Four VC firms agree on AI-native services. That's the easy part.

The reframe underneath all four theses is the same: stop selling the tool, start selling the outcome.

Sequoia partner Julien Bek put the sharpest version of it in "Services: The New Software" (March 2026): enterprises already spend roughly six dollars on services for every dollar they spend on software. That's a $6T+ global professional services market sitting next to a much smaller software market.

"The next trillion-dollar company will be a software company masquerading as a services firm. Sell the tool, and you're racing the model. Sell the work, and every model upgrade makes you faster, cheaper, and harder to compete with." — Julien Bek, Sequoia Capital · "Services: The New Software," March 2026

a16z frames the same shift as a pricing problem. Per-seat stops being the atomic unit once AI can do the task itself. Their example: a support team paying $115 per agent per month for Zendesk is buying seats; once AI can resolve the ticket, the natural unit becomes the resolved ticket, not the seat. Model a market of 10,000 customers at $1,000/month, and that reframe alone can turn a $120M opportunity into a $1.2B one, because the buyer is now paying for the outcome, not the license. It's also why incumbents struggle to make the jump themselves. Cannibalizing your own seat revenue is a much harder call than starting from zero — the same reason Blockbuster couldn't out-Netflix Netflix.

Bessemer's version leans on the P&L line, not the IT budget: vertical AI reaches into the labor line of the income statement, which dwarfs the software line. Their "services-as-software" leaders are hitting $100M+ ARR in under five years while expanding margin, not by selling a tool into a services workflow, but by owning the workflow outright.

YC made the most direct break with its own orthodoxy. For two decades, "services business" was close to disqualifying at YC: linear, people-heavy, low multiples. Their Summer 2026 Request for Startups reversed that explicitly under "AI-Native Service Companies," businesses that "don't sell software — they sell the service."

"Instead of selling software to customers to help them do the work, you can charge way more by using the software yourself and selling them the finished product — at 100x the price." — Aaron Epstein, YC Partner · Summer 2026 Request for Startups

Their companion video turns that into a market filter: pick verticals with low judgment at the task level, high domain intelligence, and regulatory complexity that keeps competitors out. Then treat the operation itself as the product. Throughput and cycle time are the real metrics.

Variance in output quality kills these companies faster than anything else, because trust is the entire product.

The clearest pushback isn't coming from AI doubters. It's coming from people building AI-native companies themselves, which is why it's worth taking at face value.

Nikola Lazarov, CEO of Eilla AI, an AI-native M&A advisory, put it plainly: on LinkedIn, the thesis is settled. AI-native services are the next big thing. In the rooms where VCs actually deploy capital, it's still a live argument. His example: Harvey, the legal AI company, raised at an $11B valuation in March 2026 on roughly $190–300M of ARR, somewhere between 35x and 55x, while earlier-stage challengers like Crosby and Lawhive are fighting just as hard to build the underlying firm. Same funds, both sides of the same bet.

The Masquerade Test: are you actually a software company in disguise, or a services firm wearing an AI costume? — Nikola Lazarov, CEO of Eilla AI (after Sequoia's framing)

Equal Ventures' Rick Zullo makes the harder economic argument. Falling costs only expand a market when demand is elastic, the classic Jevons paradox. But in categories where demand is fixed — his example is insurance claims administration — falling prices don't grow the pie. They just start a race to the bottom. He quotes an insurance executive shopping 34 different claims administrators: "they are all getting cheaper as we bid them against each other — it's beautiful." Beautiful for the buyer, fatal for the vendor.

There's a third, quieter warning worth naming: revenue quality. Bill Gurley's critique of "circular" AI deals — a large investor funds a startup that then spends the money buying services back from the investor — transfers directly to AI services. ARR that isn't paid for by an independent buyer's real budget isn't proof of a market. It's proof of a subsidy.

None of this means the thesis is wrong. It means the thesis is a necessary condition, not a sufficient one.

Put the two camps side by side and the reconciliation isn't complicated. It's just more specific than "sell outcomes, not seats."

Pricing has to reflect value delivered, not cost saved. A vendor swap that just quotes a cheaper number than the incumbent inherits the incumbent's commodity dynamics. The companies that escape it price on the outcome's value to the buyer: the deal closed, the case won, the brand cited in the AI answer.

