# Sequoia’s services-as-software thesis, in plain terms

URL: https://www.thedeepfeed.ai/posts/2026-04-30-sequoia-services-as-software-thesis/
Category: Business
Published: 2026-04-15
Updated: 2026-05-22
Author: the-deep-feed
Tags: sequoia, ai-startups, vertical-saas, services-as-software, autopilot
Kind: deep

> Sequoia says the next $1T company will be 'a software company masquerading as a services firm.' The sentence buries vertical SaaS as a category — without ever naming it. What that means and which companies fit.

## TL;DR

- Sequoia's claim is a multiples arbitrage. Services trade at **1-2x revenue**, software at **8-20x**. Automate a $6T services market at software margins and the gap is the prize.
- The framing reframes AI startups: not tools sold to professionals, but firms that sell the work itself. The product is the closed books, not the accounting software.
- Vertical SaaS is the implicit casualty. Selling a workflow to a professional caps you at the tool budget. Selling the outcome captures the labor budget, which is **roughly six times bigger**.
- Named companies fit the shape: **Harvey** moving from copilot to autopilot, Crosby and Lawhive in legal, EvenUp in claims, Mercor in recruitment, Anterior in medical coding, Edra in IT, Rillet in accounting.
- The unanswered questions are regulatory licensure, error liability, and what happens when the work itself costs **$0.03 to produce**. Sequoia has published the map. The territory is not yet surveyed.

On March 5, Sequoia partner Julien Bek published a 1,400-word essay called [Services: The New Software](https://sequoiacap.com/article/services-the-new-software/). The headline claim is one sentence. *"The next $1T company will be a software company masquerading as a services firm."*

Olivia Moore at a16z framed the shift bluntly:

> RIP, software as a service (1999-2025)
>
> Welcome to the world, services as software (2026 - ?)
>
> — [@omooretweets](https://x.com/omooretweets/status/2039150266889744653), Apr 1, 2026

![Multiples arbitrage — services 2x, SaaS 14x, services-as-software ??x](/post-images/sequoia-services-as-software-thesis/multiples-arbitrage-chart.jpg)

Read past the slogan. The sentence is doing two jobs. One is the forecast (where trillion-dollar value accrues in the AI era) and that's the part everyone quoted. The other is retiring a category Sequoia spent the last decade funding. Vertical SaaS, the entire "Toast for X" genre, is not in this future. Bek does not say so directly; he does not have to. If the next $1T company sells the work itself, by definition it does not sell a workflow tool to the people doing the work. Different businesses, different customers, different multiples.

The harder read, the one circulating among investors who have spent ten years underwriting vertical SaaS, is that the rubric just got an asterisk next to it. "Services-as-software" works as shorthand. The architectural break is that the buyer changes. You stop selling features to the practitioner and start selling outcomes to whoever signs the practitioner's invoice. On those grounds, it's the most consequential piece Sequoia has published since *Generative AI's Act Two*.

The piece spread fast in finance and VC circles:

> Services: The New Software
>
> Sequoia on the next $1T opportunity map.
>
> — [@TheShortBear](https://x.com/TheShortBear/status/2039039380015227278), Mar 31, 2026

## The claim

Bek's argument starts from a ratio. *"For every dollar spent on software, six are spent on services."* Global software is roughly $700B. Global professional services is north of $6T. The SaaS era was built on the smaller pool. The thesis is that the AI era captures the $6 instead.

The mechanism Bek calls the autopilot. *"A copilot sells the tool. An autopilot sells the work."* A copilot is software a professional uses to be more productive. An autopilot is software that replaces the engagement entirely. The professional is not the customer. The company that needed the professional is the customer. The product is the outcome.

The example Bek anchors on is unflashy and effective. A company spends about $10K a year on QuickBooks and about $120K on an accountant to close the books. *"The next legendary company will just close the books."*

> **$1T** projected market for services-as-software (Sequoia's framing)
>
> **$6T** global professional-services spend the thesis is aimed at
>
> **$700B** global software spend the SaaS era was built on
>
> **30 minutes** — Bret Taylor's reported time to write a software brief at Sierra, down from days

The QuickBooks-vs-accountant comparison is the whole essay in two numbers. $10K of software, $120K of labor. The next legendary company is not selling a better $10K SKU. It is billing the $120K and keeping most of it.

