How to Monetize Your Build

FIELD MANUAL 03 · 16 MIN

The first two Field Manuals were about why the ground is open and how to take it. This one is about what decides whether you are still standing in a year: how the money reaches you. An agent business can get paid in eight ways, one question that decides which fits, and a margin problem no one has had to think about for twenty years.

Software has been sold by the seat for two decades because the seat was a good proxy for value. One human, one license, one login, and the more people a company hired, the more it paid, which meant the vendor grew when the customer grew, and nobody had to argue about it. The proxy worked because humans were the bottleneck. A support agent could handle so many tickets an hour, a sales representative could work so many accounts, and the seat count was a reasonable stand-in for the amount of work getting done.

An agent breaks that proxy in both directions. For example, one person can run 50 agents, which means the seat count understates the work by a factor of 50. If one agent can serve 10,000 customers, then the seat count isn’t measuring anything. Charge by the seat in that world, and you either leave most of your value uncollected or price a single seat so high the buyer laughs in your face.

What the vendors closest to the problem actually did is more interesting than abandoning the seat, and it’s more valuable to review their pricing pages rather than the commentary about them. Intercom still sells seats at $29, $85, and $132 per seat a month, because the people in the inbox are still people. Alongside them, it prices its Fin agent at 99 cents per outcome, so when the agent handles the job alone, the customer pays for the job rather than for a chair, and Fin can be bought on its own with no seat at all for a company that already has a help desk. It also sells Copilot, which is the agent helping a human rather than replacing one, at $29 per agent per month, priced per seat again because there is a human in the loop whose capacity is the thing being expanded. Microsoft does the same thing in the same catalog: Copilot at $30 per user a month, and the agent-building product next to it in capacity packs at $200 a month for 25,000 credits. Salesforce started Agentforce at $2 per conversation regardless of complexity, replaced that with credits priced at about ten cents an action, so a typical exchange of three to six actions costs thirty to sixty cents, and still offers an unmetered per-user license for buyers who want a predictable bill.

The rule falling out of those pages is simple enough: price per seat where a human is still the constraint. Price per unit of work where the agent is. Most of these companies sell both to the same customer, because both are true at once in the same account, and a builder who thinks the choice is either/or is leaving one of the two on the table.

The pricing shift is a mix, not a replacement. ICONIQ surveyed roughly 300 software executives for its 2026 state of AI report, and the share pricing on consumption had gone from 19% to 35% while the share pricing on outcomes had gone from 2% to 18%, with subscription still the most common model at 58% and 37% of the companies surveyed saying they intended to change how they price within the year. The numbers add to more than a hundred because companies run several models side by side, which is the finding, not an artifact.

So the first question is not what to charge. It is what to charge for, and how many different things you are charging for at once.

Everything below is used today by a live business. Most builds use two of them, and the second one is usually the one that saves the company.

Per unit of work. The agent does a discrete thing, and the price is attached to the thing: a call, a run, a lookup, a page rendered, a document read. This is the model the pay per call marketplaces run, where a WHOIS lookup goes for a tenth of a cent and a dossier for fifty-five cents, and it is the only model that works when your customer is itself an agent, because an agent will not choose to fill in a signup form or agree to a monthly plan. It is an honest model. It scales without negotiation, but its weakness is that revenue is invisible until volume arrives.

Per outcome. You charge only when the job is finished to the customer's satisfaction, and the customer pays nothing for failed attempts. Intercom's outcome price, starting at 99 cents, is the clearest published example, and the key point is that it travels without a seat contract, since a company with its own help desk can buy the outcomes and nothing else. Sierra, Zendesk, and Decagon all describe versions of the same model without publishing a rate. Buyers love it because it moves the risk onto you, which is exactly why you should read the margin section below before you offer it.

Per action. The compromise between the two above, where the customer pays for each step the agent takes rather than for the whole result. Salesforce landed here after starting somewhere else, and the reason is instructive: a flat price per conversation is easy to explain and impossible to defend when one conversation costs you fifty times what another does.

A share of what you recover. If the agent brings the customer money that was already theirs, take a cut of it. This is the oldest pricing model in the world, and it needs no explanation to a customer, because they only pay out of money they did not have this morning. It is also the model with the most obvious incumbent to undercut, since AirHelp charges 35% of a flight compensation claim plus a further 15% if the case goes legal, medical billing advocates take 15 to 35% of what they recover, and collection agencies routinely keep a quarter to half of an old debt. An agent doing the same work for a tenth of that is an easy sell.

