Clients have always paid accounting firms, law firms and agencies for finished work: books that close, contracts that can be signed, campaigns that go live. No client has ever asked how many people it took on the inside.
Now some people have spotted that gap and started a new kind of company. These companies let AI do most of the work, then sell the finished work at a price set against the cost of hiring people. Investors call this kind of business AI-native services (service businesses where AI does the work itself). Y Combinator, or YC, one of the largest startup accelerators, has called for companies like this in three rounds in a row since mid-2025.
At least two playbooks for this space already exist: a video by Charlie Warren (a Visiting Partner at YC, speaking from what YC has seen across the companies it backs) and an article by Jake Saper, Lotti Siniscalco, Rishub Nahar and Kabir Sial of Emergence Capital. Both say up front who they are written for. Warren says it directly: "It's aimed at people thinking about starting a company, not if you're already running one."
But the people who will compete with you for the same budget are reading every line of these playbooks. If you already sell work to clients, this playbook is your competitor's plan, free to read.
In this article, Productize summarizes the AI-native services playbook (a step-by-step guide) from four investors: YC, Emergence Capital, Sequoia and Foundation Capital. One thing to say once, right here: every source is an investor describing companies in its own portfolio, and none of them shows a case that failed.
Part 1Which service businesses will AI-native services reach first?
This work fits the pattern when clients already outsource it, when it breaks into pieces, and when most pieces need no judgment call, yet the whole job is hard enough that clients won't accept it without a skilled person in charge. Warren sums this up as 4 traits.
| Trait | What it means | Ask yourself |
|---|---|---|
| 1. Clients already outsource it | They care about the result, not how it gets done | Do clients pay for a finished result? |
| 2. The sub-tasks need little judgment | Most pieces follow rules; judgment sits at only a few points | If every piece needs a person to decide, this work can't scale |
| 3. The whole job is hard enough | It takes both the model and people before clients accept the work | If the model could do it all alone, clients could do it themselves |
| 4. Regulation becomes a moat | Laws govern the work, expectations are high, newcomers struggle to enter | Who is liable if this work is wrong? |
Trait 1 is the one to read slowly. Julien Bek of Sequoia writes: "Replacing an outsourcing contract with an AI-native services provider is a vendor swap." For you, this sentence means the work clients hire you for is the easiest work to switch vendors on, because clients don't have to change how they work at all.
You can test trait 2 by breaking the work into pieces. We tried this with accounting; see which accounting tasks have already moved to AI. Other businesses can use the same method; only the list of tasks changes.
Warren adds one more question: "as the models get better, does your service get stronger, or does the model itself commoditize you?" If it's the second, every new model release leaves you weaker.
Foundation Capital sees it differently: the companies that survive are the ones doing messy work full of exceptions that the model owners can't see. Neither view has data to back it yet.
Part 2How do they price when the real competitor is labor?
AI-native services companies set prices against the cost of hiring people, not against the cost of software. Warren explains: "you're not competing with other software providers, you're competing directly with the cost of labor, internal or outsourced."
Bek gives an illustrative figure: a company might spend $10,000 a year on accounting software, but $120,000 a year on an accountant. The second amount is the one these companies are after. The price clients compare you against is a salary, not a software fee.
| Pricing approach | Example | Why |
|---|---|---|
| Per unit of work | Per tax return, per claim | The clearest, and the easiest to explain |
| By outcome | Per completed study instead of per hour | Client and vendor win in the same direction, but revenue is hard to forecast |
| Start with labor-based pricing | Hourly rates the market already knows | A workable start, but set a deadline to move to outcome-based pricing. Otherwise, the more you automate, the more you eat into your own growth (Emergence) |
| Never cost-plus | Your cost plus a set markup | Every cost cut lowers your price right away, so revenue is capped for good |
| Never straight undercutting | Cheaper than the incumbent for the same work | The work looks cheap and low quality |
The examples these sources give span several industries.
- FDA submission consulting (the US Food and Drug Administration): Panacea hires experienced FDA consultants to work alongside an AI platform and charges per completed study, while the market charges by the hour.
- Law: Crosby sells nondisclosure agreement drafting directly to companies, without going through a law firm.