The rest of the essay is mechanical. List the markets that already pay for outcomes. Filter for high "intelligence ratio," meaning rule-following work where AI can already operate autonomously. Filter for fragmented incumbents and aging workforces. Land on a $1.4-1.7T addressable surface in the US alone. Name twenty companies. Publish the deal map and let founders self-sort.

The interesting question is what the essay implies that it does not say.

## The multiples math

Why is "selling the work" a $1T idea and "selling the workflow tool" not? The answer is in public markets, and it is uncomfortable for anyone who built a vertical SaaS company in the last decade.

Software businesses trade at 8-20x revenue. Snowflake is roughly 14x forward revenue. ServiceNow has spent most of the last five years at 14-18x. Even battered SaaS comps cluster at 6-10x. The reason is gross margin. Public software businesses run 70-85% gross margins, so revenue compounds into earnings with very little drag.

Professional services businesses trade at 1-2x revenue. Accenture, the category-leading consulting and outsourcing operator, trades around 2x revenue. The Big 4 firms, when divisions of them have transacted, have priced at 1.5-2.5x. BPO firms (Cognizant, Infosys, Genpact) sit at 2-3x. The reason is again gross margin. Services firms run 25-35% gross margins because their cost of revenue is people. Headcount scales with revenue almost linearly. Costs grow in lockstep with revenue, so there is no margin gain to price in.

| Comp class | Forward revenue multiple | Typical gross margin | Cost of revenue scales with |
| --- | --- | --- | --- |
| Public SaaS leaders (ServiceNow, Snowflake) | 14-18x | 70-85% | Compute, modestly |
| Battered / mature SaaS | 6-10x | 65-75% | Compute + support |
| BPO (Cognizant, Genpact, Infosys) | 2-3x | 30-40% | Headcount, linearly |
| Category-leading services (Accenture) | ~2x | 25-30% | Headcount, linearly |
| Big 4 audit / advisory | 1.5-2.5x | 25-35% | Headcount, linearly |
| **Services-as-software (theoretical)** | **10-18x?** | **60-80%?** | **Inference, sub-linearly** |

The arbitrage Sequoia is describing lives in the gap.

Take accounting. The US outsourced accounting and audit market is $50-80B. A traditional services firm capturing $1B of that revenue is worth $1-2B. A software business capturing the same $1B at ServiceNow's multiple is worth $14-18B. A single firm that captured $20-30B of the revenue base by automating most of the labor, and traded at software multiples because most of the cost line was GPU time rather than CPAs, would land somewhere in a $200-400B valuation range. Run the same exercise in insurance brokerage ($140-200B), recruitment ($200B+), and IT managed services ($100B+). One of them produces a $1T outcome.

That is the math behind the slogan. It is a margin-conversion story. A services business reframed as software. Priced at software multiples because the labor cost has been replaced.

Jason Shuman at Primary, who has been underwriting this shape at seed for years, put the multiples gap in one tweet:

> Coatue just put a number on what we've been seeing at seed for 3 years.
>
> Software = $0.2T market.
>
> Services-as-software = $5.5T.
> 25x.
>
> The shift is from selling tools (per-seat) to selling work (per-output).
>
> This is why the best vertical AI companies don't compete with
>
> — [@JasonrShuman](https://x.com/JasonrShuman/status/2037540280140714190), Mar 27, 2026

![Industry-by-industry TAM — where the trillion-dollar companies emerge](/post-images/sequoia-services-as-software-thesis/industry-tam-bars.jpg)

The bear case here is straightforward and worth holding next to the bull case. Linas Beliūnas, in [a March 13 piece](https://linas.substack.com/p/sequoiathesis), called this the "$0.03 problem." If the work can be done by inference for three cents, what is the sustainable margin once five startups in each vertical are competing on price? Software margins on the income statement assume software prices on the invoice. The autopilots have not proven they can hold those prices.

Two reminders worth keeping in mind for founders pattern-matching this thesis. First, the multiples arbitrage holds only if the invoice can be defended: inference at $0.03 per job sets a floor near $0.30, and the question is who owns the customer relationship that lets you charge $300 instead. Second, services pricing is not bound by COGS — it is bound by the labor budget being replaced. Once two AI competitors in a vertical race to "fair" pricing, the economics revert to BPO. The verticals where that race is slow are the ones worth picking.

![Two architectures — services firm vs services-as-software](/post-images/sequoia-services-as-software-thesis/services-vs-saas-architecture.jpg)

## What "services-as-software" means

The phrase is doing a lot of work. Strip it back.

