Where AI belongs in the process, and where it does not
Art direction stays human. Deciding what a brand should feel like. Which idea a page is built around. Whether a composition is right, and when something is simply not good enough. Those are judgments a model cannot make on a client’s behalf. Production runs differently. Generating a layout variant, producing imagery to a specification, drafting copy against a brief, filling a design with real content instead of placeholder text, are all work that rewards volume and speed.
Systems that hand over the judgment and keep the humans producing are why so much AI-assisted design looks the same. We built ours the other way around.
Any source, including a generated one
Our ingestion pipeline is deliberately indifferent to where a layout came from. Generated by a language model, drawn by an agency, exported from whatever tool a designer prefers, or hand-coded by somebody thinking in CSS. Markup and styles go in, a working pattern library comes out. Because the contract is narrow, a team can use whatever produces their best work today and change their mind later without anything downstream breaking.
For AI-generated layouts specifically, the pipeline closes the gap that usually kills them. A model produces a plausible composition, and turning that into a maintainable site inside a content platform is where the time savings disappear. The classification pass handles the conversion. It reads the design, types every editable region against a closed vocabulary, and writes the manifest the platform consumes.
Imagery generated for the subject, not selected from a shelf
Stock photography has a structural problem that has nothing to do with quality. The same photographs circulate through an entire industry, so a page illustrated from a library looks indistinguishable from its competitors, while a visual index treats repeated imagery the way a web index treats duplicated text. We generate imagery per page from a mapping of setting, product type and context, so a particular application in a particular environment gets visuals built for that combination rather than the nearest available approximation.
Everything produced lands in the media library as a properly managed asset, carrying alternative text, responsive sources, and one-click replacement referenced by identity rather than by file path. Provenance and license are tagged, so nobody reconstructs months afterward whether a particular visual can be used, resold or reproduced.
Real content during the build, not after it
Every designer knows the specific betrayal of a layout that looked correct with placeholder text and came apart against the client’s actual copy. The headline that was two words is now nine. The card blurb that ran to one line runs to four. Because copy and imagery are generated into the page during the build, with the composition and direction already set, the real thing shows up early. Long headlines reveal themselves with time still left to adjust the type scale. Ragged card heights appear in review rather than after launch. The gap between a finished layout and the same layout carrying real words used to be measured in weeks. Now it closes inside the build.
Customization without a specialist for every change
The granular work is where design budgets historically disappeared. A variant needed for one product family. A template reproducing a form a client’s customers have received for years, or a seasonal treatment applied across a category. Each of those is small on its own, though collectively they are why a design retainer never ends. AI absorbs a meaningful share of the granular work, producing variants against the constraints the design system already enforces, with a person reviewing rather than authoring from scratch. A generated variation is written against a closed vocabulary and a validated pattern library, so the system rejects anything malformed rather than relying on somebody’s attention.
The rules that keep generated work from becoming sludge
A closed vocabulary
Bounds what can be produced, so composition stays open while the parts a machine must understand come from a defined set that can be checked mechanically.
Hygiene gates
Run before anything is packaged, covering landmarks, heading order, contrast range, focus visibility, self-hosted fonts, plus media budgets.
Human review at defined checkpoints
Rather than arriving as a vague final glance.
A differentiation rule
Writes each page against its siblings, because independently generated work converges on sameness with remarkable reliability.
Where AI-assisted work connects
The pattern library and named values that constrain generated output live in design systems, the ingestion contract and manifest are covered under the Model Context Protocol, and the writing discipline applied to generated copy is covered in AI-crafted context.
How the engagement runs
Design production happens through screen shared working sessions, with real content and real imagery in the build from early on rather than arriving at the end. If you have tried generated design work and got something that looked like everybody else’s, the missing piece was the system around it.