AI-native services: a $100B opportunity

AI-native services are here, and I think they're the best business you can start right now. There's probably $100 billion up for grabs. This is the full guide: what AI native services are, how to build one, and the opportunity map.

This is pretty DENSE article so grab a coffee/tea, give it 15 minutes, and let's get into it.

Background

For most of my career, running a service business meant pretty good living and a bad company (at least in the eyes of many VCs). You sold hours, the revenue stopped when you stopped, and the whole thing lived on a few people who could leave. That's changing, and the reason is simple enough to fit in one example.

A business pays about $10,000 a year for QuickBooks. It pays about $120,000 a year for the accountant who uses QuickBooks. For twenty years, software companies fought over the $10,000, because that was the part you could sell at scale. The $120,000 was locked behind a human, and you couldn't scale a human without hiring another one. That's the money that just opened up. AI can now do most of what the accountant does, which means you can sell the finished work, the closed books, instead of the tool, and you can do it at software margins.

That's what an AI-native service is. The work is done mostly by agents, and a small team handles the parts that still need a person. How you charge is up to you, a monthly retainer, a per-unit price, usage-based, whatever fits the customer. The thing that makes it AI-native is that the delivery runs on software instead of headcount.

This essay walks through every piece of it: what it's made of, how to build one, and the opportunity map.

What an AI-native service actually is

The simplest way to think about it: a normal service business sells a person's time. A SaaS business sells a tool and makes the customer do the work. An AI-native service does the work for the customer, and agents do most of it.

The customer doesn't want bookkeeping software. They want their books closed. They don't want a contract-review tool. They want to know if the contract is safe to sign. An AI-native service delivers that, and the customer experiences it exactly like hiring a firm, except it's faster, cheaper, and it doesn't get tired. Whether you bill them monthly, per job, or on usage is a separate decision.

The trick is that under the hood it runs like software. The work is done by a system you built once, not by a headcount you grow forever. That's why it can be priced like a service and valued like software, which is the whole opportunity.

Why now

This wasn't possible 2 years ago, and the reason is a few things landing at the same time.

The models got good enough at the actual work. Reviewing a contract, coding a medical claim, drafting a demand letter, closing a month of books. Two years ago the output was a rough draft a person had to redo. Now it's a finished draft a person checks.

The cost of running them collapsed. A task that cost dollars in tokens now costs cents, which means your cost to deliver one more unit of work is close to zero. That's the number that turns a service into something with software margins.

And the money was always there. Businesses in the US spend about $4.6 trillion a year on services, roughly six times what they spend on software. That's the accountant, the broker, the paralegal, the bookkeeper, the compliance specialist. For two decades that pool was untouchable because every dollar of it was attached to a person. Now it's the biggest open market in the economy, and it's kinda crazy how open these markets truly are.

You can see it in the companies that moved first. Harvey, doing legal work, went from roughly $100 million to $190 million in annual revenue in about five months. EvenUp sells demand letters to injury firms at around $500 each, for work that used to eat eight to twelve hours of an associate's day, and crossed $50 million in revenue. Kick does small-business bookkeeping for $300 to $500 a month, undercuts a human bookkeeper by half, and still runs gross margins above 70 percent. These are service businesses. They just happen to have the economics of software.

Why services, of all the things you could start

Most people building right now are choosing between a SaaS product, an app, or an agency. I'll make the case why an AI-native service beats all three.

SaaS is running up a down escalator. You sell a tool, and the tool competes with a foundation model that gets better and cheaper every quarter. Every time a lab ships, your product is worth a little less. An app is dependent on the App Store, so you're always playing a dance (and margins aren't as high as a result because you're giving Apple 30%). And a classic agency still sells hours, so its revenue is capped by its headcount and its margins are stuck at 20 to 30 percent forever.

An AI-native service flips all of that. You sell the finished work, so when the models get better, your business gets better. Model improvement works for you instead of against you.

The budget already exists. You're replacing a line item the customer already pays, so you don't have to convince anyone they have a problem or invent a new category. Sales is "we do what your current firm does, faster and cheaper." That is the easiest pitch in business.

It cash-flows from day one. A SaaS founder guesses at product-market fit for a year before revenue. A service founder gets paid for the first job. You learn what to build by getting paid to do it. It also much more of a personal type business. Anecdotally, my friends who run high margin, smaller businesses powered by AI seem happier.

The wedge

The coolest part is that a service is the best wedge there is.

You use it to get into a customer, a niche, an industry, before you know exactly what to build. You do the work, you get paid, and every job teaches you where the real pain is, what "correct" means, and which parts a machine can own. Then you productize what you learned. Then, if you want, you turn that into software. You climb from service to product with a business paying you at every rung, instead of raising money to guess.

The biggest companies in AI figured this out already. The forward-deployed engineer model, an engineer embedded with the customer who builds instead of advises, is a service used as the wedge into an enterprise. It's how the winners get in the door. You can run the exact same play at a smaller scale in a niche you know.