Services-as-software is an architecture. It has three properties.

First, AI is the worker, not the tool. In a SaaS architecture, a human professional is the unit of production and software augments that human. Each seat sold is one human made faster. In a services-as-software architecture, the unit of production is an agent. Each unit of revenue is one outcome delivered. The headcount line on the income statement does not grow with revenue. The inference line does.

Second, the customer is different. SaaS customers are the practitioners. The end-user of Clio is a lawyer. The end-user of QuickBooks is a bookkeeper. Services-as-software customers are the people who hire the practitioners. The CFO who wants the books closed. The COO who wants the claims adjusted. The general counsel who wants the NDAs signed. These are budget owners on the labor line. They are not interested in features. They are interested in whether the work happens.

Third, pricing is per-outcome, not per-seat. Per filing. Per claim. Per hire. This breaks how SaaS firms have always billed. It also unlocks a TAM that seat-based pricing cannot reach: companies that have no professional in the role to begin with. An SMB without an in-house bookkeeper is not a Clio prospect. They are a TaxGPT prospect.

The shorthand most VCs are using ("AI-native services" or "vertical AI") obscures the architectural break. The break is that the product is the work, sold to the person who buys the work. Pricing, GTM, hiring, and gross margin profile all follow from that.

| Dimension | Vertical SaaS | Services-as-software |
| --- | --- | --- |
| Unit of production | A human seat | An agent run |
| Buyer | The practitioner (paralegal, bookkeeper, recruiter) | The buyer of the practitioner (GC, CFO, COO) |
| Pricing | Per seat / per month | Per outcome (per filing, per claim, per hire) |
| TAM ceiling | Tool budget per seat | Labor line on the customer's P&L |
| Cost of revenue | Compute + support staff | Inference + a thin licensed-human signoff layer |
| Gross margin | 65-80% | 60-80% (with regulatory overhead) |
| What model improvement does to you | Compresses your wrapper | Reduces your COGS |
| Sales motion | Demo to a user, expand seats | Pilot on outcomes, expand SLAs |

The category is not "AI for X." It is "buyers of X labor finally have a vendor that's not a temp agency." That is a different sales motion, a different sales-cycle length, and a different reference customer. A useful mental model: services-as-software companies should look like a Big 4 firm acquiring a software company, not a software company looking at a services workflow backwards. Software-first orgs typically fail at outcome SLAs because the org chart is wrong for them.

## The companies fitting the shape

Bek named twenty in the essay. Cluster them by vertical and the shape becomes legible.

**Legal.** [Harvey](https://www.harvey.ai/) is the canonical copilot-to-autopilot transition, and Bek is candid that the transition is hard because it cannibalizes the firm's own buyer. Harvey today sells to BigLaw associates. The autopilot version sells contract drafting and regulatory filings as outcomes to general counsel, which removes the BigLaw associate from the chain of trust. [Crosby](https://www.crosby.ai/) and [Lawhive](https://lawhive.co.uk/) are autopilot-native. They are not selling lawyers a better Word. They are filing the documents.

**Claims and personal injury.** [EvenUp](https://www.evenuplaw.com/) is the cleanest pure-play. The product is a finished demand letter, generated from medical records, sent to the carrier, priced per case. The customer is the plaintiff law firm. The spend it replaces is the paralegal hour.

**Recruitment.** [Mercor](https://mercor.com/) sells qualified candidates, often vetted entirely by AI screening. Juicebox does the sourcing layer, also outcome-priced. Neither is a recruiter copilot. They compete with the recruiter line item itself, except the cost of fulfillment is GPU time and the gross margin profile looks nothing like Robert Half's.

**Accounting.** [Rillet](https://www.rillet.com/) and Basis sell the closed books. The customer is the SMB CFO. The replaced spend is the fractional controller. The US has lost roughly 340,000 accountants over five years, and 75% of CPAs are nearing retirement. The structural shortage is what makes this market move faster than the others.

**Medical coding.** [Anterior](https://anterior.com/) is the example. Medical coding translates clinical notes into the ~70,000 ICD-10 codes that determine what a hospital gets paid. It is rule-bound, outsourced at scale to BPO firms in India and the Philippines, and priced by the chart. The autopilot does the chart. It is sold to the revenue-cycle team.

