The agency business model is shifting from deliver-and-move-on to accumulate-and-compound. Under the old model, an agency sold skilled hours, finished the project, and started the next one from a blank page. Under the new one, every engagement leaves behind specifications, context, and orchestration patterns that make the next engagement start from a better position. AI Digital’s 2025 data found only 13 percent of agencies were successfully creating AI revenue streams, so most of the industry hasn’t made the shift yet, and that gap is the opportunity.
What the traditional agency business model was built on
For most of my career, agency growth meant hiring more people, billing more hours, and keeping everyone busy. Skilled execution was hard to get, so agencies built companies around supplying more of it, estimated in hours, staffed by utilization, and invoiced on time spent. Growth was linear, and every layer of people added coordination.
That model produced the margin structure the industry now lives with. Promethean Research’s 2026 State of Digital Services found net margins of 19 percent at agencies under ten people, falling to 12 percent at ten to twenty-four, 9 percent at twenty-five to forty-nine, and 8 percent at fifty and up. Revenue per employee averaged about $163,000, with top performers near $250,000. Those are the economics of a business that sells time and keeps very little after the project closes.
What actually accumulates in an AI agency business model
During our first year working with AI, every new engagement still felt like starting over. Then, halfway through solving a problem, my co-founder and I realized we’d solved almost the same problem before, on two earlier projects, and the answers were sitting in old project folders nobody would open. What changed for us was becoming deliberate about whether those answers survived the project.
Three kinds of assets compound. Specifications carry forward what “done” means for a type of work, including constraints, review criteria, and the edge cases that broke earlier versions. Context libraries hold what someone needs to know to work on an account: preferences, brand voice, decisions already made, and the client quirks that used to live in one person’s head. Orchestration patterns are the workflows themselves, the handoffs that failed three times and were redesigned, and the review criteria that stopped a repeated error.
Two agencies can deliver similar work today and build very different businesses underneath it, and the difference is hard to see after one project and obvious after several years.
| Deliver-and-move-on model | Accumulate-and-compound model | |
|---|---|---|
| Unit of value | Hours of skilled labor | Outcomes plus the system that produces them |
| What remains after a project | Invoice, case study, memory in people’s heads | Specifications, context library, workflow improvements |
| How the next project starts | Blank page | From tested material, by the fourth or fifth engagement |
| Growth mechanism | Add headcount, add coordination | Raise output per person, keep headcount flat |
| Margin trend with scale | 19% under 10 staff to 8% over 50 | Depends on reuse rate, not headcount |
| R&D | Whatever time is left over | Protected: $1,000 to $3,000 a month plus a weekly afternoon at our agency |
| Pricing | Hourly or hours-derived fixed fee | Retainer, fixed fee, subscription, outcome |
Why reasonable decisions stop the compounding
Around 2007 I built a small mobile app called Stranger Danger that let people see who was nearby and start conversations. Client work picked up, the app needed attention I didn’t have, and I told myself I’d get back to it when things slowed down. They never did.
I don’t think choosing the agency was a mistake. I had clients, a team, and work that paid the bills.
Internal systems die the same way. A specification gets better after a project, and then a client deadline arrives and the improvement moves to next week. Nothing is broken, so there’s no reason to stop. After enough weeks the system stops improving.
Most organizations are sitting in that gap. KPMG’s 2025 data found that 91 percent of leaders expected AI to significantly improve operations within two years, while Deloitte’s 2026 State of AI found only 34 percent of organizations fundamentally rethinking how they operate. McKinsey’s 2026 State of AI survey of 1,719 respondents found that 74 percent of its high performers were redesigning workflows because of AI, compared with about 25 percent of everyone else. AI adoption is common. Operational redesign is not.
When repeated services become products
Once an agency preserves what it learns, some of it becomes useful outside the agency. The opportunity develops along two lines: the work you repeat and the kind of client you repeatedly do it for. A specification library refined across ten early-stage startup engagements might be a framework another agency would pay for.