The pieces

Every AI-native service has the same parts. If you understand these, you can design one for almost any niche.

1. The unit. This is the single most important decision. You sell one clearly defined thing: per claim, per filing, per contract, per month of books, per report. Never per hour. A retainer is fine, as long as it's for a clearly defined scope, because the unit is still what you deliver, you're just billing for it monthly instead of per job. The unit is what lets you price it, deliver it repeatedly, and eventually turn it into software. Pick something with a clear finish line, where the customer can look at it and say "done."
2. The intake. This is how the work comes in, and it's where most service businesses bleed money without noticing. An agency takes three calls to scope one job. An AI-native service takes a form. The customer uploads the document, fills in five fields, and describes what they need. That's it. If you can't define intake as a form, your unit isn't clear enough yet.
3. The engine. The AI that does the work. This is the model plus your instructions, your examples, your context about this specific industry. On day one it's mostly a good prompt and a few dozen examples. Over time it becomes something a competitor can't replicate, because it's shaped by every job you've run.
4. The rulebook. This is the real product, and almost everyone underrates it. It's the written list of what "correct" means in your niche and every way the AI gets it wrong. "A home health note is incomplete without vitals." "A demand letter must cite the treatment dates." "This kind of claim gets denied when the code doesn't match the diagnosis." You build this list one mistake at a time, by reading outputs and writing down what you caught. After a few hundred jobs, this rulebook is the thing that makes your output trustworthy and your business defensible.
5. The review layer. Where a human still looks. You decide, per unit, what ships automatically and what a person checks first. Low-stakes, high-confidence work goes straight out. Anything with money, legal exposure, or a customer's reputation attached gets a human eye. The review layer is how you own the mistakes without drowning in them, and the rulebook shrinks it over time.
6. The delivery. How the finished work gets back to the customer. A dashboard they log into, an email with the file, a portal that shows status. This replaces the account manager. The customer should be able to see where their job is without asking anyone.
7. The pricing. You've got two good options and one bad one. Per unit, where you charge for each claim, filing, or letter, works when volume is predictable and the customer wants to pay for exactly what they use. A flat monthly retainer, where you charge a fixed fee for a defined scope, works when the customer wants budget certainty and you want revenue you can count on. Kick's $300 to $500 a month for bookkeeping is a retainer, and it's a great business. The only thing to avoid is pricing tied to hours, because that chains your revenue to your headcount, which is the exact trap that made agencies bad businesses. Whichever you pick, price against the human alternative, not your costs. If a firm charges $2,500 for the thing, you charge $800. It looks like a bargain to the customer and it's a fortune to you, because your cost to deliver one more unit is close to zero.
8. The distribution. How you get customers. For most of these niches the honest answer is cold outbound to a tight list of the exact person who writes the check, plus one lead magnet: do the first job free. A home health agency that gets ten notes reviewed for nothing, and sees three problems they missed, becomes a customer that afternoon. Content works too if you're already making it, and so do partnerships with the software these businesses already use and referrals in a niche where everyone knows everyone. You don't need an audience. You need a channel to the buyer, and a free first job is the fastest one there is.

One built all the way through

Here's what the pieces look like assembled, with real numbers.

Take home health note review. A mid-size agency produces around 2,000 visit notes a month and pays a nurse reviewer roughly $70,000 a year to check them, because an incomplete note gets the claim denied or flagged in an audit.

The unit is one reviewed note. Intake is an upload, the agency drops in the day's notes. The engine runs each note against the rulebook you built, the twenty things that get a note denied: missing vitals, a vague medication change, care that doesn't support the billed level. The review layer sends clean notes straight back and flags the risky ones for a quick human look. Delivery is a dashboard showing every note, what passed, and what needs fixing, so the agency never has to ask.

You charge $2 a note. That's $4,000 a month from one agency, and your cost to run a note through the model is a few cents. Ten agencies is $40,000 a month. Fifty is $2.4 million a year, and one person can run it, because the intake is an upload, the checking is a rulebook, and the delivery is a dashboard. You started by reviewing notes for five agencies by hand, and the rulebook you wrote doing that is now the whole product. That's the shape of it. A real business, in a niche nobody's fighting over, that you can start this month.

How to build one

First, pick a box on the map (below). You want work that a company already pays an outside firm to do, and that has a checkable right answer. Both of those matter, and I'll explain why in a second.

Second, find five customers and do the work for them, mostly by hand. Use AI as the engine, but read every single output yourself before it goes out. This is not the product yet. This is how you learn what correct looks like, where the AI breaks, and what the customer actually cares about. You'll get all of that in a few weeks by doing the work, which you'd never get by guessing in a pitch deck.

Third, write down every mistake. Each time you catch something wrong, add it to the rulebook. After a few dozen jobs you'll have a list of the twenty things that keep going wrong in this niche. That list is your product. Nobody else has it.

Fourth, productize. Fix the scope, fix the price, turn intake into a form, turn delivery into a dashboard, and run the rulebook automatically. Now one person can handle fifty customers, and the business has the margins of software with the sales cycle of a service. This is where most of these businesses should live, and it's a great place to stay.