**IT managed services.** [Edra](https://sequoiacap.com/article/partnering-with-edra-context-for-agents-at-scale/) (announced as a Sequoia investment thirteen days after Bek's essay) and Serval sell "your IT runs." Datto and ConnectWise sold software to the MSP. Edra is selling the MSP's job.

**Procurement.** Magentic, AskLio, and Tacto attack what Bek calls "abandoned work." 80% of suppliers in a typical enterprise get zero negotiation attention. Contract leakage runs 2-5% of procurement spend. The wedge is found money. Coupa sells procurement workflow to the buyer team. Magentic sells the negotiation.

What every one of these companies has in common, and what separates them from a vertical SaaS pitch on the same surface, is the line item they replace on the customer's P&L. The line is not "software." The line is "labor."

| Company | Vertical | Replaced spend line | Pricing unit | Buyer |
| --- | --- | --- | --- | --- |
| Harvey (autopilot mode) | Legal | BigLaw associate hours | Per matter / per filing | General counsel |
| Crosby | Legal contracting | Outside counsel review | Per contract | GC, head of legal ops |
| Lawhive | Consumer legal (UK) | Solicitor fee | Per matter | End consumer |
| EvenUp | Personal injury | Paralegal hours | Per demand letter | Plaintiff firm partner |
| Mercor | Recruitment | Agency placement fee | Per qualified candidate | Head of talent |
| Anterior | Medical coding | Offshore BPO chart-coder | Per chart | Revenue cycle leader |
| Rillet / Basis | Accounting | Fractional controller | Per close cycle | SMB CFO / founder |
| Edra | IT managed services | MSP labor hours | Per ticket / per seat | IT director |
| Magentic | Procurement | Buyer time + leakage | Share of savings | CPO / CFO |

Selling the work itself means taking liability. The autopilots that win are the ones that price the SLA in, not the ones trying to sell software with extra steps.

![The shapeshifter firm — half blueprint, half services co](/post-images/sequoia-services-as-software-thesis/shapeshifter-firm.jpg)

## Why vertical SaaS is the casualty

![The vertical-SaaS casualty list — six category tiles (Legal, Finance, HR, Healthcare, Insurance, Real Estate) with the right column crossed out as services-as-software absorbs the workflow.](/post-images/sequoia-services-as-software-thesis/vertical-saas-casualty.jpg)

The implicit obituary in Bek's essay is the bigger story than the explicit forecast. Vertical SaaS has the wrong shape for this thesis to be true.

Vertical SaaS sells a workflow tool to the human doing the work. Toast sells a POS to the restaurant operator. Procore sells a project tool to the construction PM. Veeva sells a CRM to the pharma rep. The TAM is bounded by the seat count of the profession and the willingness to pay for productivity per seat. It is the $1 in Bek's ratio.

Three things make vertical SaaS structurally trapped against services-as-software.

The first is the ceiling. Revenue per customer is bounded by what the tool budget will tolerate. The labor budget is six times larger and accessible only by replacing the labor.

The second is the wrong customer. Vertical SaaS GTM is calibrated for end-users. The seller knows how to demo to a paralegal, an underwriter, a project manager. Selling outcomes to a CFO or COO is a different motion. It requires liability assumption, SLA contracts, and outcome metrics that vertical SaaS has historically treated as nice-to-haves. The org chart of a $5B vertical SaaS company is wrong for selling the work.

The third is the model risk. Bek's most rhetorical line is aimed at the copilot crowd. *"If you sell the tool, you're in a race against the model."* Every model improvement from Anthropic, OpenAI, or Google compresses the value of the workflow wrapper. If Claude 5 can read a contract and flag risks, the contract-review SaaS company has six months to become a service business or six months to be commoditized into a feature. The autopilot has the opposite incentive. Every model improvement reduces its cost of fulfillment.

This is why the innovator's dilemma in Bek's essay is so unsentimental. The copilots that built real businesses on top of GPT-4 (Harvey, Rogo, Hebbia) are sitting on customer relationships they cannot use without cannibalizing themselves. Their best customers are the practitioners they would replace.

A handful of vertical SaaS companies will pivot. Most will not. The ones that do not are the venture casualties of the next 36 months, and the funds that wrote those checks already know it.

The cleanest test for whether a vertical SaaS company is going to make it is to ask the CRO who they sell to. If the answer is the practitioner, they are a feature. If the answer is the practitioner's manager, and they are already pricing on outcomes, they have a chance. What makes the Bek essay unusual is that it is unsentimental about Sequoia's own portfolio: the vertical SaaS thesis the firm helped write in 2014 is, on the same logic, the casualty in 2026.

