There is a version of the AI business that gets written about constantly - the founder, the model, the exponential curve - and a version that quietly pays people's mortgages. This is about the second one.
The premise is unglamorous. Companies now have a list of tasks they know AI could handle and nobody internally who will do it. They do not want a model, a licence or a workshop. They want the thing finished: product photos shot, ads produced, tickets triaged, contracts summarised, listings written. They have a budget for that. Selling them the finished work is a normal service business, and it is the most reliable money in the entire AI economy right now.
Start with the arithmetic, because it is the whole argument
Ten clients on a $8,000 monthly retainer is $960,000 a year.
Sit with that for a second, because it is the least impressive sentence in this article and the most important one. It is not a moonshot. It is ten relationships. Most agencies that size are four or five people working normal hours, and thousands of them exist in every service category that has ever been invented.
Now the part the thread-writers leave out: that is revenue, not profit. A realistic shape at that scale looks roughly like this.
Revenue $960,000. Salaries for four people, including yours, around $420,000. Model and tooling costs, which for image and video work are the real line item, $60,000 to $120,000. Sales, software, accounting, legal, insurance, around $80,000. Left over: somewhere near $340,000 to $400,000 before tax.
So the million in profit is roughly a three-year outcome, not a one-year one. That is still an excellent business. It is simply not the business described by anyone selling you a course about it.
What the client is actually buying
The single most common failure in this category is selling the technology.
No procurement department has a line item for "AI". They have line items for photography, for advertising production, for customer support, for content, for data entry. You are not creating a new budget - you are taking an existing one at a better price, or removing a bottleneck that has been annoying someone for two years.
This changes how you describe the work. "We use generative AI to produce product imagery" is a sentence that invites scepticism and a comparison to free tools. "Two hundred product shots on white, styled and retouched, delivered in four days, $12 each" is a purchase order. The second one never mentions the method, and the method is not what they are paying for.
It also changes who you talk to. The person who cares about AI is the CTO. The person with the budget for photography is the head of e-commerce, and they have never thought about you at all.
Six categories that are already working, and who is doing them
These are not hypotheticals. Each of these has real companies operating in it today, and looking at how they package the work is more instructive than any strategy advice.
**Product imagery.** E-commerce brands need thousands of images and re-shoot them every season. Studio photography runs from tens to hundreds of dollars per shot with a lead time in weeks. Tools like Photoroom, Flair and Pebblely built products around this exact gap, and behind them sits a much larger population of small agencies quietly delivering catalogue imagery to Shopify and Amazon sellers. The pitch is never the model - it is cost per image and turnaround in days.
**Headshots and portraits.** Aragon, HeadshotPro and Secta Labs sell a set of professional-looking portraits for something in the region of thirty to sixty dollars, against a studio session costing several hundred. The interesting version is not consumer self-serve; it is the corporate one, where a company with 300 employees wants a consistent set of team photos across four offices and has no way to organise that physically.
**Video and UGC advertising.** Direct-to-consumer brands burn through ad creative because performance decays with repetition. They need dozens of variants a month, and each one traditionally means a person, a camera and a script. Arcads, Creatify and HeyGen built products around generating that volume. Agencies sell the same thing as a managed service: a monthly quota of ad variants, with the brand only seeing finished creative.
**Voice, dubbing and localisation.** A course, a support library or a product catalogue in one language, delivered in nine. This used to mean voice actors and a studio per language. The economics collapsed, and the buyers - training companies, publishers, e-learning platforms - are the least AI-native audience on this list, which is precisely why the work is available.
**Written work at volume.** Product descriptions, category pages, technical documentation, translation. This is the most commoditised category and the one with the most competition, so the money is in the least fun corners: regulated industries, technical domains, anything where the writer has to understand what they are describing.
**Agents that do a job.** The newest and highest-value category. Sierra, Decagon and Intercom's Fin sell customer support that resolves tickets rather than software that helps humans resolve them - and several of them price per resolution, not per seat. 11x and Artisan sell outbound sales agents described explicitly as digital workers. Harvey does legal work. The pattern is identical everywhere: the pricing is attached to an outcome the buyer already understands and already pays for.
Notice what every one of these has in common. The buyer is not asked to learn anything, adopt anything, or change how they work. They approve something and receive something.
How to price it
Three models, in ascending order of how much you will eventually earn.
**Per asset.** Twelve dollars an image, ninety dollars a video, four dollars a product description. Easy to sell because it is trivially comparable to what they pay now. It is also a treadmill: your revenue stops the moment you stop producing, and your price is under permanent downward pressure from whoever is willing to do it cheaper.
