The studios that waited on AI are the ones that are losing now. Not because their work is worse — it often is not — but because their throughput is. They are pricing themselves at the bottom of a market that is reshaping at the top. The studios that adopted early are quoting half as many days for the same scope and shipping at higher fidelity. Within eighteen months, that gap will not be a gap. It will be the market.
This is the framework we use at Two Words to put AI inside the design process — what it touches, what it doesn't, and where we are deliberate about keeping it out. It is a working document. We update it every quarter because the underlying tools change every quarter. The shape, though, has been stable for a while.
Five zones where AI genuinely helps
The first mistake teams make is treating AI as a general-purpose tool — a magic spreadsheet that should be on every step. It isn't. It is a set of capabilities that map well to certain tasks and poorly to others. The map we use has five zones.
1 · Ideation and divergence
Early in a project, you want the search space to be wide. Twenty concepts is better than two. AI is excellent at this because it has no ego and no taste — it will gladly produce the obvious version, the obvious-plus-rotation version, and the version someone wrote about online once. The designer's job here shifts from generating options to recognising the interesting ones.
We use Claude and GPT for verbal exploration ("give me twenty value propositions for an investor portal in the UK property market, ranging from utilitarian to aspirational") and Midjourney / Stable Diffusion for visual exploration ("editorial photography, dark background, single glass object, brand pharma launch"). Volume is the point. Curation is the craft.
2 · Pattern recall
Half of design is remembering what has been done before — by you, by your team, by the field. AI is good at being the institutional memory you never built. Feeding a model your component library and asking it to suggest a layout for a new screen is faster than scrolling through Notion. Asking it to surface similar prior work to a brief lets you start from a real reference instead of from scratch.
We built our own internal tool — the Two Words Co-Pilot — partly for this. It knows our design system, our previous projects, our voice guidelines. When a designer asks it for a dashboard pattern, it doesn't propose something generic. It proposes something that already looks like us.
3 · Iteration speed
The middle of a project is iteration. Make a version, react to it, make the next one. Historically this loop is throttled by how fast the design hands can move. With AI in the loop, the loop can run ten times in the time it used to run once. Vibe-coded prototypes in Cursor or v0, Figma plugins that generate variants from a single frame, prompt-driven copy revisions — all of these compress the cycle.
The risk here is that you can iterate so quickly you stop thinking. We have a rule: every fifth iteration, the team has to stop, look at the original brief, and answer in writing whether we are closer to it or further away. It sounds simple. Without that pause, the iteration can drift sideways forever.
4 · Production prototyping
The frontier zone. Modern AI-assisted coding lets a designer with rough technical instincts ship a working prototype to a real URL in hours. Not a Figma file pretending to be a product — an actual browser-based, responsive, data-fetching prototype. This is the change that is reshaping the field.
For a Klay Securities engagement, we used the Co-Pilot to generate a fully working onboarding prototype from a written brief in five working days. The client demoed it to their board the next Monday. Six weeks of design fidelity collapsed into a week. This was not slop. It was a real product, in a real codebase, that we then refined into the eventual ship.
5 · Validation and synthesis
After research interviews, transcription used to take a day. Synthesis took a week. AI does the first step in minutes and the second in hours. Feed a set of recorded user sessions to Claude and ask it to surface the top three friction points by frequency. The output is not a finished research deck — but it is a starting draft you can edit, which is much faster than a starting blank.
The same applies to usability scoring, accessibility audits, copy reviews against tone guidelines, and competitive teardowns. None of these end with the AI's output. All of them start with it now.
Where the craft still has to come from a human
Five zones is not the full process. There are at least four places where we are deliberate about keeping humans in charge, because AI is either bad at them or because doing them with AI corrodes something important.
Taste and narrative
AI can produce a layout that satisfies the brief. It cannot tell you that the layout, while correct, is a beat too earnest for the brand you are building. That judgement comes from years of looking at work, working with editors, building a sense for what a brand should feel like to a specific audience. The model can render a hundred competent options. The designer's job is to know which one is the right one — and to be able to defend that to a client.