## The unanswered questions

Bek's essay is short for a reason. It is a thesis statement, and thesis statements are most powerful when they leave the hardest questions unaddressed. There are three.

| Question | The Bek essay's posture | What's actually unsolved |
| --- | --- | --- |
| Regulatory licensure | Acknowledged in passing | Who signs the return / brief / binder; how is the licensed human paid |
| Error liability | Not addressed | Insurance products for AI-generated work do not yet exist at scale |
| Labor backlash | Not addressed | Organized professional bodies are already lobbying |
| Foundation-model encroachment | Not addressed | Anthropic / OpenAI shipping outcome-priced products directly |
| Pricing floor | Not addressed | The "\$0.03 problem" — once five competitors per vertical race on price |

**Regulatory licensure.** Most of Bek's named verticals are professions with statutory practice rights. CPAs sign tax returns. Lawyers sign legal advice. Brokers sign insurance binders. Each licensing regime was written around a human practitioner with personal liability. An autopilot that "just closes the books" runs into a real question of whether it is practicing accounting without a license, and the same question lives in legal, insurance, and healthcare. The path most likely to work is a licensed human-of-record sitting on top of an AI fulfillment stack. The autopilots that win at scale will probably be the ones that figure out how to make the licensed human a $50-an-hour signoff layer instead of a $300-an-hour worker. That compresses the margin Sequoia is forecasting. It is doable. It is not free.

**Error liability.** A copilot's mistake is a human's mistake. The human is liable. An autopilot's mistake is the company's mistake. EvenUp's wrong demand letter is EvenUp's lawsuit. Anterior's wrong ICD-10 code is Anterior's audit. Ryan Gaines at N6 Finance [made the point well](https://n6finance.com/insights/aiforservices): the liability surface for autopilots is structurally larger. Insurance products to cover AI-generated services work do not really exist yet. They will. The question is who pays the premium and how much it eats the gross margin.

**Labor backlash.** The verticals Bek lists employ a combined 6-8 million people in the US alone. Accounting has a labor shortage and will not push back. Recruitment, claims adjusting, and legal services have organized professional bodies, and those bodies have already begun to lobby. The political surface looks more like the 2010s ride-hail fight than the 2010s SaaS rollouts. Sequoia's portfolio is not modeled for that.

There is a fourth question Bek skipped. What if the foundation models eat this themselves? Anthropic and OpenAI have both signaled interest in selling outcomes directly. A "ChatGPT for tax filings," with the work done at the model layer and no application company in between, would compress the autopilot moat to whatever proprietary data the application company accumulated. That data moat is real but unproven. It is the same moat Sequoia bet on for Uber and DoorDash, and it is the moat that takes the longest to compound.

## The Sierra benchmark — what a working autopilot looks like

Bret Taylor's [Sierra](https://sierra.ai/blog) is the most-referenced live example of a services-as-software company at scale. It is worth pulling apart because almost every founder pattern-matching on Bek's essay should be benchmarking against Sierra's shape, not against a generic "AI for X" pitch.

Sierra sells customer-experience agents. The customer is not a CX tool buyer. It is the VP of customer service at a Fortune 500. The pricing unit is the resolved interaction. The contract is an SLA. The inference happens on whatever model the agent currently routes to. The proprietary substrate is the runtime, the QA pipeline, the agent reasoning patterns Sierra calls "Agent Development Lifecycle," and the customer-specific behaviors that accumulate over a six-month deployment.

Three things separate Sierra from a vertical SaaS competitor selling into the same buyer.

First, the engagement starts with a paid pilot scoped against a containment metric, not a feature checklist. The deliverable is "we can autonomously resolve 47% of your tier-1 tickets at 92% CSAT." Vertical SaaS pilots are demos. Sierra's pilots are services proposals.

Second, the cost line on Sierra's P&L is dominated by inference and a small expert workforce that handles QA and edge cases, not by an engineering team building features. This is the architecture Bek is pointing at. The company is a software company in its filings and a services company in its operations.

Third, Sierra prices in dollars per resolved contact, not dollars per agent seat. That number is already comfortably below the offshore-BPO comparison, which is the line item it replaces. The economics work because the gross margin is software-shaped, not because the price is software-shaped. Sierra's per-resolved-interaction pricing is the cleanest services-as-software data point we have: undercutting offshore CX BPO on price while delivering software margins.