- Insurance: Strala takes on only claims processing for insurance carriers: a narrow scope, but the steps repeat.
Part 3Growing revenue doesn't mean the AI is doing the work
Emergence has a name for this, Mirage PMF: "The illusion of product-market fit created by revenue growth that's powered by human labor rather than AI leverage."
If you book the cost of people reviewing work in the wrong category, sign 1 is the easiest one to be fooled by. Emergence warns: "Too many founders offload labor costs to operating expenses, but since this is a service, they are absolutely COGS." Warren also splits COGS into three buckets: model costs, infrastructure costs, and the cost of people reviewing the work. Each bucket needs someone who owns the number.
One example of the core metric in sign 5 is Crosby's HURT (Human Review Time), which measures the minutes a person spends reviewing one document after the AI has finished it, without letting quality slip. Emergence says the closer HURT gets to zero, the closer margins get to those of a software business.
But cutting review time to make the numbers look good is risky in the other direction. Warren says: "Customers will fire you for variance faster than they will fire you for being a bit slower or a bit more expensive than the incumbents." Inconsistent work gets you fired faster than a high price. A survey of 1,150 full-time employees across industries, cited by TechCrunch, also found that 40% have to take on extra work because of AI output that looks polished but has no substance.
We ran into this ourselves when we had 11 AI agents help read the source documents for this article. Every one of them reported that its quotes matched the originals exactly, but when a program checked them against the real sources, 36 of 639 did not match.
The same principle is behind a service we designed ourselves: AI that reads your bills, on your own machines. The service sells an Excel file you can import into your accounting system, not a model. The service page says: We don't let AI do the math for you
. AI only reads the numbers off the bill; the addition, subtraction and tax are the job of a fixed calculator. When a bill's numbers don't add up, the system sends it to a person instead of guessing. For where to place the person who reviews, read Human in the Loop: a three-tier model.
Every investor quoted in this article is based in the US: Bek's $10,000 versus $120,000 figure is an illustrative case priced in dollars, and the FDA is a US regulator. Wages in Thailand are lower, so the gap between software spend and labor spend described in Part 2 does not carry over as is. Two ideas still hold for your service business, though: knowing your cost per piece of work (item 2 of the checklist in Part 4) matters more, not less, when prices are already low and there is little room for error. Rules as a moat (trait 4 in Part 1) also has Thai equivalents, because many service fields in Thailand are governed by laws and professional licenses too.
Part 4What can an existing service business use tomorrow?
Warren warns founders not to buy an existing service business and add AI to it, because "You just can't acquire a product market fit." An existing business has its own way of measuring results, its own way of hiring, and its own expectations, and adding AI doesn't change those right away.
That warning tells you two things. The playbook tells new competitors to build something new to compete with you, not to buy your company. And you can't just lay AI on top of how you already work and wait for results either. What you have that they don't is real work with real clients. Use that advantage by starting with one piece of work, not the whole company.
Once you find work that should move, the next question is whether to fix the process, build your own, or buy. The Fix, Build, or Buy self-assessment takes 8 questions and tells you which way that work should go, with three-year numbers. If the pieces you can move chain together, we offer a service for this, Standing up an agent fleet, which helps design the scope of an agent team (AI that can carry out several steps in a row on its own) and sets up the review system at the same time.
Think about the work clients paid you for last month. Which pieces did they pay for because of the finished result, without ever asking how many people it took? Those are the pieces a competitor will come for first, and the ones you should try breaking down before they do.
Sources and references
- Charlie Warren (YC), How to Build an AI-Native Services Company, Startup School, June 3, 2026 Source
- Jake Saper, Lotti Siniscalco, Rishub Nahar, Kabir Sial (Emergence Capital), The AI-Native Services Playbook, March 30, 2026 Source
- Julien Bek (Sequoia Capital), Services: The New Software, March 5, 2026 Source
- Ashu Garg, Jaya Gupta (Foundation Capital), When model providers eat everything, November 5, 2025 Source
- Connie Loizos (TechCrunch), The AI services transformation may be harder than VCs think, September 28, 2025 Source
- Y Combinator, Requests for Startups Source
Read next: All articles · Which accounting tasks have moved to AI · Human in the Loop: a three-tier model