The accountant, not the software.

Greg Isenberg says the best business to start right now is an AI-native service: sell the closed books, not the bookkeeping tool, and let agents do most of the work. He is right, and his list of what one is made of is a list of what we have been building. Here is where his guide and AI3 meet, piece by piece, and where we still fall short.

Joe Charlesworth, founder, AI3 · 27 September 2026 · 10 min read

The argument

Greg Isenberg published a long guide this week to what he calls AI-native services, and it is the clearest statement I have read of the thing we are building for. His example fits in a sentence. A business pays about $10,000 a year for QuickBooks and about $120,000 a year for the accountant who uses it. Software companies spent twenty years fighting over the $10,000, because that was the part that scaled. The $120,000 was locked behind a person. Now, he argues, agents can do most of what the accountant does, so you can sell the closed books instead of the tool, at software margins.

That is an AI-native service: the customer buys finished work, agents do most of it, a small team handles what still needs a person, and the delivery runs on software instead of headcount. He puts the pool at about $4.6 trillion a year of US business spending on services, roughly six times what goes on software, and names the companies that moved first: Harvey in legal, EvenUp selling demand letters at around $500 each, Kick doing small-business bookkeeping for $300 to $500 a month at gross margins above 70 percent. All figures are his. The whole guide is worth the fifteen minutes he asks for.

What I want to do here is take his list of what one of these businesses is made of, eight pieces, and hold it against what is on ai3.co today. Partly because the fit is close, and partly because the places where it is not close are the work.

Two readers

The guide is written for a founder who will start one AI-native service: pick a niche, do the first five jobs by hand, write down every mistake, productise. We have a second reader in mind as well, and she matters more to us. She is the owner of the business that will hire the service. The plumber, the dentist, the hotel in the Mission. She does not want to build a firm. She wants the phone answered, the books closed and the letters sent, and she wants someone on the hook when it goes wrong.

AI3 is built for both. The founder runs the service as an organisation on AI3, with its workers, its books and its door for agents to pay through. The owner hires a worker from the directory and gets the finished work in her dashboard. The layer under the two of them is the same: the operating system both run on, and the market where they find each other. Isenberg's eight pieces describe what that layer has to provide.

The eight pieces, and where each one lives

1. The unit. He says this is the most important decision: sell one clearly defined thing with a finish line, per claim, per filing, per month of books, never per hour. On AI3 a worker's job is defined the same way. Ivy answers the phone and books the call. James writes the week's posts. The unit of work is a task with a deliverable the owner reads and judges, and the unit of money is what a business sells: ask, book, hold, quote, each with a price a machine can read.

2. The intake. His rule is that if you cannot define intake as a form, your unit is not clear enough. Our hire form asks one question per worker. Ivy asks for your hours, services and prices and how you like bookings taken. That single answer becomes her day-one task: the call script she will answer with, every fact listed with where it came from, and every fact she still needs from you. Three screens, no model to choose, no budget to set.

3. The engine. The model plus your instructions and your context. Ours routes each job across more than one model and measures what each one actually cost and whether the result was accepted, so a worker that is cheaper on one model and better on another is run on the right one. The owner never sees this. She sees the result.

4. The rulebook. Isenberg says this is the real product and almost everyone underrates it: the written list of what correct means in your niche, built one caught mistake at a time. He is right, and this is the piece where we are furthest behind. Every worker on AI3 runs under a written set of rules: never invent a fact, a price or a promise; name your sources; say exactly what was out of reach rather than pretending it was done; put the result first because a busy owner reads five lines. Each role has a playbook on top of that. But those playbooks are written by us, per role. The rulebook he describes is written by the service, per niche, out of hundreds of jobs' worth of mistakes. Ours does not yet grow on its own from what an owner rejects. More on that below.

5. The review layer. Where a human still looks. He says to decide, per unit, what ships automatically and what a person checks first, and that anything with money or a customer's reputation attached gets a human eye. This is the piece we have thought about hardest, because it is where the money is. On AI3 nothing is sent to a customer, published or paid without the owner's yes; drafts are fine, sending is not. Spending is gated by authority the owner set, not by the worker's judgement. And when a worker delivers, the owner's verdict is recorded against it, and a disputed verdict has somewhere to go.

6. The delivery. A dashboard the customer logs into, replacing the account manager. Ours is the organisation's home: the tasks page where each deliverable lands, the customer record the receptionist writes calls and bookings into, the calendar she books against, and the books where the money shows up. The owner should never have to ask where a job is, and on AI3 she does not.

