Vertical AI Isn’t Software Anymore
A field note on the rise of AI-native companies
Vertical AI isn’t software anymore. It’s an agentic company that completes workflows end to end. The bet isn’t to sell tools to a vertical. It’s to do the work the vertical was paying humans to do. That sentence is doing a lot of work, so let me unpack it.
The frame that’s becoming obvious in 2026 is this: for every dollar a company spends on software, six dollars go to services. The accountant gets $120,000 a year. QuickBooks gets $10,000. The law firm gets $200,000. Westlaw gets $30,000. The Big Four ESG advisory engagement is $200,000. The disclosure software underneath it is $50,000. The labor budget is six times the software budget, and software has been politely sitting beside it for thirty years pretending that’s just the way things are.
The thesis underneath the new wave of AI-native companies is that the labor budget is the actual prize. Sequoia partner Julien Bek’s Services: The New Software (March 2026) captures the framing cleanly: the next trillion-dollar company won’t sell a tool. It will sell the work that the tool was meant to help do. Software masquerading as services. The next legendary accounting company doesn’t sell better QuickBooks. It closes the books.
Bessemer’s vertical AI playbook (January 2026) makes the same point in different language: vertical AI competes for labor budgets, not IT budgets. Business and professional services are 13% of US GDP, roughly 10x the size of the software market, and that’s the pool AI-native companies are now able to address. Unlike vertical SaaS, which captured a fraction of Fortune 500 IT spend, vertical AI taps directly into the labor line of the P&L.
This is a structural reframe, not a feature improvement.
Copilot products help professionals do their existing work faster. Harvey helps lawyers draft. Rogo helps bankers analyze. Watershed helps sustainability consultants prepare reports. They sell to the professional, the professional uses the tool, the professional signs off, the professional takes liability. Pricing is per seat. Margins are software margins capped at IT-budget rates.
Autopilot products skip the professional entirely. Crosby doesn’t sell to law firms. It sells the drafted NDA to the company that needed one. WithCoverage doesn’t sell to insurance brokers. It sells the policy to the CFO. Rillet, which Sequoiabacked with $25M Series A in May 2025 and which then raised a $70M Series B from Andreessen Horowitz and ICONIQ ten weeks later, doesn’t sell ERP software to controllers. It sells your books closed to the CFO. The mission, in Sequoia’s words, is a zero-day close that lays the foundation for a one-person finance function. Pricing is per outcome. Margins capture labor-budget rates. Liability sits with the AI company, which is also why the moats run deeper: once you’ve taken liability for a customer’s compliance for two years, switching to a competitor means starting from zero on trust.
Insurance is the cleanest example of the same pattern in a regulated industry. Lemonade, the publicly-listed AI-native carrier (NYSE: LMND), reported in 2025 annual filing that its AI agent Jim handled 96% of first notices of loss without human intervention, and 55% of claims were fully automated end-to-end as of December 2025. Pet insurance in-force premium grew 55% year-over-year, from $283M to $439M. The model isn’t “use AI to make insurance faster.” The model is: chatbots Maya and Jim are the insurance company, and the carrier infrastructure exists to support what they do. Then in May 2026, Corgi , a Y Combinator-backed AI-native carrier for business insurance, raised a $160M Series B at a $1.3B valuation, doubling its valuation in four months. Corgi’s pitch: rebuild the fragmented insurance stack (TPAs, MGAs, reinsurers, carriers) from scratch as an AI-native carrier, with AI handling underwriting, claims and policy operations directly.
What makes a company genuinely AI-native, rather than a traditional company that “added AI”? Five things, all downstream of one decision: the company is structured around the work, not around the tool.
First, the data model is built for agents, not humans. A traditional accounting system stores ledger entries the way an accountant would write them down. An AI-native ERP stores them the way an agent needs them to reason across entities, currencies, and consolidations. The schema itself is the product.
Second, the unit of value sold is the completed outcome, not the seat license. You don’t pay per accountant using the tool. You pay per closed month. You don’t pay per consultant drafting a CSRD report. You pay per CSRD report filed. The pricing model maps to what the customer actually wants. Bessemer’s AI pricing playbook (February 2026) tracks this shift explicitly: AI-native companies are abandoning seat-based pricing in favor of usage, output, and outcome-based models.
Third, workflows are agentic by default. The default state is “done.” The human exists as an exception handler, not as the operator. Compare this to a SaaS app where the default state is “you have to do it” and the software exists to help you do it slightly faster.
Fourth, the product gets better with usage in ways the customer feels. Every report drafted, every transaction reconciled, every appointment booked makes the next one tighter. The moat is the customer-specific learning, which is impossible to replicate by acquiring a competing product. Lemonade calls this their“Customer Cortex“ More customers means more data, which means better pricing and underwriting, which attracts more customers.
