How to build a service-as-software company

A new kind of company is being built right now, and it is not a software product. It is a service: tax, audit, insurance, law, parts of healthcare, rebuilt from scratch with AI doing most of the work and a human in the loop where it counts. Y Combinator calls these AI-native service companies; the broader name for the model is service as software. Instead of selling software the customer operates, you deliver the finished outcome. The markets are measured in trillions, and the opportunity barely existed two years ago.

YC visiting partner Charlie Warren laid out a playbook for founders starting one from scratch. Here is the distilled version: how to pick the market, build the team and the product, sell, price, and read the P&L, plus the traps that sink these companies.

What "service as software" means

In normal software, you sell a tool and the customer does the work with it. In a service-as-software company, the human is the interface and the software scales them. The customer hands you the problem, a tax return, a contract, an FDA submission, and you hand back the finished result. Better models made this possible: they are now good enough that models plus a few humans can deliver an outcome the customer will accept and pay for, not just a co-pilot the customer has to drive.

Why it matters for founders: traditional services firms top out around 30 percent margins. The bet here is that AI operating leverage (more on that below) pushes you toward software-like margins, 50 percent and up, on markets two to three times bigger than software.

Pick a market with these four traits

The usual startup advice still applies: pick a market you would happily work in for a decade, because these companies take ten years or more. On top of that, the best AI-services markets share four traits:

• Low trust, already outsourced. The work is already sent to a vendor, and the customer cares about the result, not how it was made. You are displacing a vendor, not asking anyone to change behavior. You show up where the budget already lives.
• Low judgment at the task level. Break the work into steps. If every step needs a human exercising real judgment, you cannot scale. You want most steps automatable, with judgment concentrated in a few places where humans stay in the loop.
• A high intelligence threshold. The overall work has to be genuinely hard, hard enough that models plus humans are required to deliver it.
• Regulation can be a moat. Regulated industries carry higher expectations and legal accountability, which raises the bar for everyone and protects you once you clear it. Panacea, a YC company, delivers FDA regulatory services for biotech and medtech by pairing experienced FDA consultants with an AI platform.

Then run two checks. First, the Sam Altman test: as the models get better, does your service get stronger, or does the model commoditize you? You want the first camp. Second, an honesty check: are you using humans because the work genuinely needs judgment, or to paper over gaps in your product? Be honest, or you will scale a problem instead of a company. And be careful with anything involving equipment and on-site labor, where the software-margin math does not apply because you own physical things.

Build the team: domain, model, and operational fluency

Start with people you have already worked with. If you are solo, ask the best people you know; you would be surprised who says yes. For AI services specifically, the strongest founders share three traits:

• Domain fluency. You are selling to skeptical buyers in regulated spaces, so you have to bleed credibility. Direct experience is best, learned is acceptable.
• Model fluency. Know what frontier models can do today and design the product to ride the curve as they improve. There is no substitute for strong technology here, and founders underestimate it.
• Operational rigor. Variance, throughput, cycle times, SOPs. Not glamorous, but you are running an operation and you have to respect that skill set.

A good example is General Legal, an AI-native law firm YC backed. The founders pair law-firm experience (Cooley, Fenwick) and legal-tech depth (CaseText) with shift work that cuts cycle times. The operations thinking is the point. You can see the broader pattern on our AI for legal hub.

The product is the operation

This is the part that breaks software founders' instincts. The product is not the thing the customer uses; it is the operation that produces the outcome. So you build like an operator:

• Build for the bottlenecks. Throughput and cycle time are now product metrics. Track them the way a SaaS team tracks active users.
• Variance will kill you. Non-uniform output is the existential risk. Customers will fire you for inconsistency faster than for being a little slower or pricier than the incumbent. Inconsistency destroys trust, and trust is the whole product. As a CISSP, I would go further: in regulated work an auditor may read everything you produce, so treat reliability and a clean audit trail as features, not afterthoughts, and hold client data with least access and real logging.
• Humans must scale non-linearly. If revenue only grows in step with headcount, you have a consulting firm, not a software-margin business. The humans in the loop are also your users, so the internal software has to be good enough that they like using it.

It is fine to do things that do not scale at the very start. But automating the process is the product, so eventually you have to.

Sell outcomes, and dodge the early-demand trap

The biggest early mistake is the demand trap. When you have nothing, it is easy to sign a pile of pilots, and then you drown serving them by hand and never build the product that scales. Cap your first pilots to a small handful, on purpose, and resist signing too many too fast.

Then sell the outcome, not seats or tokens. For the first few customers, the pilot is the product, so do not standardize too early. Use the pilots to find where AI gives you a real edge versus where you are just automating something obvious, and build accordingly, fast.

Price on value, and read the P&L

Pricing is harder than in software because you are not competing with other software, you are competing with the cost of labor. Two approaches work: per-unit pricing (per return, per claim, per loan), which is the cleanest to explain, and outcome-based pricing, which aligns incentives but is harder to forecast. Panacea, for instance, prices on the completed study rather than by the hour. Two approaches to avoid: cost-plus, which caps your upside forever, and straight-line undercutting, which makes your work look cheap and low quality.

Then watch the P&L, because these companies live or die on it. Your cost of goods sold has three parts: model costs, hosting, and the humans in the loop. Give each a number, a trend line, and an owner from day one, and be suspicious of zero- or negative-margin pilots. The core bet is AI operating leverage: the more of the product you build, the lower your COGS and the better your gross margin. You do not need software margins on day one, but the trajectory has to be believable.

Don't buy your way in

There is a tempting shortcut, especially for founders with an operating background: buy an existing services business, bolt AI on top, and skip the slow revenue build. It almost never works. You cannot acquire product-market fit, and a legacy firm carries legacy expectations on metrics, hiring, and performance that AI does not magically change. The one decent reason to buy is a fast regulatory moat, like insurance licensing. Otherwise, building beats buying.

What to do this week

1. List markets that are already outsourced, regulated, and hard, where you have or can build credibility. Score each on the four traits.
2. Run the Sam Altman test on your top pick: do better models make you stronger or commoditize you?
3. Sketch the work as steps and mark which ones need human judgment. If it is most of them, keep looking.
4. Name the three people you would want as co-founders for domain, model, and operational fluency.
5. Write a one-page P&L guess: a per-unit price, the three COGS lines, and where margin improves as you automate.

The throughline is that the process is the product and the product is the process. That is the same operating discipline we teach in AI Operating System for Startups: put AI to work across the company, keep a human on the calls that matter, and build for reliability. For the broader version of this shift, see how to build an AI-native company, and for the mindset that powers it, how to build an AI-first company.

Sources

• How to Build an AI-Native Services Company, the Y Combinator video by visiting partner Charlie Warren that this article distills.
• Company examples: Panacea (AI-native FDA regulatory services) and General Legal (AI-native law firm). Background on General Legal's founding team: Law.com.