A marketing agency is almost pure labor. Automating the production work is where AI reaches an industry first.
A marketing agency sells work its clients cannot do themselves. It runs the sites, the emails, the posts, and the campaigns for a roster of businesses. It pays for that work almost entirely in people. Payroll is half to two-thirds of what an agency takes in [1], the largest cost it carries. After it, the average agency keeps about 13 cents on the dollar [2]. Giant Context is what I am building to change that number. It is autonomous marketing. You hand it a client's documents. It produces and runs that client's marketing on its own.
This is a theory. I have no customers and no evidence the market wants it. What follows is the reasoning that made it worth building, set down so it can be checked. Everything I write after this is the record of testing it.
kept per dollar — the average agency margin
billable utilization in 2024, a five-year low
loaded cost of a five-person team, per year
An agency grows by adding clients, which means adding people. The work is billed against hours. A producer has only so many. Utilization, the share of an agency's paid time that is actually billable, ran under 70% across the industry in 2024, a five-year low [3]. One account handler carries fewer than ten clients before the service starts to slip [3]. Every new client on the roster pulls in more salary. The margin that was thin to begin with gets thinner as the shop scales. The largest agencies, the ones with the most clients, keep the least of all [2].
Growth is the number one thing agencies say they need [4]. It is also the thing the model punishes. To take on more work, an agency hires. The hire is the cost. The cost is a person, because the tools are not what produces the marketing. A person does.
Now put the production into software. Most of what a producer makes, the posts, the pages, the emails, the ad copy, is exactly the kind of work these systems are good at now. Automate the bulk of it and the cost that scales with the roster comes off the books.
A five-person production team, two writers, a social manager, a designer and an SEO specialist, costs about $460,000 a year fully loaded [5]. Automate 60 to 70% of what they make and the agency saves $280,000 to $325,000 a year.* A fifty-client shop with a larger team saves $450,000 to $520,000 [5].
| Production role | Annual cost |
|---|---|
| Writer | $78,000 |
| Writer | $78,000 |
| Social media manager | $70,000 |
| Graphic designer | $61,300 |
| SEO specialist | $70,000 |
| Base total | $357,300 |
| Fully loaded (about 1.3x) | ~$464,000 / year |
Salaries are from Indeed, the US Bureau of Labor Statistics, Career.com and Webflow [5]. Labor is half to two-thirds of what an agency spends [1]. There are about 41,000 marketing agencies in the United States, 88% of them under 50 people [6]. The same cost structure sits under almost all of them.
* The savings figure assumes a standard fully-loaded-cost multiplier and that most production work automates. Both are conservative. Even at half the automation, a small agency still saves six figures a year.
Agencies are not waiting to be sold on AI. Better than nine in ten are already using it or trying it [7]. The question is what they are buying. On one side are the platforms that run an agency's many clients, the sub-accounts and the billing and the dashboards. GoHighLevel and Vendasta lead there. The AI in them is an assistant on top of a fulfillment shell. On the other side are the content tools, Jasper and Copy.ai, that write well but cannot keep one client's work separate from another's. Both leave a person in the seat, deciding what to make and approving each piece. The agency still bills the hour. The hour still has a person in it.
What none of them does is the whole job at once. That job is one system that keeps every client separate, works from each client's own facts, and produces and publishes the marketing without a person driving it. The strong-AI tools are single-client. The multi-client tools have weak AI. No one has put the two together. The best-funded newcomer, Peec AI, raised $21 million late last year to help brands surface in AI search, not to run an agency's roster [18]. No one has funded the intersection either.
The timing is not subtle. Late last year the two largest agency holding companies merged in a deal worth more than thirteen billion dollars, cut four thousand jobs, and named AI as the reason [10]. The industry press is already calling the billable hour dead. The agencies that move to the new model instead of defending the old one are the ones that come through the repricing.
| What it is | Runs a whole roster | Produces the marketing |
|---|---|---|
| GoHighLevel, Vendasta | Yes | No. An assistant on a fulfillment shell. |
| Jasper, Copy.ai | No | Writes well, one client at a time. |
| Giant Context | Yes | Yes, from each client's own facts. |
So the platform sells the outcome rather than a better tool, which changes which budget it competes for. The whole marketing-software market is about $550 billion a year worldwide [9]. A system that does the work competes for the labor budget instead, the trillions that businesses and their agencies spend on the people who get work done. Foundation Capital put the market for services and labor that software has never reached at about $4.6 trillion [8], many times the size of the software pool.