Fifth, integration depth equals workflow depth. The AI-native company sits inside the customer’s existing systems, pulling from their ERP, HRIS, utility billing, lab system, hospital information system, rather than being a destination they have to visit. The customer never “logs into” the AI-native company the way they log into Workday. The work just happens.
The other critical reframe is on which verticals are actually attackable. The lazy version of this thesis is: anywhere there’s a big labor budget, build the autopilot. The accurate version is more nuanced. The opportunity isn’t just labor budget. It’s labor budget intersected with task structurability, capture rate, and defensibility.
Task structurability means the work has to be the kind AI can actually do. Bek’s framework splits cognitive work into intelligence (rule-based, procedural: medical coding, NDA drafting, financial close, ESG disclosure mapping) and judgement (accumulated experience, ambiguous contexts, stakeholder management). The first is autopilot-able now. The second isn’t. Big labor budgets without task structurability produce pilots that never scale.
Capture rate is whether the labor budget actually flows to the AI vendor when the work shifts. Fragmented buyers, startups, mid-market firms, individual hospitals, let you capture a meaningful share. Consolidated buyers with monopsony power often keep the savings themselves. The fragmentation of the buyer side determines how much of the labor budget you actually receive.
Defensibility is the most overlooked. Where labor budgets are large but there’s no moat, no proprietary data, no workflow lock-in, no regulatory advantage, the labor savings get competed away to near-zero margin. Customer support is the canonical example: huge global labor pool, ten well-funded autopilot startups, margins compressing fast. The verticals that actually compound are the ones where the regulatory complexity making the work expensive also creates the moat for whoever masters it first. Insurance, healthcare, legal services, financial services, ESG reporting. These are big because compliance is hard, and the compliance hardness is also what keeps competitors out.
There’s a sobering counterpoint to the labor-budget thesis that doesn’t get emphasized enough. Linas Beliūnas argued in a sharp Substack response to Bek’s piece that machine rates are roughly 97% cheaper than human labor, meaning the labor budget does not transfer to AI vendors at par. The $120,000 accountant doesn’t move to your top line. You capture maybe $30,000-$40,000. The rest evaporates into customer surplus. The winners aren’t whoever extracts the biggest absolute dollar amount per customer. They’re whoever (a) achieves defensible scale fastest at the new lower price point, and (b) expands into adjacent labor budgets: tax, audit, FP&A, treasury from accounting; insurance pre-auth, lab delivery, pharmacy follow-up from front desk; supplier audit, transition planning from ESG. Wallet share expansion is the math that actually works.
The pricing question is where most founders get stuck. Three structures all show up, suited to different stages and markets.
Pure outcome pricing, per report, per appointment, per closed book, is theoretically the cleanest match for the autopilot thesis. It’s how Big Four advisory firms have always billed. It signals “we’re not software” to the buyer. But it’s premature for most v1 products because customers can’t predict their bill, your costs aren’t perfectly variable, and measurement disputes will eat your relationship time. Outcome pricing becomes viable once you have 50+ customers and stable unit economics.
Hybrid pricing, base platform fee plus per-doctor or per-employee or per-entity charges plus overage tiers, is what works in practice for early-stage AI-native services. It’s predictable enough for finance teams to budget, scales naturally with customer size, and protects you from the small-customer-eating-your-fixed-costs problem. This is what enterprise SaaS evolved to, and it’s what most AI-native companies will end up at by Series A.
Pure per-seat pricing is the trap. It caps you at software-budget economics and signals to the customer that you’re a tool, not a service. Any company pricing per seat is fighting for the IT budget, which means the next pricing cycle competes them down to commodity rates. Avoid.
The framing that works for the customer conversation isn’t pricing. It’s capacity unlock. You’re not telling them how much they save. You’re telling them how much more they can do. The hospital’s front desk handles 80% of patient demand today; the other 20%, after-hours inquiries, missed calls, never-made follow-ups, lapsed pre-auth conversations, is invisible lost revenue. The accountant takes 15 to 20 days to close the month, which delays everything downstream. The ESG consultant takes 16 weeks to produce a report, which is so slow the company can’t actually use the data for transition planning. AI-native services don’t just compress costs. They expose latent capacity that was always there but unreachable at human labor rates.
This is also why the AI-native company’s pricing isn’t capped by what the customer is currently spending. The customer is paying for what they get today. The AI-native company sells what they could be getting tomorrow. The ceiling is set by the value of the unlock, not by the cost of the labor displaced. A mid-market Indian hospital paying ₹1.6L/month on front desk staff might happily pay ₹85K to the AI vendor while also recovering ₹9-10L/month in previously-lost revenue from better inbound conversion and reduced no-shows. That math reads as net positive impact, not cost replacement, and it scales because every adjacent workflow added is another unlock.
The verticals where this is currently playing out in 2026 are clearer than they were even 18 months ago.
