There is a story most of the agency industry is telling about AI right now. It goes like this: humans do the human work, agents amplify it, the best delivery comes from both. It's optimistic, it's reasonable, and it's roughly the version on the homepage of every agency in our category.

It's also incomplete in a way I think is going to matter a lot, and soon.

The story leaves out a question we should be asking out loud. Where does judgment actually come from? Where does taste come from? Where does the architectural instinct of a senior engineer, the spatial sense of a senior designer, or the strategic intuition of someone who has run twenty engagements come from? The boring answer is the right one: it comes from doing hard problems badly, over and over, until you stop doing them badly. Debugging production at two in the morning. Bad rounds of exploration that go nowhere. Strategy decks that misread the brief. Refactors of systems you do not yet understand. Owned failures, taken slowly, paid for in time.

That's the substrate. Not the headline work. The substrate. It is the work nobody puts in a case study. It is also, demonstrably, where every senior practitioner in this industry developed the judgment they sell.

Here's the part that isn't being said: that substrate work is exactly what agents are best at eating.

Not the headline work. The substrate. First-pass research. Status synthesis. Boilerplate code. Layout exploration. The unglamorous reps where the practitioners we will need in five years would have, until recently, been quietly building the foundation under everything else they will do. Each individual delegation feels like a productivity win. On the day, it is. Compounded over a year, then five, the practitioners who came up without the substrate will not have the same foundation underneath their work. The judgment layer they sell rests on reps that increasingly are not happening.

From outside an agency, none of this looks wrong. Velocity is up. Output is up. Deliverables look crisp. The judgment layer underneath, the thing the deliverables are supposed to be expressing, is just structurally different than it was five years ago. It still looks right. You can no longer assume it bears the same weight.

The market signal is real, but it isn't extinction

It's worth being precise about what the data actually shows, because the discourse on this swings between "agencies are doomed" and "nothing has really changed," and neither is true.

What is true is that buyer psychology has shifted. The 2025 RSW/US Professional Services New Business Survey describes the climate as "the work is out there, but it's taking longer to win and coming in smaller chunks," with more than half of agencies reporting that landing new business is harder than the year before. AgencyAnalytics' 2025 benchmark research found that nearly three-quarters of agency leaders believe generative AI has indefinitely changed how people discover content, and that paid advertising has now overtaken SEO and web design as the most widely offered service among surveyed firms. Demand isn't disappearing. It's moving.

The AI builder category has reached a scale that's hard to dismiss as toys. Vercel raised $300M at a $9.3B valuation in 2025; v0 has reported more than 3.5 million users. Lovable's ARR moved from $300M to $400M in a single month earlier this year, with 200,000 new coding projects created daily. Bolt's own marketing copy speaks directly to agencies, promising they can "deliver more projects, faster, without scaling headcount." That positioning is not anti-agency. It is a productivity wedge. It is also a useful tell about where the category thinks the wedge lands.

These tools do not need to fully replace agencies to change the economics of the work. They only need to reduce the mystery, perceived effort, and price tolerance around common execution tasks. And on that, they are succeeding.

The clearest framing I have come across in some recent industry analysis is this: AI and no-code aren't an agency extinction event. They are an execution-margin compression event. That is a more useful diagnosis than the doom version, because it tells you which parts of the business are exposed and which parts aren't.

What's actually exposed, and what isn't

The exposed work is the work where a client can now plausibly ask, "why are we paying a team for this?" Marketing sites. Quick MVPs. Prototype-to-demo work. Simple dashboards. Brochureware. Campaign variants. Content repurposing. Shallow internal tools. The work where AI builders can produce a credible 60 to 80% in an afternoon. The last 20 to 40% is still hard, often much harder than the first 60%, but the buyer's price anchor has already shifted by the time that becomes obvious.

The work that isn't exposed, or is exposed much less, is the work where the question isn't "make the thing" but "make the right thing, safely, coherently, and inside a real business." Product strategy. Brand positioning. Design systems. Enterprise architecture. Data, security, and compliance. Integrations. Migration planning. Complex UX. Research synthesis. Experimentation. Governance. Ongoing product operations. None of this is glamorous on a homepage either. All of it is harder to fake.

This is also where the answer the industry is currently selling, things like context architecture, prompting craft, and agent orchestration, runs into a problem. It is real skill, and it matters. It also doesn't work without the underlying judgment. Pattern-matching on what makes a good prompt is not the same thing as pattern-matching on what makes a good system. Without the substrate, context architecture produces work that looks plausible to people who don't yet know enough to recognize what's wrong with it. Which is, by definition, the same demographic doing the work.