Founders building in this category should ask the same question Sierra answered. The framing is not "what's our software?" but "what is our equivalent of a resolved contact?" Per filing, per chart, per close cycle. Pick the unit before picking the model.

The other useful tell from Sierra is sales velocity. Enterprise software sales cycles in CX historically run nine to fifteen months. Sierra's pilot-to-production motion runs six to ten weeks because the buyer is approving an outsourced operation, not a software rollout. The procurement department treats the contract more like an MSA with a BPO than a standard SaaS order form. That alone is worth a 2x improvement in ARR-per-rep over a vertical SaaS competitor with a comparable ACV.

Founders modeling this category should pay attention to which line of the customer's P&L they will be billed against. SaaS lives on the IT line. Services-as-software lives on the operating line of whichever function it replaces. The operating line is where the budgets are bigger, the procurement is faster, and the displacement story actually closes.

## The data moat question

The most-asked follow-up to Bek's essay, especially from existing portfolio companies with vertical SaaS DNA, is whether the data accumulated by an autopilot constitutes a defensible moat against the foundation models themselves.

The answer is "sometimes, in some directions, and not the directions most founders are pitching."

What does not work as a moat. Generic interaction logs. Anonymized chat transcripts. Annotated reasoning traces sold as "proprietary training data." Foundation-model labs already have orders of magnitude more of these than any application company will accumulate in five years. Anthropic's own evaluation infrastructure consumes more annotated reasoning per quarter than a Series-B autopilot will see in its lifetime.

What does work as a moat. Outcome data tied to a specific customer's downstream system, where the autopilot's correctness can be verified against the customer's own ground truth. EvenUp's letters get accepted or rejected by carriers, and that acceptance signal is uniquely visible to EvenUp. Anterior's coded charts get audited by CMS and commercial payers, and the accept/deny signal is uniquely visible to Anterior. Rillet's closed books match or do not match the audited financials at year-end. These are signals foundation-model labs do not have access to without doing the application company's job, which is the whole point of the application company existing.

| Type of "data moat" | Defensible against foundation models? | Why |
| --- | --- | --- |
| Generic interaction logs | No | Labs have more |
| Annotated chain-of-thought | No | Labs generate this in-house |
| Customer's own private documents | Weakly — depends on contract | Customer can BYO model |
| Outcome verification tied to external ground truth | Yes | Lab cannot get this without becoming the application |
| Workflow expertise encoded as agent design | Yes, until labs ship vertical agents | Race against lab roadmap |
| Regulatory + compliance certifications | Yes | High switching cost on customer side |

The moat in services-as-software is not the model, the data, or even the agent design. It is the customer relationship that lets you see whether the work was right. Foundation labs cannot see that without becoming the application company themselves. The cleanest example is recruitment: a Mercor-shaped autopilot knows which candidates got hired and which did not six months later, a signal that does not exist anywhere else in the recruitment stack. Every model improvement above that data layer makes the pipeline cheaper to run, not weaker.

The corollary, and the part Bek does not say, is that services-as-software companies that try to be defensible at the model layer are fighting a war they cannot win. The defensibility lives in the integration with the customer's downstream verification system. That integration is hard, slow, vertical-specific, and exactly the work foundation labs do not want to do because the integration does not generalize.

## What this means for the rest of the venture stack

Three structural shifts follow from the thesis if it holds.

The first is GTM. Most application-layer venture-backed companies were built with sales orgs designed for SaaS. Reps demo to end-users, expand seats, and close on annual contracts in the \$25K-150K range. Selling outcomes to operating-line buyers is closer to enterprise services sales. ACVs are higher. Cycles are longer at first, then shorter once a vertical reference is established. Reps need to understand the customer's operating P&L, not the feature list. The number of hires that survive the transition is small.

The second is the partner model. Boutique consultancies, system integrators, and outsourced services firms are the natural channel for autopilots in the regulated verticals. They already have the licensure, the customer relationships, and the trust budget. The autopilots that scale fastest are likely to do so by riding existing services firms rather than competing with them. This was the Y Combinator-friendly version of the Bek essay that Michael Seibel articulated in March: build the autopilot, sell to the regional firm, let the firm sell to the end customer.