This is also where Zullo's Jevons argument needs to be answered carefully. He's right for fixed-demand markets. But the marketing category is experiencing demand elasticity driven by a structural shift in the buyer journey:

• 89% of B2B buyers consult AI before making a purchase decision ^[4]
• 94% of AI answers cite third-party sources, not brand websites ^[1]
• $6B AEO/GEO managed-services market by 2030, up from $1.3B today ^[2] Companies invisible to AI in 2026 face a harder problem than companies invisible to traditional search, because AI answers compound: a source cited today trains the answer tomorrow. That's elastic demand growth, not a fixed pie being redistributed. The investor data makes the elastic-demand argument concrete. $261M+ in disclosed VC funding has entered the GEO/AEO category already. Profound reached a $1B valuation on $6.8M ARR, a 147× multiple, because early capital recognized the category before the market normalized. The broader B2B AI-visibility segment is $9B today and growing to $22B by 2030. ^[6] That's not a fixed pie compressing. It's a new market being created by a structural shift in how buyers find vendors. The margin has to be real, not borrowed. If the business still runs on a large, linear headcount behind the AI layer, it's a services firm with a chatbot bolted on: exactly the "costume" Lazarov is testing for. The test is simple: if revenue per operator grows as client count grows, the leverage is real. If it's flat, there are people behind the curtain, and the model doesn't hold. And the moat has to survive a well-funded copycat showing up on day one. Three candidates, but they're not equally durable. Workflow lock-in is table stakes, not a differentiator. Every managed service has it in year two. The real race is between the other two: domain expertise encoded in edge cases and a data flywheel that compounds with every delivery. A general-purpose model can produce a passable article. It cannot simultaneously calibrate for a specific client's editorial policy, brand voice, regulatory context, competitive sensitivities, and AI citation strategy, reliably, at volume, with a proof loop before every output. That calibration layer is the product. The underlying model is the machine tool. The distinction is the same reason Datadog survived AWS CloudWatch, and Cursor survived GPT-4.

"The model can attempt this" is a different claim from "the model can reliably produce this for your client, in production, with a proof loop before every output." The gap between those two claims is where the business lives.

The data flywheel is subtler but more durable: every client interaction, which prompt variant outperformed, which source type gets cited for this query, what edge cases look like for this editorial policy, is training signal for the operator model. A purpose-built workspace starts with the compounded decisions of every client that came before. Invisible in year one. Very difficult to replicate in year three.

"Claude Code doesn't compete with us — it's our factory floor. Frontier agents are generically smart but institutionally ignorant: they can't sign an SLA, can't see which sources ChatGPT actually cites in your category, and carry zero liability when they fabricate a citation. Our product is the layer models can't learn." — Humanswith.ai internal analysis

We're not going to claim we've solved the equation above. Nobody gets to claim that yet; the category is eighteen months old. What we can show is one real data point on the operating leverage side, because we ran the before-and-after ourselves.

That's a 10x increase in output on half the budget. Not a projection, a comparison of two periods of the same business.

The more interesting number is what happens on the client side. Visitors who arrive via an AI citation convert at approximately 5× the rate of organic search visitors, because the AI has already made the trust recommendation on their behalf. ^[3] That's not a throughput argument. It's an argument about what the production output is actually worth.

Early client results are consistent with the operating model:

• Humanswith.ai · Dogfood — 2 → 1,000+: AI citations across 9 engines in 12 weeks. 15.4% AI mention share, 5–10× ahead of every competitor in category.
• Birdview · B2B SaaS (PSA) — 23× lift: ChatGPT mention rate 0.9% → 21.5%. Now cited alongside Monday.com, Wrike, Asana. "ChatGPT recommends us by name." — Head of Marketing
• CodHob · Kenya fintech — #3 of 1,216: 0 → 35 AI citations. From zero presence to #3 of 1,216 domains in niche in 15 days. 22% share of category queries. The bridge from Stage 2 to Stage 3 is one engineering milestone: collapsing onboarding from weeks to hours. When that happens, one operator serves many clients simultaneously, and the business stops looking like a services firm. What we do know: a services business still needs a human for the part that isn't production. The relationship, the judgment calls a client won't make with a chatbot, the trust that keeps them from switching to whoever's cheapest this quarter. That human layer is real. But it also means the production side, everything up to it, is still a long way from its limit. The next 10× is pointing agents at that 80%.

We already proved the first 10×: one operator running agents does what six people used to do. The question worth asking is what the next 10× looks like, and where the category goes once autonomous agents replace the production bottleneck entirely. This is what the evolution of marketing services actually means: not an improvement to the existing model, but a replacement of it, stage by stage.

The margin progression here isn't internal projection. It reflects where managed services businesses land at each stage of automation. Traditional agencies run at 20–30% gross margin because most of the cost is people. Platform-and-operator delivery pushes above 45% because the fixed cost shifts to the platform, not the team. The autonomous model, where agents handle production continuously, is where comparable managed services categories show margins in the 60–85% range. That's the structural argument for why this category attracts capital at software multiples: services revenue approaching software economics.