7. The pricing. Per unit or a flat retainer, never hours, and priced against the human alternative rather than your costs. Here we differ, and on purpose. A service built on AI3 prices its unit however its founder likes, and his advice is good. AI3 itself charges the organisation: $50 a month, workers free to add, and what a worker uses billed at the provider's price with no markup. We sell the organisation, not the inference, so a founder's margin is his to keep.

8. The distribution. Cold outbound to the person who writes the cheque, plus one lead magnet: do the first job free. Our free first job is the readiness report. Enter a website, and a business gets a report on what its customers say and whether an agent can find it, read its price and pay it, with the first three fixes and a worker beside each one. That report is the lead magnet for every service listed on AI3, and the directory is the channel to the buyer.

What AI3 is

Isenberg's guide stops at the edge of the firm. It assumes, reasonably, that the founder has a bank account, a bookkeeper, a contract and a way to be paid, and it leaves the founder to build the eight pieces from scratch for each niche. We read the list differently. The eight pieces are not a firm. They are a template. A note-review service and a lease-abstraction service are the same eight parts with different content: a different unit, a different intake form, a different rulebook, a different price. Everything else repeats, and what repeats should be built once.

That is what AI3 is: the operating system a business runs on when most of its work is done by agents, and the market where those businesses find their customers and each other. The template that turns a rulebook into a running service. The books each one keeps. The door through which people and agents pay it. The record of its customers and its calendar. The venue that answers when a job goes wrong. Not bookkeeping for agents, though it has books. The whole business, and the trade between businesses.

If he is right about the size of the pool, and we think he is, then the businesses that run this way will outcompete the ones that do not, in every niche he lists and thousands he does not. We want every one of them running on AI3. That is a large claim, so here is exactly where it stands.

The template is the work

So the next thing we build is the template, and we build it rather than running services ourselves. A founder who has done Isenberg's first five jobs by hand should be able to open an organisation on AI3 and fill in eight fields, and have a running service by the end of the afternoon. In the order the pieces unblock each other:

• The unit and its price. One clearly defined thing, per unit or on a monthly retainer, and every completed unit a line in the organisation's books. Not hours, ever.
• The intake as a form. The fields and the files a customer supplies for one unit, and a batch of files becoming a batch of units.
• The rulebook the organisation owns. Every rejected deliverable carries a reason. The owner promotes a reason to a rule with one click. The engine reads the rules before every unit. The list is private to the organisation, leaves with it, and is the thing a competitor cannot download.
• The review setting, per unit. Ship automatically when the rules pass, hold for a person when money, legal exposure or a customer's reputation is attached, and sample what went out unchecked. This is the switch that turns a cheaper agency into a business with software margins, and it is off until the owner turns it on.
• The delivery. A status page per customer of the service, so nobody has to ask where a job is.
• The guarantee. One sentence on the listing saying what happens when a unit is wrong: redone, refunded, or taken to Recourse. His answer to "why not just use ChatGPT" is that customers want someone on the hook, and the listing should say who.

Then the listing goes on the directory beside our own workers, and the readiness report, which is already the free first job for a local business, points at the services that fix what it finds.

Why this is the same thesis

Two posts ago we wrote that the unit of the web is becoming the request, and that agents will pay for what they use. Last week we wrote about the agent that shops. Isenberg's guide is the other side of the same change: if agents can do the accountant's work, the $120,000 is no longer locked behind a person, and thousands of small firms will form to sell that work. Each of them needs the eight pieces. Each of them also needs books, a door, a record of its customers, a market and someone on the hook. The eight are the firm. The rest is the operating system, and it should not have to be built eight thousand times.

He closes by telling you to go find your $120,000. We would add one line. When you find it, build it on AI3, where the books are already kept, the door is already open to agents, the customers are already looking, and the mistakes become rules.

What to do about it

• If you are starting an AI-native service: open an organisation and add the workers your unit needs. The service template is next, and we want the first services on it to be real ones. Write to us if yours should be one.
• If you run a business: connect your business. It is the free first job. The fixes on it each have a worker beside them.
• If you build agents: the API and the MCP server put the books, the customer record and the calendar behind your own agent, whatever it runs on.

More from the blog · RSS · Questions and corrections: contact. A correction that changes a fact is noted above with its date.