An agency is where that shift bites first, because an agency is almost pure labor. Take the labor out of delivery and the cost of providing the service changes rather than the price of it.
So that is what I am building. One system that reads each client's own material until it understands the business, decides what marketing that client needs, makes it, and puts it out. It holds every client separately and works from each one's facts, so an agency runs a whole roster from one place and each client gets its own marketing rather than filler about some business in the same trade.
One system across a whole roster beats a pile of tools for the same reason it beats a stack for a single business. Every tool in an assembled setup knows only its own slice. One system works from a single understanding of each client, so the same facts move that client's pages, emails and posts without anyone keying them in three times, and the overhead of wiring tools together goes with them.
The work stays good because it runs on the client's own facts. You hand it the client's material, and it produces that client's marketing rather than confident filler about some company in the same trade. That is the difference between output an agency can hand a client and output it has to redo.
The first customer is me. The platform runs my own marketing, this site and the blog it hosts, and the saving from not building and staffing all of that by hand is what makes a one-person build possible at all. What it does not prove is the multi-client case, since running one business's marketing and keeping fifty separate are different problems.
There is a standard worry about a business like this. The classic version, made by a16z in 2020, is that AI companies run gross margins of 50% to 60% where software companies run 60% to 80% or better [11], for two reasons. Serving each customer burns real computing cost, where serving one more software customer costs almost nothing. And the output usually needs a person in the loop to check it. On that logic an AI business is structurally worse than a software business.
The cost of serving a customer here is the cost of the model doing the work, with no person in the loop being paid by the hour. And that cost is falling faster than almost anything else in technology. Epoch AI measured the price to run a given quality of model dropping from $20 per million tokens to 7 cents in about two years, close to 285 times cheaper [12].
When the main cost of serving a customer is compute, and compute is collapsing at that rate, the margin widens over time rather than thinning. The 2020 objection assumed a cost curve that has since moved.
That also settles how you pay. You pay for the marketing produced, not for seats or licenses. For an agency running many clients, the platform costs what the work costs. And that cost falls every year, as the model underneath it gets cheaper.
Everything above is the case for selling to agencies. The same automation points at a second market, the small business that could never afford an agency or a marketing hire at all.
There are about 35 million small businesses in the United States, roughly 28 million of them with no employees at all [16]. Their owners spend an hour a day or less on marketing and mostly doubt it is working [15]. They buy a stack of six or seven tools and use a fraction of each [13], because the point-tool era sold them the enterprise's software without the enterprise's staff [14]. For a business like that, the platform replaces the marketing manager it was never going to hire, whose fully loaded cost runs past $160,000 a year [17].
The trouble is reach. Selling to 35 million businesses one at a time is its own expensive problem, which is why the agency comes first. An agency is a single buyer that feels the saving multiplied across a roster, and it carries the small business in as its client. Neither market has proved itself. The agency is the sharper of the two.
An agency sells two things bundled together, the tools and the people who run them. The people are the expensive half, and the half the model cannot scale, because every client adds a little more payroll and leaves the margin a little thinner.
Automate the work and that stops. The roster grows without the salary growing under it. The agency keeps the relationship, the clients, and the judgment about what good looks like, and stops paying a team to produce what a system can.
Marketing is a plausible first candidate because it is made of a business's own facts turned into content, which is the shape of work these models handle now. Whether the same holds for other kinds of work is not something this thesis needs to be right about.
The math points to a large market no one has taken. Autonomous marketing, one system that runs a business's marketing end to end, is the thesis. I am building the platform to find out whether it holds.
Remote, Pacific time, full-time or contract. Get in touch.