**Retainer.** A fixed monthly fee for a defined quota - "up to 300 images a month, four-day turnaround, unlimited revisions within scope". This is where the $8,000 figure comes from and it is the model to aim for. Revenue becomes predictable, the client stops re-deciding every month, and your margin improves as you get faster while the price stays put.
**Per outcome.** You are paid per resolved ticket, per booked meeting, per qualified lead. Highest ceiling and highest risk, because you have taken on the client's uncertainty as well as your own. Do not start here. Move here after a year, with one client you know well, and only in a category where the outcome is unambiguously measurable.
The mistake almost everyone makes is pricing against their own cost. Your cost is nearly zero and falling, which is exactly why it is irrelevant. Price against what they pay today for the same result - the studio invoice, the agency retainer, the salary of the person currently doing it.
The step-by-step
**Pick one output and one industry.** Not "AI services". Product photography for furniture brands. Support agents for B2B software. Localisation for e-learning publishers. Narrowness is not a limitation, it is the entire sales advantage: you can say "I do this for companies exactly like yours" and that sentence closes deals that no amount of general capability will.
**Build the pipeline before you have a client.** Whatever the output is, produce it end to end at least twenty times on your own. You are looking for where it breaks - the products the model gets wrong, the accents it mangles, the edge case that needs a human. Those failure points are your actual product knowledge and they are what separates you from the client's intern with the same subscription.
**Produce work for five real companies who did not ask.** Take their actual products, their actual catalogue, and do the job. Then send it to the person whose budget it is with no pitch attached - here is what your category page looks like with these images, here is the cost, here is the turnaround. This converts at a rate that no cold email ever will, because you have removed every step between them and seeing the result.
**Price the first one low and say why.** Explicitly a founding-client rate in exchange for a case study and a reference. Not a discount you will regret - a trade you named out loud.
**Deliver obsessively for ninety days.** The first client is not a revenue source, it is the proof that makes the next four possible. Overdeliver in ways that cost you time rather than money, since time is what you have.
**Then raise your price and do it again.** Every new client should be more expensive than the last until you find resistance. You will find it much later than you expect.
Why clients leave, and what stops them
The churn in this business has one dominant cause and it is not quality. It is the client realising they could do it themselves.
That realisation is usually wrong - they have neither the pipeline nor the failure knowledge - but it does not need to be right to cost you the account. The defence is not secrecy. It is being embedded in something they do not want to rebuild: their brand guidelines encoded into your process, their product data connected to your system, their approval workflow running through your tooling, three years of their assets organised in a way they now depend on.
Put differently, the retainer survives on switching cost, not on skill. Deliberately accumulate it.
Three ways this goes wrong
**Selling the tool instead of the work.** The moment a client is thinking about which model you use, you have lost the frame. They will conclude they can buy it directly, and they are right.
**Competing on price in a commodity category.** Written content is already there, and generic imagery is heading there. If your only differentiator is being cheap, your margin has an expiry date visible from here.
**Building on one provider's feature.** If your whole service is a thin layer over one capability, assume it gets absorbed into the provider's product within a year and priced at zero. Everything durable in this list has human judgement, industry knowledge or integration work between the model and the invoice.
Why we think this is the highest-probability path
We build AI trading agents, which is a far more fashionable business than the one described above and a considerably worse way to reach a million dollars. It is worth saying why.
In a service business, someone hands you an agreed amount of money for a defined thing. You are paid for effort and judgement. In a trading business, you are paid only for being right about an uncertain future, and the market has no obligation to pay you at all. Turning fifty thousand dollars into a million requires multiplying capital twentyfold, which at twenty per cent a year - a return that would place you above most professional funds - takes over sixteen years.
We also have a running demonstration of how hard the second thing is. Our published forecasts state a 50% range; across 56 resolved forecasts the outcome landed inside 47 times. Eighty-four per cent, against a target of fifty. That is a failure, not a success - a range that catches almost everything carries no information - and because every outcome kept landing inside, it produced no symptom for weeks. We publish that number on the same page as the good ones.
If a system that timestamps every claim before the event and scores itself in public can be that wrong about its own uncertainty, the service business starts looking considerably more attractive. It is boring because it works often enough to have become boring.
The one-line version
Find a job businesses already pay for, do it with AI at a fraction of the cost and a fraction of the time, charge against their old invoice rather than your new cost, and never once mention the model.
Educational content - not financial advice.