The piece reached past the VC cohort fast. Luke Harries flagged the London follow-on:

> The next AI Tinkerers London is on the 7th of April - featuring a fireside chat with @JulienBek, Partner at @sequoia, after his recent viral article on "Services: The new software" alongside technical talks from leading startups
>
> — [@lukeharries](https://x.com/lukeharries/status/2038546827209806034), Mar 30, 2026

The third is the LP conversation. Funds that underwrote vertical SaaS at \$50M ARR exit assumptions will not get the markups Bek's essay implies for the autopilots. Fund construction shifts toward fewer, more concentrated bets at Series A and B in named verticals, with longer time-to-revenue and bigger reserves for follow-ons. Several funds have already started repositioning. The honest LP letters from early-stage funds with 2018–2022 vertical SaaS portfolios this year say some version of "the multiples we underwrote against do not exist anymore in our category." The leading indicator that the thesis is being internalized in the seed market is the W26 YC batch, which is pitching outcome-priced from day one in noticeably higher numbers than W25.

## The founder checklist — does my company fit the shape?

![The founder checklist — five questions to test fit: high-touch services replaced by AI, per-outcome pricing, integration into legacy ops, data moat from operations (the load-bearing question, in editorial red), and human escalation for edge cases.](/post-images/sequoia-services-as-software-thesis/founder-checklist.jpg)

If you are pattern-matching on the Bek essay, the honest test is not whether you can describe your company in services-as-software language. It is whether the architecture, the customer, and the pricing all line up. Most companies that talk this way pass one or two of the three. The ones that pass all three are rare, and they are the ones that fit Sequoia's thesis.

| Question | Pass | Fail |
| --- | --- | --- |
| What does the customer's invoice from you say? | "Per closed book / per filing / per resolved interaction" | "Per seat per month" |
| Who signs the PO? | The operating-line owner (CFO, COO, GC, CRO) | The IT or department-tool budget |
| What line of the customer's P&L do you replace? | Labor / outsourced services | Software subscriptions |
| What does your gross margin scale with? | Inference + a thin licensed-human signoff layer | Engineering headcount + sales |
| What happens when the underlying model gets better? | Your COGS goes down | Your moat shrinks |
| What is the SLA you sign? | Outcome-quality + volume guarantees | "99.9% uptime" |
| What does the buyer reference look like at year two? | "We retired N FTEs / replaced our BPO contract" | "Our team is more productive" |

A company that answers in the left column on five or more of these questions is the shape Bek is describing. A company that answers in the right column on three or more is, in his framing, a vertical SaaS company with AI features. The honest assessment matters more than the marketing copy. The multiples are decided by the answer to the first question, not the deck.

The single most useful filter when evaluating a pitch in this category is to ask the founder what unit they bill on. If the answer takes more than five seconds or starts with "well, it depends," the company is a SaaS business telling itself a services-as-software story. There is also a subset of vertical SaaS founders pattern-matching the thesis by flipping a pricing page from per-seat to per-outcome and calling it a day. That is not the work. The work is rewiring GTM, contracts, and the cost structure to match.

## The decoded thesis

The next $1T company is, by Sequoia's framing, not really an AI company. It is a service business that AI made cheap to scale. The product is the work. The buyer is the budget owner who used to hire labor. The economics look like software because the labor line has been collapsed into compute. The valuation looks like software because public markets price gross margin.

The implicit obituary for vertical SaaS is the part of the essay that will age hardest. Sequoia funded most of vertical SaaS in its current form. Bek's piece does not say "we were wrong." It says the bigger version of the same insight is one floor up the org chart. Vertical SaaS sold software to the worker. Services-as-software sells the work to the worker's manager.

Trent Hughes captured the implication founders are running with:

> My predictions for AI this year:
>
> 1. AI reliability becomes the most urgent problem in enterprise. Deploying AI is easy. Knowing if it works is not.
>
> 2. Services as software explodes. One person agencies doing the work of 20. Margins that look like software companies.
>
> 3.
>
> — [@trentjhughes](https://x.com/trentjhughes/status/2045203890669486421), Apr 17, 2026

If the thesis holds, the trillion-dollar question is no longer "which industry has bad software." It is "which industry has expensive labor that follows rules." That is a different map, and Sequoia just published it. Some founders will read the map as a permission slip and spend the next 36 months losing $20M of LP money to sales cycles they did not understand. A handful will read it as the brief it actually is, pick a vertical with a high intelligence ratio and a tractable regulatory surface, and build the business that proves the thesis right. The second group is small. It is also the group Sequoia is writing for.

