Your Agency Sells Hours. Computing Just Stopped Counting.

The average organisation now spends $85,000 a month on AI. That figure is up 36% year-over-year, and the share planning to push past $100,000 monthly has more than doubled. If you run a digital agency, that number should make your stomach drop—not because your clients are spending money on AI, but because they're spending money on intelligence that used to come packaged as your invoices.

For 25 years, the agency model has rested on a straightforward proposition: we have skilled developers, you don't, so you pay us by the hour to translate your business problems into software. The entire value chain was denominated in time. Sprints, retainers, day rates, SOWs measured in person-days. The scarce resource was the developer, and the agency was the access layer.

That model just hit an expiry date.

The Instruction Is Dead

For sixty years, the fundamental unit of computing was the instruction. Deterministic, sequential, human-authored. A developer wrote code, a machine executed it, and the value sat in how cleverly someone could sequence those instructions. The developer's job was translation: take this business logic, turn it into machine logic. One function at a time. One Jira ticket at a time.

That unit of work has been replaced. The fundamental material of computing is now the token—a unit of purchased intelligence. You don't tell the machine what to do step by step. You describe what you want, feed it context, and buy enough intelligence to get a result. The machine works out the steps on its own.

This is not a tools upgrade. It's not "developers with better autocomplete." It's a categorical change in what computing actually is. And once you see it, the second-order effects on careers, companies, and entire markets become impossible to ignore.

The price curve tells the story in hard numbers. GPT-4 equivalent performance cost $20 per million tokens in late 2022. Today it costs roughly 40 cents. That's a 98% reduction in less than four years. And consumption hasn't decreased in response—it's exploded. This is Jevons' paradox applied to intelligence: when a resource gets cheaper, you don't use less of it. You use enormously more. Steam engines got more efficient; coal consumption exploded. Cloud computing got cheaper; AWS bills went up. Intelligence is following exactly the same curve.

According to reporting by Ed Zitron, Anthropic spent $2.66 billion on AWS through September 2025, against an estimated $2.55 billion in cumulative revenue over the same period. More than 100% of top-line revenue going straight to infrastructure—before even accounting for their Google Cloud spend. Perplexity burned 164% of its entire 2024 revenue across AWS, Anthropic, and OpenAI combined. These aren't companies being reckless. They're operating a fundamentally different model, one where intelligence is a purchasable input with a price curve, a consumption curve, and a set of second-order effects that reshape how organisations operate.

And they're growing fast enough to justify it. The bet is that the intelligence they buy today compounds into revenue that eventually outpaces the cost. For many of them, it's already working.

Agencies Are Selling Ice in a Warming World

Here's the uncomfortable bit for anyone running or working at a digital agency.

The agency model was built for a world where the bottleneck was human translation capacity. You had business problems on one side, machine logic on the other, and a layer of skilled humans in between converting one into the other. Agencies industrialised that translation layer. They hired developers, organised them into teams, sold their hours at a markup, and called it a service business.

When intelligence was scarce and expensive—when you needed a human brain to write every function, debug every edge case, architect every system—that model made sense. The agency genuinely added value because there simply weren't enough skilled developers to go around, and coordinating them required operational infrastructure that most brands didn't want to build.

But intelligence isn't scarce anymore. It's a commodity with a price that's falling faster than anyone predicted. And when the core input to your service offering becomes a commodity, the value migrates away from access and toward everything else: domain expertise, distribution, proprietary data, customer relationships, trust.

The agency that sells "we have good Shopify developers" is selling access to a translation layer that's being automated out of existence. Not tomorrow. Not next quarter. But the trend line is unambiguous and it's accelerating.

Consider what StrongDM's CTO Justin McCarthy recently disclosed: his three-person team targets $1,000 per day in token spend. No hand-written code. Three people. A thousand dollars a day in purchased intelligence, directed at outcomes. That's not an agency engagement. That's a small team with a clear spec and an intelligence budget. The developers aren't writing code—they're specifying outcomes, managing context, and evaluating quality.

