-vision startup for parking operators. Long-dated incumbent contracts slowed direct sales, so it bought Premier Parking in May 2022 (about 2,000 employees, 600-plus locations), then raised $1.5 billion to take SP Plus private for $1.8 billion in October 2023, adding 3,300-plus locations and about 2 million spaces. Founder Alex Israel's point: a growth buyout succeeds on revenue synergies from deploying technology, not just cost cuts.
Long Lake, founded in 2023, built Nexus, a horizontal AI platform for automating workflows across acquired businesses. It bought dozens of service firms from 2023 to 2026, then closed a $6.3 billion acquisition of Amex GBT. Founder Alexander Taubman's framing: when teams are 30 to 40% more efficient, the business can grow with high incremental margins, which changes the whole mindset toward growth.
This model blurs venture and buyouts, which are usually two distinct asset classes with very different risk profiles. Each stage needs a different type of capital, from experimental venture risk to something closer to plain buyout capital, and it is a type of business that is incredibly difficult to underwrite, and so, some venture capital funds have decided to incubate their rollups than underwriting one (e.g., Thrive Holdings)
Engineering is no longer the moat in the application layer
In an age where intelligence is increasingly commoditized, the cost of writing software continues to fall while the cost of services and distribution remains largely unchanged. As a result, the relative value of everything around the software layer only increases.
Value is increasingly shifting away from pure SaaS and toward:
Services that turn software deployment into realized customer value.
Data that compounds with every interaction and becomes increasingly difficult to replicate.
Marketplaces that own demand and expand into the supplier's workflow.
Outcomes delivered by service businesses that use software to make the underlying delivery model structurally cheaper.
The question is no longer simply who can build the best software. It is who can own the layer around it that becomes harder to commoditize.