A flat fee for the job. One price, one job, no percentage. Claimable charges $39.95 to write and file a health insurance appeal, on terrain where the advocate's cut on a $1,000 denial would have been $150 to $350. Flat pricing wins whenever the customer is frightened of the percentage, and it is the right answer when your cost per job is stable and the value varies wildly, because a percentage on a large claim will get you accused of profiteering by the same customer who was delighted at the small one.

A subscription for standing watch. The agent is not doing one job. It is watching for the moment a job needs doing, which might be a price drop, a renewal date, a filing deadline, a rate change, or a competitor's move. Nobody wants to think about the watching, which is precisely why they will pay a small amount every month to stop thinking about it. This is the most durable revenue on the list because it doesn’t depend on the customer remembering you.

A take rate on a market you run. If you are not the agent but the place agents and customers meet, you can charge a percentage of what passes through. Apple takes 30% and 15% under its small business program; Google Play takes 30% above the first million dollars and 15% below it; Etsy takes 6.5% plus a listing fee. Each of those numbers is a reason somebody is building something to route around them, so if you choose this model, choose your number knowing it is now a target painted on your own back.

Selling to the agents rather than to the humans. The least crowded position on the list. Somebody has to supply the agents with data, tools, verification, compute, and answers, and the agent will buy without a demo, without a procurement cycle, and without a quarterly business review, provided the price is in the request, and the payment travels in the response. If you already run an API, a dataset, or a scraper, this is not a new business. It is a second doorway into the one you have.

Choosing among those eight is really choosing your unit, which is the thing you count and charge for, and the unit determines everything downstream: how you sell, what you have to measure, when you get paid, and who carries the risk if the work goes badly. Three tests will get you to the right one.

The first test is whether the customer can see the unit. A resolution, a recovered claim, a booked load, and a filed appeal are all things the customer already counts, which means you don’t have to teach them anything before you can send an invoice. Tokens, compute minutes, and API calls are things you count, and a customer who has to learn your unit before they can approve a budget will take three months (or longer) to buy.

The second test is whether the unit moves with value. If a customer gets 10x the benefit, does your revenue go up? A monthly subscription fails this test, which is why so many software companies watched a customer expand tenfold and collected nothing extra. A per-outcome price passes it automatically.

The third test is whether your cost per unit is stable. This is the one builders skip, and it is the one that kills them, because the moment your price is fixed per unit and your cost per unit is not, you have written a blank check against your own margin (We are looking at you, Amazon Bedrock). That is the subject of the next section, and it is worth learning the words for, because the words are how you will explain the business to an investor, a bank, or the first serious hire who asks whether this thing works.

Unit economics means the revenue and the cost of a single one of whatever you sell, stripped of everything else the company spends. Cost to serve is everything it costs you to deliver that one unit after the sale, e.g., the model calls, the storage, the payment fee, the share of support. Price minus cost to serve is contribution margin, the money each sale contributes toward the fixed costs of the business, the salaries and the rent and the compliance work, before any of those are counted. The same figure, written as a percentage of revenue across everything you sell, is your gross margin, and it is the first number an investor will often ask for because it tells them whether growth makes you richer or poorer. Take rate is the share of a transaction you keep when the money is really somebody else's, which is what a marketplace charges. Customer acquisition cost is what you spend to win one paying customer, and payback period is how many months of that customer's contribution margin it takes to earn that spending back. A business with a twelve month payback can grow on its own cash. A business with a thirty month payback is a fundraising exercise wearing a product.

For twenty years, software had almost no cost of goods sold. You wrote it once, the marginal cost of the next customer was a rounding error, and gross margins in the high seventies and eighties were the normal expectation. That is not the business you are in. Every run of your agent buys tokens from a model provider at a published price, and those tokens are your cost of goods, as real as steel is to a bridge builder.

The scale of the difference is visible in the data. The same ICONIQ survey put the average gross margin at an AI application company at 52% in 2026, up from 45% the year before and 41% the year before that. Improving, and still nowhere near what a software investor was trained to expect. If you price like a software company and pay like a manufacturer, you will discover the gap at the worst possible moment: when volume finally arrives.

Two things follow, and they point in opposite directions, which is why this needs thinking about rather than a rule.

