Your team adopted AI across production last year. Designers generate layout directions in an afternoon. Copy drafts arrive before the content strategist has opened the brief. Developers scaffold sections faster than they can name them. The launch date moved by roughly nothing.
That gap between generation speed and delivery speed is the most common pattern in agency production right now, and the cause is not the tooling. Google’s 2025 DORA research on AI-assisted development, drawn from nearly 5,000 technology professionals, found that AI functions primarily as an amplifier of an organization’s existing strengths and weaknesses, and that the largest returns come from the underlying system rather than from the tools themselves.
AI changed the cost of producing options. It did not change the cost of absorbing them. Everything downstream of generation still runs at the speed your web design system allows, and most agencies are running on a system nobody designed.
A web design system is not a process document
A web design system is the set of decisions your team has already made before a project starts. It defines what the build unit is, where styling is controlled, who is allowed to change a page, and what happens when a change arrives in week seven. Those decisions either exist in a form the whole team can act on, or they exist as a habit.
Most agencies have a habit. There is a folder of past projects, a Figma file that drifted from the build three sprints ago, and one developer who knows why the spacing works.
That arrangement produces good websites. It produces them by absorbing effort nobody scoped, which is why every retro identifies the same problems and every next project repeats them.
Agencies did not choose this workflow because it produces better work. They inherited it. The custom theme era rewarded teams who could build anything from nothing, and the tooling reinforced that for fifteen years. The habit outlived the constraint that created it.
Generation got cheaper and review did not
The constraint in most agency projects has moved from producing work to deciding about work. Your designer can now generate twelve credible homepage directions in an afternoon. The person reviewing them owns the client relationship and the consequence of choosing wrong, and their capacity has not changed since last year.
This surfaces first in client presentations. More options reach the client because more options exist, and the decision gets harder rather than easier. Rounds multiply, and the timeline AI compressed at kickoff expands again at approval.
It surfaces second in build. Across 500+ builds at Forge and Smith, the phases that consistently overran were never the ones where someone was typing. They were the handoffs, the approvals, and the rework triggered by a decision that landed too late to be cheap. We mapped where website projects actually lose their hours in more detail, and the distribution has not changed since AI entered the workflow.
Output without a component layer is faster one-off work
Every AI-generated layout still has to be built by a person, and where nothing exists as a reusable build unit, each variation becomes a fresh build with fresh QA attached.
The cost people miss here is translation. A designer imagines a section, a developer interprets it, and QA verifies the interpretation. The client revises, and the developer rebuilds against a new interpretation. Five people handle the same section, and every handoff is a chance for it to arrive slightly wrong. AI adds volume to the front of that chain without removing a single step from it.
A component layer collapses those translations into one object. When the thing your designer explores and the thing your builder assembles are the same object, there is nothing left to interpret. Refoundry’s reusable components function as the actual build unit, which turns a layout direction into a page rather than a ticket.
Forge and Smith measured the difference directly. Before the component layer existed, a mid-sized site consumed 80 to 120 build hours and another 40 to 100 in QA. Across roughly 170 Refoundry builds since, the same scope runs 40 to 50 build hours and 20 to 25 in QA. The scope did not change. The translation cost did. The before and after of component-based web design covers what else shifts once that layer exists.
Speed at the start does not survive the first real change request
Late-stage change is where website projects lose their margin, and AI has no effect on it. The client sees staging in week seven and wants the service section restructured. That request costs what it has always cost, because the cost lives in how your build responds to change, and AI never touched that part of the system.
In a custom theme build, restructuring means a developer, a ticket, a queue, and a regression risk across every page sharing that template. Six weeks of AI-accelerated production gets consumed by one structural change. In a component-driven build, the same request is a reassembly. The build does not break under change. It absorbs it.
Faster drift is still drift
AI-assisted output multiplies inconsistency in any system where styling lives at the page level. Every generated section arrives with its own spacing logic, its own type scale, and its own idea of what a button is. Someone reconciles all of it, and that someone is usually sitting in QA with a launch date behind them.
DORA’s summary of the same research is blunt about the tradeoff: AI improves throughput, but often at the cost of stability when the foundation underneath is not solid.
Global styles are the control that resolves this. When type, colour, spacing, and buttons are governed centrally, generated output inherits the system instead of negotiating with it. QA returns to being a check, and consistency stops depending on who was paying attention that week. We have broken down where build quality actually originates separately.
Four system decisions matter more than your AI tool
Four decisions determine whether AI output converts into delivered work, and none of them involve buying an AI tool.
A component layer that is the actual build unit. The library your designers pull from and the library your builders assemble from have to be the same library. Where those diverge, every project pays a translation cost that scales with the number of ideas in play.
Global styles with central control. Type, colour, spacing, and buttons governed in one place, with local overrides available where a project genuinely needs them. This is the mechanism that stops volume from becoming drift.
A named review point with an owner. More options require faster decisions, and faster decisions require someone whose job it is to make them. Without that role defined, additional generation capacity converts directly into additional rounds.
A change path that does not route through a developer. When a content strategist can restructure a page in the editor, late-stage change stops being an escalation. This single condition moves more hours than any generation tool currently on the market.
Forge and Smith rebuilt around those four conditions over 500+ builds, and Refoundry is the version of that system we now ship.
What this means for your next project
AI is a multiplier applied to whatever your web design system already does. If your system rebuilds every page from scratch, AI helps you rebuild faster. If your system routes every change through a developer, AI fills that queue sooner. If your system assembles from components under central style control, AI compounds into delivery gains you can put on an invoice.
The variable you control is the system, and it was the variable long before AI arrived. What changed is the cost of ignoring it. Every AI tool eventually runs into the architecture underneath it.
Try it on a real project
The fastest way to test any of this is on work you are already doing. Start a Refoundry build, put a component layer under a live project, and measure what happens to the second and third page rather than the first. Every system looks the same on page one.
If you want the numbers before you build anything, our agency case studies break down what changed at Forge and Smith across build time, launch volume, and profitability.