Now imagine you're the CMO or CTO of a mid-market ecommerce brand. You're currently paying an agency £15,000 a month for a team of three developers working on your Shopify Plus store. You see StrongDM doing it with three people and a token budget. You see your competitors spinning up AI-assisted development internally. You start doing the maths.

The maths doesn't work in the agency's favour.

The Competent Middle Is the Kill Zone

The emerging developer market is splitting into three tracks, and agencies sit squarely in the worst one.

Track one: the orchestrator. These developers don't write code. They specify outcomes and manage the intelligence that produces those outcomes. System design, specification writing, quality evaluation, token economics. They think in agent architectures, context windows, evaluation frameworks, cost per outcome. Their value scales with the volume of intelligence they can direct, which means compensation correlates with token budgets rather than billable hours.

Track two: the systems builder. These are the people who build the infrastructure that orchestrators use—agent frameworks, evaluation pipelines, context management systems, routing layers that send the right task to the right model at the right cost. Deep technical work, closer to traditional systems engineering than application development, but on an entirely new stack. Small in number, highly specialised, enormous compensation ceiling because their work creates company-wide capability.

Track three: the domain translator. This is the track almost nobody is discussing, and it may be the largest of the three. These are people—increasingly people who don't identify as developers—who combine enough technical fluency to work with AI systems and enough deep domain expertise to know which problems are worth solving in a specific market. The dental practice management specialist who can now build tools. The construction scheduling expert who can now automate workflows. The insurance compliance analyst who creates custom applications instead of just filling in spreadsheets.

Now ask yourself: where does the typical agency developer sit in this map?

Not in track one. Agencies are incentivised to bill hours, not to optimise token budgets. The entire commercial model fights against the orchestrator role because efficiency reduces revenue. An orchestrator who solves the problem in two hours with £50 of inference is worth less to the agency than a developer who takes forty hours to build the same thing by hand.

Not in track two. Most agencies don't build platforms or infrastructure. They build applications on top of platforms that other people created. The systems builder track requires deep, specialised expertise that the generalist agency model doesn't develop or reward.

And not in track three, either. This is the painful one. Domain translation requires deep vertical expertise in a specific market. But agencies spread their attention across dozens of clients in different industries. The agency that works with a pet food brand on Monday, a fashion retailer on Tuesday, and a B2B parts distributor on Wednesday never develops the depth of domain knowledge that makes a domain translator valuable.

The agency developer sits in what the market is increasingly calling the "competent middle"—developers who write perfectly adequate application code but lack either the deep systems expertise of track two or the deep domain expertise of track three. This is precisely the profile identified as most exposed. Not because AI replaces them overnight, but because the value of generic code production is heading to zero at the same rate as the cost of tokens.

Your Client Is About to Become a Tech Company

The pressure isn't just coming from the supply side—it's coming from the demand side too. Your clients are about to stop needing you.

There's a bet circulating in Silicon Valley about when the first solo-founder billion-dollar company will emerge. Some people think it's happening this year. Whether or not that specific milestone lands, the underlying dynamic is real: the minimum viable team for building software is approaching one.

If you're a brand operator with deep knowledge of your market—if you understand your customers, your supply chain, your competitive dynamics—and you develop enough AI fluency to specify outcomes and manage intelligence, you don't need an agency anymore. You don't need a team of eight developers for six months. You need a clear spec, a token budget, and the domain knowledge to evaluate whether the output is good enough.

This is already happening. AI-native companies are running at three to five times the revenue per employee of traditional SaaS businesses. A $10 million ARR startup might operate with 15 people where a traditional company would need 55 to 70. Klarna's CEO went on record saying the world isn't ready for the impact AI will have on knowledge work—and he was describing his own company's journey, not theorising about someone else's.

For ecommerce specifically, this plays out in a very concrete way. The Shopify merchant who used to need an agency to customise their theme, build their integrations, and optimise their checkout can increasingly do it themselves. Not because they learned to code—because coding is no longer the bottleneck. The bottleneck is knowing what to build and why. And that knowledge lives with the merchant, not the agency.

OpenAI is reportedly planning agent pricing tiers: $2,000 a month for knowledge worker agents, up to $20,000 for research-grade intelligence. When your client can hire an AI agent for $2,000 a month that handles the work you were charging £15,000 a month for, the conversation about value gets very short very quickly.