We built CampaignSpark, which monitors Google Ads accounts through the day, and CodeRaven, which reviews code as it’s written, because we needed them ourselves. CodeRaven reviewed 1,164 pull requests in its first 148 days at one client for $161.20 in total AI spend. Neither started as a product idea. Consulting exposed a problem, project work forced us to solve it repeatedly, and the repeated solution became part of how we deliver other engagements.
We watch for three signs before taking an internal system seriously as a product: it works across multiple clients without major changes, the team uses it on nearly every engagement of a type, and outsiders have started asking how we do the thing it automates. All three true means it’s worth a closer look, not an automatic launch, because product businesses bring their own distractions.
The market puts a price on this. FE International’s 2026 agency M&A analysis reports that agencies with 80 percent or more recurring revenue trade at 5 to 7 times EBITDA, compared with 3 to 4.5 times for project-heavy shops, with owned IP adding another 1 to 2 turns. Setup work, meanwhile, is getting cheaper. More clients can connect a model themselves. A three-year library of tested specifications is much harder to reproduce.
Protecting R&D time
None of this happens reliably if systems only improve when someone has free time. At our agency that has meant somewhere between $1,000 and $3,000 a month for models and tools, plus an afternoon each week for experimentation when the schedule allows. The dollars are manageable. Protecting the hours is harder.
The process is simple. Someone explores an idea, we document what happened whether it worked or not, useful experiments move into production, and failed ones leave a record so we don’t repeat them six months later. After a year, specifications were more complete, workflows more reliable, and quality checks caught problems we used to miss. Some tools we built were later replaced by commercial products, and building them taught us where the commercial tools would fail.
How to know whether the agency is compounding
Every experienced agency gets better at something, so it’s easy to describe that improvement as compounding. A more useful test is whether the next engagement begins from a better position than the last. We watch four things.
Hours required for the same kind of deliverable, comparing an early engagement with a recent one. If the fifth site build in a category takes the same effort as the first, the learning hasn’t reached the production process. Revision rate on first drafts, because if drafts arrive faster but get rewritten more, the efficiency just moved into review.
Specification reuse, meaning how much of a new engagement starts from something already known. And the exception queue: if output doubles and the decisions needing a person double with it, there’s more capacity but no more independence.
None of these needs to be precise. Direction matters more than the number, and it has to be visible to someone besides the founder, because a feeling that the team is faster survives long after the numbers stop supporting it.
The adoption gap is the opportunity
On a typical day our agency runs fifteen to thirty AI workstreams across about ten people, eight of them full-time. Our revenue still looks a lot like a traditional agency’s, because the business is AI-assisted, not autonomous. Developers still decide what gets built, project managers still review client communication, and designers still decide what’s good enough to show.
Early 2026 Forbes reporting on CB Insights and Redpoint data put AI-native companies at two to four million dollars in revenue per employee, against $150,000 to $200,000 for a traditional agency. Those are software businesses, and the comparison mostly shows how different the economics are. The useful comparison is between agencies: Promethean found 34 percent had implemented AI across the business and 28 percent were still doing so, while narrowed-focus agencies posted 30 percent net margins on 13 percent growth.
Most agencies are using AI. Far fewer have rebuilt the business around it, and building systems that remove human effort while still selling human effort caps how much of the gain the agency keeps.
What this means for an agency founder
The AI agency business model isn’t a software pivot. It’s a service business that keeps what it learns. Pick one service you’ve delivered at least three times, compare the first engagement with the most recent, and look at the hours. If the newer one is faster and you can say why, there’s already something worth preserving, and if not, the old project folders will show you what the team keeps rediscovering.
The full argument, including how the pricing model has to change alongside the operating model, is in The Cognitive Agency. You can also see what one of these accumulated systems looks like in practice in our breakdown of five months of AI code review. And if you want help working out what your own agency should keep, systematize, and price differently, our AI consulting and development team does that work with agencies and their clients.