The shift in what clients show up with

There is a second-order effect worth naming, and it is more interesting than the doom narrative.

Clients increasingly show up with AI-built prototypes already in hand. A quick Lovable build. A v0 sketch. A Bolt-spun demo their PM put together over a weekend. They have already crossed the "is this idea possible" threshold without an agency. What they need help with is everything that comes after: auditing what they've built, integrating it with their actual stack, securing it, branding it properly, making it scale, turning it into something a customer would actually pay for.

This isn't a threat. It's a reframe. The sales motion shifts from "we can build your idea" to "we can turn your rough AI-built idea into something real, differentiated, secure, and worth putting in front of customers." The first version is what the AI builders are commoditizing. The second is what they can't.

The largest agency groups in the world are reading the same signal. Publicis raised its 2025 revenue forecast on the back of AI-driven products and services and stronger-than-expected Q3 performance. WPP launched WPP Open Pro to let brands plan, create, and publish marketing campaigns themselves, and entered a five-year, $400M partnership with Google to embed Gemini, Veo, and Google Cloud capabilities into its services. None of these moves is "humans instead of AI." They're all AI-enabled operating capacity plus human judgment. The pure-execution shop is the model under pressure. The judgment-and-integration shop, augmented by AI internally, is the one growing.

The organizational design question nobody wants to answer

If judgment is what gets priced higher in this market, the obvious question is how an agency keeps its judgment sharp while the substrate work that historically built it is being absorbed by tools.

The naive answer is to tell individual practitioners to choose the hard work. Just push back on the easy delegation. Slow down. Fail at things yourself. This advice is correct and almost completely useless, because every incentive structure in a creative services business punishes it. Velocity. Deliverables. Client scope. "Let me slow down and fail at this myself" reads as a cost center to anyone holding a P&L. No individual practitioner survives that posture for long without organizational backup.

Which means this isn't an individual problem. It is a problem of organizational design.

The agencies that come out of this transition stronger, the ones priced on judgment in five years, will be the ones who treat developmental friction as a thing the organization explicitly pays for. Not aspirationally. Not in a slide. As a line item, owned by someone, defended against the velocity pressure that would otherwise eat it. That looks like senior practitioners deliberately working through problems they could have delegated, because the delegation costs the team's collective judgment. It looks like junior practitioners owning hard problems end-to-end instead of being handed scaffolded solutions. It looks like leaders defending the slow, failure-rich work as load-bearing, not as overhead.

It is not free. It costs against velocity, which is the metric the rest of the market is racing on. That is the point.

Where Metalab fits

This is the part where I am supposed to make the pitch. I'll keep it short, because the analysis above is the actual argument and I would rather it stand on its own.

Metalab has been doing this for a while. Since 2006 we've shipped over 475 products, including 24 unicorns, used by more than 2.2 billion people. The way we got there is not glamorous and it has not changed. Small focused teams. Senior-heavy from day one. A bias for shipping over abstracting. Our founder once described it as "we shut up and build stuff, get feedback, and iterate until both our team and the client love it." We've been a remote-first, judgment-priced, craft-obsessed shop for nearly twenty years, well before any of those words became fashionable.

The recent work has been mostly AI-native, by the way. Suno, Cognition's Windsurf rebrand, Modular, and others we'll talk about when we can. We are not the agency arguing humans should win this round at AI's expense. We are the agency that hires the people whose judgment is worth amplifying, and uses the best tools in the industry to amplify it.

That posture used to be table stakes for a premium firm. In the world we are heading into, it is increasingly a moat. Not because the people are unreplaceable, but because the organizational choices that produce them are unreplicable on a timeline that matters. You cannot crash-build that bench in eighteen months.

We are going to keep telling this story publicly, because the industry needs the conversation. We are not afraid of what AI does to creative services. We were already built for what comes after, and we are increasingly excited about the version of the agency model that emerges on the other side of this compression event. Smaller teams. Senior-heavy judgment. Faster prototyping. Stronger technical architecture. Tighter strategy. Pricing tied to outcomes. Less hours-resold-as-value, more judgment-priced-as-value.

The quiet trade everyone else is still making, substrate for velocity, judgment for output, foundation for ship date, is going to look obvious in retrospect. The agencies that thrive will be the ones who recognized which work was actually being commoditized, repositioned around the work that wasn't, and paid the organizational cost of keeping their judgment sharp on purpose. We think that's a good place to be.

---

Curtis is a Lead Software Engineer at Metalab, working at the intersection of AI tooling, engineering process, and collaborative design and engineering practice. He writes occasionally about how the agency industry is adapting to AI-native workflows.