AI-Native Service Firms: The Startups That 'Do the Work,' Not Assist

For three years the default AI startup pitch was a copilot: software that sits beside a professional and makes them faster. The bookkeeper still does the books, just with an assistant. The claims adjuster still adjudicates, just with suggestions. That model is now being explicitly retired. Y Combinator's Summer 2026 Request for Startups names "AI-Native Service Companies" as a category it wants to fund — companies that do not assist a professional but replace the service outright, delivering the finished outcome with AI agents and a thin layer of human review. The bet is simple and aggressive: do not sell a tool to the accountant. Become the accountant.

"Does the Work" Is a Different Business Than "Assists You"

The distinction sounds semantic. It is not. A tool that assists a professional is sold into that professional's existing budget for tools. Its ceiling is whatever line item the firm allocates to software — typically a few percent of the cost of the humans it makes faster. A company that does the work, by contrast, is sold against the cost of the humans themselves. The buyer is no longer choosing between two pieces of software; they are choosing between hiring a firm of people and hiring a firm of agents. That is a vastly larger budget, and it is the budget AI-native service firms are built to capture.

Consider bookkeeping. The assist model sells a smart ledger to a bookkeeper who reconciles 200 accounts a month. Best case, the tool lets them reconcile 280, and the firm pays a software fee for the lift. The do-the-work model takes the 200 accounts directly: ingests the bank feeds, categorizes the transactions, flags the ambiguous ones for a human reviewer, closes the books, and delivers a finished monthly statement. The customer never hires a bookkeeper at all. They pay per closed month. The revenue that used to flow to a staffing firm now flows to a software company — and it flows at software margins, because the marginal cost of an agent processing one more account is a few cents of inference, not an hour of human labor.

This is why the category is being named now rather than two years ago. For the assist model to become the do-the-work model, the underlying agents had to cross a reliability threshold on bounded, repetitive tasks. They have. Not on open-ended reasoning, and not without supervision — but on the specific, rule-governed, high-volume work that fills the back office of every professional services firm, modern vertical agents now complete the task end to end often enough that a thin human layer can catch the rest.

The Economics: Services TAM at Software Margins

The reason investors are leaning in is arithmetic. Software is a roughly $700 billion global market. Professional and business services — accounting, legal, consulting, insurance brokerage, staffing, back-office outsourcing — are measured in the trillions. For the entire history of SaaS, software companies have been fenced out of that larger pool because software sells tools, and services firms sell outcomes delivered by people. AI-native service firms are the first credible attempt to cross the fence: to capture a slice of the services market while retaining the gross-margin profile of a software business.

The margin story is the crux, and it is more nuanced than pure software. A classic SaaS business runs 75–85% gross margins because the product is bits and the cost of serving one more customer is near zero. An AI-native service firm does not get there immediately. It carries two recurring costs a SaaS company does not: inference (tokens are a real, variable cost of goods sold) and the human-in-the-loop reviewers who handle exceptions. The thesis is that both costs fall over time — inference per task drops as models get cheaper, and the share of work needing human review shrinks as the agents and their feedback loops improve — while the revenue stays anchored to the value of the outcome, not the cost of producing it.

Outcome pricing is what makes the model legible to buyers. A customer cannot easily evaluate "$40,000 a year for an AI bookkeeping platform." They can immediately evaluate "$0.40 per reconciled transaction" or "$X per closed month" against what they pay a bookkeeping firm today. Pricing on the outcome aligns the vendor's incentives with the customer's — the firm only gets paid when the work is actually done — and it sidesteps the seat-based logic that AI is busy dismantling. We unpacked that shift in our analysis of the collapse of per-seat SaaS pricing: when an agent does the work, charging per human seat stops making sense, and the unit of value migrates from the person to the outcome.

Where It Lands First

Not all professional services are equally vulnerable to being rebuilt as an agent. The first wave concentrates in work that shares three properties: it is high-volume (the same task repeats thousands of times), it is rule-bound (there is a correct answer governed by policy, regulation, or accounting standards), and it is document-heavy (the inputs and outputs are text, forms, PDFs, and structured records that an agent can read and produce). Where all three converge, the agent has a clear target, the human review burden is manageable, and the outcome is verifiable.

Insurance is the canonical example. A claim is a document. The policy that governs it is a document. The adjudication is a rule-application problem with a verifiable answer: is this claim covered, for how much, and what is missing. The work is enormous in volume and intensely repetitive, and the brokerage and adjustment layers have historically been staffed by armies of people doing exactly the kind of structured reading and writing that vertical agents now handle. An AI-native claims firm does not sell the insurer a claims tool. It processes the claims and bills per claim resolved. The insurer sees a lower cost per claim and a faster cycle time; the firm sees services revenue at a fraction of the human cost.

The same shape holds for compliance, where the work is reading regulation, mapping it to a company's activity, and producing filings; for bookkeeping, where the work is categorizing and reconciling; and for back-office support, where the work is executing well-defined operational processes against a system of record. In each, the professional services firm's actual product was always an outcome, and the humans were the means of production. Swap the means and the outcome can be delivered at a different cost.

