Why every services domain is being rebuilt, AI-first.

For four decades, venture capital has chased software — pure software, sold as a subscription to a CFO or a CIO. The vast services economy next door was left to operators: law firms, accounting practices, consultancies, recruitment agencies, clinical advisors. In 2024, the ground moved.

Services firms produce deliverables — a memo, a shortlist, an audit, a tax return, a diligence report — and their costs scaled linearly with headcount. Their margins were capped. Their clients complained about quality. And for decades, nothing changed. Until it did.

Andreessen Horowitz (a16z) has spent the last eighteen months articulating a single, load-bearing thesis: large language models, reliable tool use, and agent orchestration have finally crossed a threshold where the deliverable itself can be produced by software — with a human in the loop for judgment. Their framing of the opportunity is sometimes called service as software. Not AI-as-a-tool inside a law firm. Not a chatbot bolted onto an ATS. An entirely different architecture for a services business.

Y Combinator's founders have voted with their batches. Of the last four cohorts, a striking share have launched AI-native vertical services firms — legal discovery, medical coding, financial close, due-diligence memos, paralegal research, sales engineering, recruiting, benefits compliance. Individually small. Collectively, a wave. The common thread: the customer does not care how the deliverable is produced. They care that it is better, faster, and cheaper. Everything else is implementation detail.

01 / Why nowThree curves that made it economic

Three things changed in the same eighteen months. None of them was the AI narrative. They were all about unit economics.

Context windows grew from 4,000 tokens to 1,000,000 — enough to ingest a full company dataroom, a candidate's entire public record, or ten years of case law in a single call. What used to require document chunking, vector retrieval, and tedious assembly can now be done in one shot.

Tool use became reliable. Models can now orchestrate dozens of steps — retrieve a document, run a computation, call an API, cross-check against a rubric, hand off to a human, resume — without the plumbing breaking at step four. The "agent" is no longer a demo. It is a production workflow.

The cost per million tokens fell by roughly 80% between mid-2023 and early 2026. A workflow that used to require a senior associate now runs in eleven minutes for the price of a cup of coffee. The economics of a services firm — any services firm — invert.

Those three curves — more context, reliable agents, collapsing cost — are what made services-as-software possible. Not hype. Math.

02 / The legacy playbookWhy "AI inside the agency" doesn't work

Consider recruitment. A traditional boutique searches a network, writes job descriptions, filters CVs by keyword, conducts 30-minute screening calls, assembles a shortlist, and charges 20–30% of first-year compensation. Every step is a person's hour. Every person is a fixed cost. The industry's gross margin has barely moved in forty years.

Now compare an AI-native recruitment firm. The same work is decomposed into software steps — candidate sourcing, a 45-minute structured AI interview, a rubric-graded skill assessment, a practitioner QA, a Talent Analysis Graph shipped to the client. The cost of the interview is sub-dollar. The practitioner spends eight minutes on QA, not eight hours. The deliverable is better because it is more structured, more evidence-based, and more defensible than a partner's one-page summary.

Clients are not sentimental about their vendors. They are sentimental about their outcomes. — The defining line of the AI-native services thesis

Incumbents confuse the two moves. Robert Half and Michael Page each announced "AI partnerships" in 2024 and 2025. Both still charge partner rates. Both still deliver a one-page consultant summary. The AI is decoration, not architecture. A 10-to-15% productivity uplift on the same cost base does not create a structural advantage. It just delays the reckoning.

The pattern repeats in every vertical. Legal firms add "AI research assistants" for their associates while the work product remains a 40-hour memo. Consulting firms announce "AI-augmented engagements" while the pricing stays per-partner-per-day. The assumption underneath all of it — we will ship the same deliverable, just faster — is the expensive assumption. The new firms ship a different deliverable at a different price point, and the buyer does not go back.

03 / Two playbooksa16z's thesis and YC's wave

The a16z playbook tends toward category leaders — firms going after a $20B+ vertical (insurance, tax, accounting, recruiting, medical coding) with a vertically integrated stack, proprietary workflows, and a bet on winning category share. They raise from seed through growth, build product and operations in parallel, and target enterprise logos. The moat is compounding: a firm that has run 50,000 tax returns has a rubric, a dataset, and a QA flywheel that a new entrant cannot match.

The Y Combinator playbook is scrappier and more diverse. Pick a single painful workflow inside a services vertical. Automate it end-to-end. Price it as a deliverable, not a seat. Own the margin. A single founder can credibly compete with a 50-person boutique in a narrow slice of work. Scale is optional; margin is not.

The two approaches converge. Whether a firm is venture-funded or founder-funded, the output looks the same: a deliverable, shipped faster and better than the legacy incumbent, at a price that reflects software economics rather than consultancy economics.

04 / What winning looks likeThe pattern across domains

A pattern is emerging across every vertical that has crossed the threshold. The AI-native firm that wins is not the one with the best model or the slickest product. It is the one that does three things well.

It picks a specific vertical and goes deep. Not "recruitment" but "specialised finance roles in India." Not "legal" but "M&A transaction diligence for private-equity funds." Depth beats breadth because the prompts, rubrics, evidence standards, and QA protocols are domain-specific. A horizontal AI hiring tool and a finance-specific AI recruiter are not in the same business.

It keeps a practitioner in the loop, visibly. The deliverable is better because a domain expert has reviewed it — and the firm says so, loudly, on the artifact itself. This is what closes the trust gap with buyers who have been burned by pure-AI attempts. The practitioner is not a fallback. The practitioner is a product decision.

It sells the deliverable, not the technology. The website does not lead with "AI". It leads with the outcome the buyer cares about — a hire, a filing, a memo, a diagnosis. This is the lesson every successful AI-native firm is converging on: the buyer of services does not want to buy AI. They want the work done.

05 / The question for executivesRebuild, or be rebuilt against

If you are running a services firm, the question is not whether AI will reshape your domain. The question is whether your firm will be the one that rebuilds, or the one that gets rebuilt against. Five years is the window. Twenty-four months from now, in most verticals, there will be an AI-native entrant that ships the same deliverable at 30% of the cost with more defensible quality, and the migration will have already started.

If you are buying services — as a CFO, a General Counsel, a CHRO, a head of hiring — the question is whether your current vendor's cost structure allows them to compete, five years out, with that entrant. If the answer is no, continuing to buy the legacy deliverable is a choice, not an inevitability.

The services economy spent forty years resistant to software. It is not going to spend another forty. The firms that understand that — and rebuild their workflows architecturally, not cosmetically — will be the ones that still exist on the other side of this cycle. The rest will be case studies.