# Update, May 22, 2026 — applied to the YC Summer 2026 RFS

The Deep Feed applied the founder checklist above to YC's Summer 2026 Request for Startups, scoring each of the 16 RFS lanes on the seven-question filter. The split is sharper than expected. Four lanes (AI-Native Service Companies, AI Personalized Medicine, Startups Selling to Huge Companies, AI for Low-Pesticide Agriculture) fit the services-as-software shape cleanly. Three are hybrids. Nine are not the shape, and the four lanes the Spring 2026 batch crowded into most heavily (Software for Agents, the AI Operating System for Companies, SaaS Challengers, Dynamic Software Interfaces) are mostly on the wrong side of the multiples line. Three Spring 2026 companies cleared the full seven-question checklist on their public one-liners: Klaimee (liability insurance for AI agents), Panacea (AI-Native FDA Regulatory Services), and Astraea (clinical-trials agents). The full mapping, with the per-lane verdict and the cross-reference to the Spring batch evidence, is in [Three weeks after the YC Summer 2026 RFS, the Spring batch is the field check](https://www.thedeepfeed.ai/posts/2026-05-22-yc-spring-2026-rfs-three-week-field-check/#the-sequoia-1t-overlay).

## Sources

- [Sequoia Capital — Services: The New Software (Julien Bek)](https://sequoiacap.com/article/services-the-new-software/)
- [Sequoia Capital — The $10 Trillion AI Revolution (Konstantine Buhler)](https://sequoiacap.com/article/10t-ai-revolution/)
- [Sequoia Capital — AI Ascent 2025 Keynote](https://www.youtube.com/watch?v=v9JBMnxuPX8)
- [Sequoia Capital — Partnering with Edra](https://sequoiacap.com/article/partnering-with-edra-context-for-agents-at-scale/)
- [Linas Beliūnas — Sequoia's Thesis Will Mint Billionaires and Bankrupt Copycats](https://linas.substack.com/p/sequoiathesis)
- [N6 Finance — Sequoia Is Right That AI Is Coming For Services, But Three Things Don't Get Talked About Enough](https://n6finance.com/insights/aiforservices)
- [Han Heloir Yan — Services Are the New Software: Building Them Is the Hard Part](https://medium.com/data-science-collective/services-are-the-new-software-building-them-is-the-hard-part-ca2d3ff9aad4)
- [Harvey — official site](https://www.harvey.ai/)
- [EvenUp — official site](https://www.evenuplaw.com/)
- [Mercor — official site](https://mercor.com/)
- [Crosby — official site](https://www.crosby.ai/)
- [Lawhive — official site](https://lawhive.co.uk/)
- [Rillet — official site](https://www.rillet.com/)
- [Anterior — official site](https://anterior.com/)
- [Sequoia Capital — Generative AI's Act Two (Pat Grady, Sonya Huang)](https://www.sequoiacap.com/article/generative-ai-act-two/)
- [Sequoia Capital — AI 50 2025](https://www.sequoiacap.com/article/ai-50-2025/)
- [Sequoia Capital — Partnering with Sierra (Bret Taylor)](https://www.sequoiacap.com/companies/sierra/)
- [Sequoia Capital — Partnering with Harvey on the AI Future of Law](https://www.sequoiacap.com/article/partnering-with-harvey-putting-llms-to-work/)
- [Bessemer — State of the Cloud 2025 (multiples comps)](https://www.bvp.com/atlas/state-of-the-cloud-2025)
- [Tomasz Tunguz — The End of Vertical SaaS as We Know It](https://tomtunguz.com/the-end-of-vertical-saas/)
- [Benedict Evans — AI eats the services stack](https://www.ben-evans.com/benedictevans/2025/ai-eats-services)
- [a16z — The AI-First Services Company](https://a16z.com/the-ai-first-services-company/)
- [BLS — Occupational Outlook Handbook (Accountants & Auditors)](https://www.bls.gov/ooh/business-and-financial/accountants-and-auditors.htm)
- [Sierra (Bret Taylor) — Building the agentic enterprise](https://sierra.ai/blog)
- [EvenUp — How AI is changing personal injury settlement](https://www.evenuplaw.com/blog)
- [Mercor — Letter from the Founders 2026](https://mercor.com/blog)

---

Canonical: https://www.thedeepfeed.ai/posts/2026-04-30-sequoia-services-as-software-thesis/
Site: https://www.thedeepfeed.ai
Full corpus: https://www.thedeepfeed.ai/llms-full.txt