The domain translator track—track three—is essentially your client becoming their own developer. The ecommerce operator who knows their vertical inside and out, who understands which problems are worth solving and which are distractions, who can point intelligence at the right problem with the right constraints. That person doesn't need a generalist agency sitting between them and the technology anymore.

Token Economics Is Now a Core Business Competency

If you take nothing else from this, take this: token economics is a real skill, it's measurable, and the organisations that build it are pulling away from everyone else.

Cursor became a billion-dollar revenue AI coding editor remarkably fast, then walked straight into a structural trap. It sent essentially all of its revenue to Anthropic in API costs. When Anthropic raised prices through its priority service tiers, Cursor's costs exploded overnight. The company was forced to gut its unlimited $20/month plan and introduce a $200/month tier. Users revolted. The subreddit turned into a complaint forum.

The lesson isn't that tokens are expensive. The lesson is that if you don't understand and actively manage your intelligence spend, you're one supplier pricing change away from a crisis. Cursor's response was instructive: it started building its own model. It recognised that dependency on a single intelligence supplier was an existential risk and moved to verticalise.

Enterprises that have figured this out are building internal platforms that route work to the right model at the right price point. Haiku for the cheap stuff, Opus for the hard stuff. Average enterprise spending on AI model APIs and platform tools hit $7.1 million in 2025, up from $4.5 million just two years prior. Projections push that into eight figures for 2026.

For agencies, this creates a fork in the road. You either develop genuine token economics capability—meaning you understand model selection, context engineering, cost-per-outcome optimisation, and evaluation frameworks—or you remain a body shop that happens to use AI tools. The first path leads somewhere. The second path leads to irrelevance on a timeline measured in quarters, not decades.

What Actually Survives

Not every agency dies. But the ones that survive won't look like agencies anymore.

The market is splitting along an axis of generalised scale versus specialised precision. At the top: enterprises and well-funded AI-native companies competing on token volume, building horizontal platforms, running agents on broad workflows. Across the enormous surface area of everything else: builders who win on specificity. The sharp angle. The niche market. The customer relationship that no amount of token spend can replicate.

The agency that knows "Shopify" doesn't have a moat. The agency that knows the operational complexity of a 50-location restaurant chain's inventory management system, and has built proprietary workflows and evaluation frameworks specifically for that problem—that agency has something defensible. But notice: it stopped being an agency somewhere along the way. It became a vertical software company.

The survivors will look like this:

• Deep vertical operators who picked one industry, developed genuine domain expertise, built proprietary tools and frameworks for that specific market, and compete on knowing their clients' problems better than anyone else. Their moat is distribution and trust, not developer headcount.

• Intelligence infrastructure firms that help enterprises build and manage their token economics capability. Not writing code for clients, but building the evaluation pipelines, the routing layers, the context engineering systems that enterprises need internally. Track two work, essentially.

• Outcome-priced consultancies that have abandoned the hourly model entirely. They charge for results, not time. They manage their own intelligence budgets internally, optimise ruthlessly for cost-per-outcome, and keep the efficiency gains for themselves instead of passing them to clients as reduced hours.

Notice what all three have in common: none of them sell developer hours. None of them are in the translation business. They've each found a way to create value that doesn't depend on the scarcity of human code-writing ability—because that scarcity is evaporating.

The honest assessment is that most agencies won't make this transition. The agency model selects for generalists. It rewards breadth over depth. It incentivises billing hours over reducing them. Those are exactly the wrong institutional muscles for a world where intelligence is cheap, domain expertise is valuable, and the competitive advantage sits in knowing where to point the tokens rather than how to write the code.

If you're running an agency right now, the question isn't whether to adopt AI tools. You're probably already doing that. The question is whether your business model—the fundamental economic logic of how you create and capture value—can survive a world where the unit of computing has changed.

For most agencies, the honest answer is no. Not without becoming something else entirely.

The unit of work changed. The unit of value changed with it. And the businesses that were built to sell the old unit of value are standing on a trapdoor that's already started to open.

Best of luck in the new world. You're going to need more than luck. You're going to need a plan.