"Some of the biggest companies of the next decade won't be software businesses. They'll be services companies rebuilt from scratch with AI doing most of the work." — Charlie Warren, YC Visiting Partner · YC Startup School, 2026

Stage 1 through Stage 2 is already in the data. Stage 3 is where the operating model transforms: when onboarding collapses and agents run production autonomously, the business stops being constrained by operator time. That's the structural unlock, and it's what separates a company that will compound from one that will plateau.

For the broader category, the underlying bet is this: AI search is not a feature added to existing search. It's a change in the buyer journey that doesn't reverse. The -30% organic search traffic drop already visible in our client data is not a temporary anomaly; Gartner projects it reaches -50% by 2028. ^[5] Citation patterns compound: a brand cited consistently today trains the next model generation to cite it again. The companies building the execution layer, not just monitoring dashboards but the actual content and distribution infrastructure that earns citations, are building the asset that captures that compound effect.

It's loud out there, the thesis is crowded, and the Masquerade Test is real. But the underlying shift isn't a thesis anymore. It's in the buyer data.

Four VC funds writing the same thesis is signal, not noise. It means the structural shift is real enough that sophisticated pattern-matchers in four different rooms landed on the same conclusion independently. That's the easy part to agree on.

The harder part, which the skeptics are correctly identifying, is that a structural shift creates a category, and categories create a lot of companies, and most of them don't survive their own success. Zullo's fixed-demand trap holds up. So does Lazarov's Masquerade Test. So does Gurley's revenue quality test. None of these are arguments against the category; they're selection filters for who will be left standing in it.

The filter — three conditions that separate real from costume:

1. Outcome-based pricing — not undercutting the incumbent on the same old service, but pricing on the value of the outcome delivered (the deal closed, the brand cited, the case won).
2. Real operating leverage — gross margin that climbs as client count grows, not borrowed from a model that still has six people behind it.
3. A moat a copycat can't buy in a year — domain expertise encoded in edge cases, or a data flywheel that started compounding before the clone showed up. Companies with all three will be the ones the next wave of investment papers is written about. Companies with only the first will be the cautionary examples that sharpen the Masquerade Test for the round after that. From inside the gap between the two camps: it's both things at once. The shift is real, and the timing is tighter than it looks from the outside. The bar for surviving it is higher than the thesis makes it sound. And the companies that get built here, the ones that prove operating leverage, then moat, then autonomy, in that order, will look, ten years from now, like the obvious result of an obvious shift. Which is exactly how the obvious results always look, in retrospect.

What does "AI-native services" mean? A company that doesn't sell software to a services workflow. It owns the workflow and sells the finished outcome (the deal closed, the case won, the article that gets cited), pricing on that outcome rather than on seats or tools.

Why do YC, a16z, Bessemer and Sequoia all back this thesis? Each fund reaches the same reframe from a different angle: Sequoia points to the $6T+ services market versus a smaller software market; a16z reframes per-seat pricing around resolved outcomes; Bessemer points to the labor line of the P&L dwarfing the software line; YC reversed two decades of treating "services business" as disqualifying in its Summer 2026 Request for Startups.

What's the strongest argument against the thesis? Equal Ventures' Rick Zullo argues that in fixed-demand markets, falling AI costs just start a race to the bottom rather than growing the market (the Jevons paradox). Nikola Lazarov's "Masquerade Test" asks whether a company is a real services firm or a software company wearing a costume. Margin has to be earned by real operating leverage, not borrowed by keeping a large team behind the AI layer.

What does Humanswith.ai's own data show? Running the same marketing-agency business through an agentic workspace instead of a 6-person team produced 120 articles/month versus 10–12 before, on $3,000/month versus $6,000: a 10x output increase at half the budget, plus AI-citation results such as a 23× lift in ChatGPT mention rate for one B2B SaaS client.

[1] Humanswith.ai Hermes platform data, cross-client monitoring, July 2026. Third-party citation bias: 94% of AI-generated responses cite sources outside the queried brand's own website.

[2] AEO/GEO managed-services market: consensus of 4 independent market studies + Gartner, 2026. Range: $1.3B (2026) → $6B (2030), 34–50% CAGR.

[3] Internal client attribution data, Humanswith.ai, Q2 2026. AI-cited visit conversion rate vs. organic search.

[4] Forrester B2B Buyer Study, 2026. "89% of B2B buyers consult AI assistants before making a purchase decision."

[5] Gartner Digital Marketing Forecast, 2026. "Organic search traffic projected to decline 50% by 2028 due to AI search displacement."

[6] Humanswith.ai Financial Model v6, market tab. SAM = B2B AI-visibility managed services, bottom-up: ~242K accounts at $30–36K ACV. Profound valuation/ARR: reported Profound financing (Series C, 2026). GEO/AEO category disclosed funding: Profound ($155M+), Otterly, Peec AI, Daydream, AirOps combined = $261M+ disclosed as of July 2026.