In regulated services, accounting (Rillet), tax advisory, legal transactional work (Crosby for NDAs, Lawhive for consumer legal), ESG and CSRD reporting, healthcare revenue cycle (Anterior), insurance underwriting and claims (Corgi, Lemonade, Pace, Strala), the autopilot pattern is taking hold fast. These work because regulation creates a structured task surface, provides authoritative taxonomies, and the EU’s CSRD ESRS framework alone has 1,144 mandatory and voluntary data points across 82 disclosures, published openly by EFRAG with full XBRL taxonomy, and screens out competitors who can’t navigate the compliance layer.
In operational labor pools, call centers, customer support, front-desk and patient communication, customs and export documentation, fund administration, RegTech for financial services, the disruption is just hitting. India’s domestic call center market alone is ~$33B annually, and India’s broader BPO industry sits at $38-54B with roughly 1.2-1.6M voice agents. AI-native voice handlers are pricing at a fraction of human rates with better coverage and multilingual capability; companies like LimeChat report handling up to 95% of customer queries without human assistance. The Indian healthcare front office sector, front desk plus admin clerks plus nurses pulled into phone duties, is similarly massive and similarly under-attacked. These are the easiest unlocks to demonstrate because the pain is visceral and daily.
In knowledge-work back office, financial close, FP&A, contract management, procurement, recruiting operations, immigration paperwork, grant writing, the autopilot version replaces multiple coordinated humans rather than a single role. Companies like Abridge (clinical documentation), EvenUp (personal injury law), Fieldguide (audit) and Legora (corporate legal) are showing what’s possible. These are harder to build because the workflows span systems, but the labor budgets per customer are huge and the moat from cross-customer data is deep.
What unifies all of these is the same pattern: the company doesn’t sell tooling to professionals. It does the work. It bills against labor. It compounds with data. It expands by adding adjacent labor budgets rather than adding features.
The unsexy verticals, waste management, traditional B2B services, local government, hospitality back office, construction operations, are arguably the largest opportunity, because labor budgets are enormous and software penetration is shallow. The investment thesis quietly emerging is that the next decade of category-defining companies will look more like Corgi (owning the regulated insurance stack with an AI-native back office) and less like a SaaS layer sitting on top of incumbents.
The companies winning here aren’t necessarily the ones with the best models. The foundation models are commoditizing, and access to GPT-5 or Claude Opus 4.7 isn’t a moat. The winning companies have a vertical-specific data model designed for agents not humans, workflow integration deep enough that customers can’t easily switch, a regulatory or trust position competitors can’t replicate by lifting code, a pricing model that captures labor budgets not software budgets, and an operating discipline that treats customer trust as the most valuable asset, because autopilot economics only work when the customer doesn’t have to audit the output every month.
The mental model worth carrying out of this moment is simple. Software was a layer. AI-native services are the work itself. The old question was “what tool do you use?” The new question is “who handles it for you?” Whoever owns the answer to that question in a given vertical owns the labor budget that vertical was spending on humans yesterday. That’s what an AI-native company actually is.
Sources & further reading
The core thesis
Julien Bek (Sequoia Capital), Services: The New Software, March 2026 — the foundational essay
Sequoia Capital, Partnering with Rillet: The Financial ERP for the AI Age, May 2025 — Bek’s worked example
Bessemer Venture Partners, Building Vertical AI: An early stage playbook for founders, January 2026 — the labor-budget vs IT-budget framing
Bessemer Venture Partners, The AI pricing and monetization playbook, February 2026 — pricing structures in practice
Bessemer Venture Partners, Roadmap: AI systems of action, December 2025 — systems-of-record to systems-of-action framework
Bessemer Venture Partners, Ten principles for building strong vertical AI businesses
Critique and counterpoints
Linas Beliūnas, Sequoia’s Services Thesis Will Mint Billionaires and Bankrupt Copycats, March 2026 — the machine-rates-evaporation critique
Medium / Data Science Collective, Services Are the New Software: Building Them Is the Hard Part — engineering-team perspective on the thesis
Press coverage of the thesis
Fortune, This Sequoia partner thinks AI-enabled services are the new software, April 2026
Company evidence
Corgi, Series B announcement, May 2026 — $160M at $1.3B valuation
TechCrunch, Insurance startup Corgi hits $1.3B valuation, May 2026
The Next Web, From $630m to $1.3bn in four months: Corgi hits unicorn status, May 2026
Lemonade, Inc., 2025 Annual Report (Form 10-K) — 96% AI-handled FNOL, 55% claims automated
Rillet, Series B announcement, August 2025 — $70M from a16z and ICONIQ
TechCrunch, Rillet raises $25M from Sequoia, May 2025
Regulatory frameworks referenced
EFRAG, ESRS Implementation Guidance and the IG 3: List of ESRS Data Points — the open taxonomy that makes ESG autopilot possible
European Securities and Markets Authority, ESRS XBRL Taxonomy — machine-readable disclosure schema
Great read