The a16z Rebuild Thesis

Andreessen Horowitz frames the broader version of this as a rebuild-from-scratch thesis. The argument is that you cannot bolt agents onto a legacy professional services firm and get an AI-native one. The legacy firm's processes, pricing, hierarchy, and incentives are all organized around billable human hours. Its business model rewards spending more hours, not fewer. Asking such a firm to deploy agents that eliminate the hours is asking it to cannibalize its own revenue, and incumbents almost never do that willingly. The opportunity therefore belongs to startups that build the firm around the agent from day one — outcome pricing, agent-first workflows, a human layer sized for exceptions rather than production.

This rebuild logic is why the category is a startup opportunity rather than an incumbent upgrade. The Big Four accounting firms will deploy AI internally and become more efficient, but their business model — leverage, the pyramid of junior staff billing hours to clients — is structurally opposed to charging per outcome at a fraction of the headcount. A startup carries none of that baggage. It can price on the claim, the return, or the resolution from its first customer, because it never built a P&L that depends on the hours.

The Build Pattern

Underneath every credible AI-native service firm is the same architecture, regardless of vertical. It is not a single agent answering questions. It is a pipeline that takes work in at one end and produces a finished, accountable outcome at the other, with humans positioned precisely where the agents are weakest. Four components recur.

The human-in-the-loop layer deserves emphasis because it is where naive builders go wrong in both directions. Remove it entirely and the firm ships errors into regulated, high-liability work — a misclassified tax position or a wrongly denied claim is not a bad chatbot response, it is a legal and financial event. Staff it too heavily and the margin advantage evaporates; you have rebuilt a staffing firm with extra inference costs. The discipline is to instrument the agent to know what it does not know, route only genuine exceptions to humans, and feed every human correction back into the system so the exception rate falls over time.

This is the same architectural reasoning that separates a real production agent from a demo. The difference between a chatbot that answers questions and an agent that completes accountable work is largely a question of architecture — durable state, verification, and human checkpoints — rather than model quality. We laid out that divide in our piece on chatbots versus agents and the architecture of ROI, and it applies directly here: the service firm lives or dies on the system around the model, not the model alone.

The Risks: Liability, Regulation, and the Last Mile

The thesis is strong, but the failure modes are specific and serious. The first is liability. When you do the work, you own the outcome. A SaaS vendor whose tool surfaced a wrong suggestion can point to the human who accepted it. An AI-native service firm that filed the return, adjudicated the claim, or closed the books has no such buffer — it is the party responsible. That raises the bar on quality from "useful most of the time" to "accountable every time," and it changes the economics: errors are not just churn risk, they are direct financial and legal exposure.

The second is regulation. Many of the richest target verticals — tax, insurance, financial compliance — are regulated precisely because the work carries public consequences when done badly. Licensing requirements, professional liability rules, data residency, and audit obligations do not disappear because an agent is doing the work. In several domains a licensed human must sign off on the outcome by law, which structurally caps how thin the human layer can become and shapes where the moat actually lives. The firms that win here treat regulation as a design constraint from day one rather than a problem to be discovered after launch.

The third risk is the last mile, and it is the one that quietly kills the unit economics of underbuilt firms. The agent handles the first 85–95% of cases beautifully and cheaply. The remaining slice — the ambiguous, the novel, the genuinely hard — consumes a disproportionate share of human time and carries most of the liability. A firm that prices for an easy average and then drowns in hard exceptions has a margin problem masquerading as a product. The long-term winners are the ones whose feedback loops genuinely shrink that last-mile slice over time, converting today's expensive human exceptions into tomorrow's automated cases.

What This Means for Builders and Buyers

For founders, the implication is to stop pitching copilots into budgets that are too small and start pitching outcomes against budgets that are large. Pick a vertical where the work is high-volume, rule-bound, and document-heavy; build the agent, the review layer, the system of record, and the outcome-pricing model as one integrated firm; and treat liability and regulation as first-class design inputs rather than afterthoughts. The hardest and most defensible work is the last mile — the instrumentation that tells the agent when to defer, and the feedback loops that make the deferrals rarer.

For buyers, the implication is that "we use AI" and "we do your work with AI" are about to be very different propositions in the market, and the second is the one worth paying for. The question to ask a vendor is no longer "does your tool make my team faster." It is "will you take this entire function off my plate, deliver the outcome, and price it per result." When the answer is credibly yes, the buyer has stopped purchasing software and started purchasing a service that happens to be delivered by agents — which is exactly the shift YC and a16z are betting the next wave of companies will be built on.

Conclusion: The Firm Is the Product

The AI-native service firm is not a SaaS company with an AI feature and it is not a staffing firm with a chatbot. It is a new kind of company in which the firm itself — the agent, the review layer, the system of record, and the outcome-based contract — is the product. Its addressable market is the multi-trillion-dollar pool of professional services that software could never reach, and its cost structure approaches that of software. That combination is rare enough to explain why Y Combinator named it as a category it wants to fund and why a16z frames it as a rebuild-from-scratch opportunity rather than an incumbent upgrade.

The execution risk is real — liability, regulation, and the stubborn last mile will sort the durable firms from the fragile ones. But the strategic direction is no longer in doubt. The most valuable AI companies of the next decade will not be the ones that helped professionals do their work. They will be the ones that did it.