First, your costs are falling faster than almost any input cost in commercial history. Andreessen Horowitz's analysis of what it called LLMflation found that, for a model of equivalent performance, the cost falls by about 10x a year, and that intelligence at the level of GPT-4 became roughly 62 times cheaper in the twenty months after March 2023. The published prices bear it out. OpenAI's cheapest current tier runs at 20 cents per million input tokens. Anthropic's smallest current model is a dollar, compared with $60 per million for comparable capability in 2021. If your build is unprofitable today because the model is expensive, and everything else about it works, you are not running a failing business. You are running a business whose main input gets cheaper every quarter without you doing anything, which is a position most founders in history would have traded a limb for.

The second is that falling costs are only your gain if you did not hand them away in your pricing. Price your product as a markup on your model spend, and every price cut from your supplier is a price cut you must pass to your customer, because you told them what the thing costs. Price it as a fraction of the toll the customer used to pay, and every cut goes to your margin instead. This is the entire difference between a business that gets better as the technology improves and a business that gets commoditized by it, and it is decided by a sentence on your pricing page written before you had any customers.

Which brings back the warning about outcome pricing. When you charge $0.99 for a resolution, and your cost per attempt is variable, you agree to absorb every hard case, every retry, every customer whose problem takes forty exchanges instead of four. That is a fine trade if you know your cost distribution and have priced the average with room to spare, and it is a catastrophe if you are guessing, because the customers with the worst problems will find you first and tell their friends. Before offering to be paid only for success, measure what failure costs you, then price the success high enough to carry it.

Consumer pricing has a hard arithmetic problem. You cannot spend $60 acquiring a customer for a $40 product. Customer acquisition cost quietly kills most builds at this price point, and there are only three ways out:

Claimable is the clearest current example of the third. The patient pays $39.95 to have a health insurance denial appealed, or pays nothing, because the company now also runs the same service as a platform for drug manufacturers, health systems, and patient access programs, and said in April 2026 that it was live across ten programs covering treatments used by more than a million patients, with over 70% of referred patients completing and submitting an appeal. Look at whose interests line up. A manufacturer earns thousands of dollars from a patient who stays on a treatment and nothing from a patient whose claim was denied and who gave up, so the manufacturer will happily pay for the appeal that the patient could not be bothered to file. The builder sits between them and collects from whichever side values the outcome more, and the patient, who has the least money and the most friction, pays nothing at all.

The pattern generalizes further than people expect. If your agent saves a business money, sell it to the business rather than to the worker whose time it saves. If your agent makes a platform's customers more successful, the platform may pay you rather than have you charge its users. If your agent reduces somebody's losses, their risk department has a budget, and their customers do not. Ask who has the most to gain from the job being done, then ask why that party isn't paying. The answer is often that nobody has offered them the chance.

That question is worth more time than the pricing page. Most builders spend a week deciding between $29 and $39 a month and five minutes deciding who to send the invoice to, and the second decision is worth 10x the first.

There is a floor and a ceiling on what you can charge, and they are not where new builders think they are. The floor is your cost to serve, which, as we have said, is falling. The ceiling is what the customer pays today for the same job to be done by the incumbent, which is the toll, and everything between the two is available to you.

Most builders price just above their costs, because that is the number they can calculate and it feels safe. It is not safe. It leaves nearly all of the value with the customer. It makes the business fragile the moment a model provider raises a price, and it teaches the market that what you do is cheap. It also removes the money you need to fight with, since the guerrilla who wins is the one still shipping in eighteen months.

Price down from the toll instead. If a billing advocate takes 35% of a $1,000 recovery, the customer's reference point is $350, and $39.95 reads as free even though it is a hundred times your cost. If an incumbent charges 30% of a developer's revenue, you can charge 5% and be both wildly cheaper and wildly profitable. If the toll is a 30-cent minimum on a card, anything you can charge under 30 cents is not a discount, but a category the incumbent cannot enter at all.

The number to write on your pricing page is the one that is obviously, immediately, arithmetically better than the toll, and not one cent lower. Being ten times cheaper wins the customer. Being a hundred times cheaper wins the same customer and starves you.

Choosing a price is half the problem. The other half is getting the money to arrive, and the rail you choose decides which prices are even possible.