Fifth, and only if you want to, turn it into software. Once the rulebook and the workflow are solid, you can let customers run it themselves. That's the jump from a productized service to a product, and it's where the highest valuations are. But you'll have earned it, because you built the software by doing the work first instead of guessing what to build. You can always optionally raise VC once you get to this point if you want.

I plan on doing an episode this week of the pod going deeper into this @startupideaspod on YouTube/Spotify/Apple.

The map

Two questions decide almost everything about whether an idea works.

One: does the customer already pay an outside firm to do this? If yes, the budget exists, the scope is already defined, and switching to you costs them nothing. If they do it in-house, you're trying to replace someone's employee, which is a much harder conversation.

Two: is there a checkable right answer? If the output can be verified against a rule, AI plus your rulebook can own it. If it needs a real judgment call, a human has to stay in the loop.

Put those two on a grid and you get four boxes.

Already outsourced and checkable: build here. This is the sweet spot. You're a cheaper, faster version of a firm they already write checks to, and you can prove the work is right.

Checkable but done in-house: a real opportunity, harder sell. Position it as a tool the team keeps, not a replacement for a person.

Already outsourced but judgment-heavy: keep a human in the loop and charge a premium. AI does the prep, a person makes the call, and you can price close to what the old firm charged.

In-house and judgment-heavy: skip it. Because that's a job, not a business.

The opportunities

Here are real top-right boxes, each with the unit you'd sell. None are glamorous. That's on purpose.

• Medical billing and coding, per claim. Clinics already outsource it. A wrong code means a denied claim, so correct is checkable and expensive to miss.
• Commercial insurance quoting, per policy. The broker's core job is shopping carriers and filling forms. It's mostly a task with a checkable output.
• Customs and freight classification, per shipment. Every import needs the right code and paperwork. Rule-bound, document-heavy, costly when wrong.
• Regulatory filings for licensed businesses, per filing. Finance, food, healthcare, all pay specialists to follow strict rules, and the rules are the spec sheet.
• Lease abstraction and title work, per document. Commercial real estate is buried in this and already farms it out.
• RFP and grant writing, per submission. Mostly structure and precedent, and the budget already exists as a line item.
• Property tax appeals, per appeal, with a share of the savings. Already outsourced, checkable, and the customer only pays when you win. (I did this recently and paid a firm 50% of savings, you can do the same service for 10% of savings!)
• Home health visit note review, per note. Notes have to be complete or billing gets denied, and agencies review them by hand today.
• Demand letters for injury firms, per letter. A junior associate spends hours on each. The structure is repeatable and the output is checkable.
• Accounts payable exception handling, per exception. Companies pay people to chase the invoices that don't match, and the matching is a rule.

Look at the pattern. Each one is work a business already pays someone to do, has a right answer you can check, and lives in an industry running on old software that hates the process. That combination is the tell. When you find it, you've found a business.

I'll keep adding these types of opportunities on Ideabrowser.com (literally free to sign up for validated ideas)

What makes it defensible

The first question every founder and every investor asks is: why won't the customer just do this themselves with ChatGPT? The answer is that they don't want a tool, they want it done, and they want someone on the hook when it's wrong. A clinic isn't going to have its office manager paste claims into a chatbot and hope. They want a firm that knows what a denied claim looks like, has caught that mistake three hundred times, and will fix it if it slips through. Your rulebook and your accountability are the product. The model is just the engine, and everyone has the same engine.

That's the real moat. When you sell the finished work, you own the mistakes. A software company ships a bug and patches it. A firm that files a wrong claim has a real problem, and the customer is right to be upset. So the moat is not the model. It's the rulebook, the hundreds of edge cases you caught and wrote down in one narrow niche. A competitor can download the model in an afternoon. They can't download three hundred jobs' worth of your mistakes.

That's also why the review layer matters. It's what lets you own the mistakes without them owning you, and it's the thing that shrinks as the rulebook grows.

The one thing to get right

Most people building "AI agencies" point the AI at production and leave everything else the way it was. That gives you a slightly cheaper agency. The reason agencies were never great businesses was the overhead around the work, scoping, checking, managing, selling, all of which grew with every client.

An AI-native service collapses that overhead on purpose. Intake is a form. Scope is a menu. Quality is a rulebook. Account management is a dashboard. Do that and the business behaves like software. Skip it and you've just built a faster version of the thing that never worked. And the coolest part is if you ever want to sell that business, your margins might be 60-80%. So you can command a life changing exit on something like a 6-12X EBITDA exit.

So, that's basically the whole thing. Pick a unit. Build the intake, the engine, the rulebook, the review layer, and the delivery. Price against the human, per unit or on a retainer. Start by doing the work by hand for five customers, write down every mistake, then productize. The $120,000 that was locked behind a person is open now! Go find yours.

Point of this article was just to get your creative juices flowing.

AI-native services are here, and now's the time.