Cards are the default, and they have a floor. Stripe's published rate is 2.9% plus 30 cents per successful charge, so a five-cent sale costs 30 cents to collect and a one-cent sale costs thirty times what it earns. Any business whose natural unit price is under a dollar cannot exist on that rail, which isn't a complaint about Stripe, since the 30 cents pays for a dispute process, a fraud system, and a network of banks designed for a shopper in a shop. It simply means that if your unit is small, the card is not your rail.

Marketplaces are a rail too, and an expensive one. If your money arrives through an app store, you pay 15% or 30% before anything else happens, which you have to price in from day one rather than discover later.

The rail built for small units is settlement in stablecoins over a protocol that puts the price in the response, and the economics are not close. Coinbase's x402 facilitator settles the first thousand payments a month free, charges a tenth of a cent each after that, and makes verification free. At that price, a one-cent sale is a real business and a tenth-of-a-cent sale is a real business, which is the whole reason the pay-per-call market exists at all.

There is a catch, and it matters most. Settlement in stablecoins is final. USDC's merchant documentation says transactions are final and irreversible, with no chargebacks. For a merchant tired of losing disputes, that sounds like a gift, and for about a year it will feel like one. What it actually means is that every check you would have relied on a card network to perform after the fact now has to happen before the money moves, because there is no after. Verifying who authorized the spend, what they were authorized to spend it on, and keeping a record that stands up when somebody later says it wasn't them is now your job rather than your processor's. That is the subject of the next Field Manual, and it's why we started FLINT.

Four things look like a monetization strategy and aren't, because each one has a season when it feels clever.

Free forever with no payer: There is nothing wrong with charging nothing, and Counterforce Health has said it will always be free for individuals, but it is funded by grants, and it says so. If your plan is free, with no third party paying and no clear moment when that changes, you don’t have a pricing model. You have a hobby with users, and users are more expensive than no users.

Growth bought below cost: Selling a service for less than it costs you to deliver in order to show a chart to an investor is a strategy with exactly one exit, and it is a fundraise. It can work if you achieve escape velocity. It is also the single most common way a build that had a real business inside it dies, because the founder never found out what customers would actually have paid.

Volume without a unit: A dashboard of impressive activity that nobody pays for is not traction. Money moving is the test the newsletter applies to every build it covers, and that’s not squeamishness about pre-revenue companies. In our experience, until somebody pays, you've learned nothing.

Waiting to monetize until later: Sometimes correct, usually not, and it is worth being honest about which one you are doing. Charging early doesn't slow you down. It tells you within a week whether the value you believe in is something anyone will pay for, and that answer is worth more than a year of growth.

The eight models above are all in use by somebody live right now, most of them by people with no funding and no staff. That is the part worth holding on to. Lovable told TechCrunch in March that it had reached $400 million in annual revenue with 146 employees, and Carta's data has solo-founded startups rising from 23.7% of new companies in 2019 to 36.3% by the first half of 2025. The tools that make one person productive enough to build an agent business are the same tools that make one person's revenue per head look like a small company's, and the constraint that used to require a sales organization, e.g., a billing department, and a pricing committee is now a decision you make on a Sunday afternoon.

Be deliberate. Decide what you count, decide who you send the invoice to, decide what the incumbent charges for the same job, and then price down from that rather than up from your bill. The Special Agent is written for the guerrilla, the disruptor who knows single actors can wield disproportionate leverage. We uncover what they are building, evaluate the economic terrain, and turn isolated actions into concentrated force.

The author of this Field Manual tried to keep all information timely, relevant, and accurate. If prices change in the future for the companies mentioned, we will try to keep this piece updated. Further, the author would like to point out that his MBA is from a state school and, as such, may lack the professionalism demonstrated by the Stanford and Harvard types. Direct all complaints to [email protected].

SOURCES

Intercom pricing · Salesforce on Agentforce Flex Credits · Sierra on outcome based pricing · Zendesk pricing · Decagon on resolution based pricing · Microsoft Copilot Studio pricing · ICONIQ 2026 State of AI, as reported · a16z, LLMflation · Anthropic API pricing · OpenAI API pricing · OpenAI on GPT-4o mini pricing · Stripe pricing · Coinbase x402 facilitator fees · Apple App Store Small Business Program · Google Play service fees · Etsy fees · USDC merchant guide on finality · AirHelp fees · Medical billing advocate fees · Claimable enterprise platform announcement · Claimable · Counterforce Health · TechCrunch on Lovable's revenue · Carta solo founders report