Ethics, regulation, and the implications no one asked about
Designing for regulated industries — finance, pharma, healthcare — means making decisions that look like UI choices but are actually compliance choices. A dropdown that defaults to the wrong option can mis-sell a financial product. A disclaimer that disappears on mobile can trigger an FCA enforcement. AI doesn't know these things in a way you can trust. The human in the loop has to be the one to catch them.
Conversations that are hard
The most important conversations in a project are with the client, the user, and the team. Tough scoping conversations, hard feedback, the moment in week six when the brief has drifted and someone has to say so. AI can prepare you for them. It cannot have them for you. Studios that try to outsource these to a tool produce work that is technically fine and politically poisoned.
The first principle of a project
What is this product actually for. Why does the world need it. What is the company really trying to do. These questions get asked once at the start and re-asked every few weeks. They have to be answered by a human with a real opinion and stake in the answer. Outsource this and the work becomes very competent and very meaningless.
AI can render a hundred competent options. The designer's job is to know which one is the right one.
The governance layer
Even within the five "yes" zones, there are guardrails. Without them, you produce work that is faster but that you cannot ship.
Data. Don't feed client material into a model whose training data policy you haven't read. We use enterprise plans with explicit no-training clauses. For sensitive clients (pharma, finance, defence-adjacent) we run smaller models locally and never let prompts leave our environment.
Model selection. Different models for different jobs. Claude for long-form writing and structured reasoning. GPT-4 for code generation and exploratory design coding. Midjourney for moodboards. Stable Diffusion for production renders we want to fine-tune. Using the wrong model wastes time and produces worse output.
Disclosure. We tell clients which parts of the work were AI-assisted and which weren't. We have never had a client complain. We have had several thank us for the transparency. It also forces us to be honest with ourselves about the line.
Provenance. Anything we ship that an AI touched goes through human review before it leaves the studio. That includes copy that was rephrased by Claude, image alts that were generated, code that was vibe-prototyped. The model is a junior. The designer is the editor.
A worked example: three days, not three weeks
A pharma client came to us in March needing a launch site for a new therapy. The first draft of the brief was vague. Historically, week one would have been turning the brief into something workable. Here is what we did instead.
Day one — fed the brief into Claude with our pharma-launch playbook. Got back a structured tree: page architecture, headline hypotheses, regulatory checkpoints, three tone variants. Pruned to one tone, one structure. Twenty minutes of human work, four hours of analyst-equivalent output.
Day two — generated three visual directions in Midjourney with art-directed prompts, two in Figma using AI plugins on our component library. Picked one. Cleaned it up by hand.
Day three — used Cursor to vibe-code a working prototype of the home page and one therapy page. Real responsive HTML, real CTAs, real navigation. The client clicked it live on a call the next morning.
Week two onwards was where the craft happened — typography, copy passes, regulatory review, the slow careful work of making something feel inevitable. AI didn't replace that. It bought us the room to do it without working weekends.
The next 24 months
Three predictions, none of which we are betting the studio on, but all of which we are watching for.
Design systems become live. Static component libraries get replaced by ones that can generate new components on demand, in your style, from a description. The system stops being a museum and becomes a factory.
Prototyping fidelity matches production fidelity. The line between "Figma file" and "shipped product" disappears for most use cases. We are already mostly here.
The clients who get the most value will be the ones who could already write a good brief. AI doesn't fix bad inputs. It amplifies whatever you start with. The studios that survive will be the ones that double down on the upstream craft — strategy, taste, narrative — and let AI do the volume work downstream.
Closing thought
This is a craft amplifier, not a replacement. The studios using it as a replacement are producing slop and undercutting themselves on price. The studios using it as an amplifier are producing better work and charging more for it. The difference is not the tool. It is the framework around the tool.
Designers who can hold both — the taste to know what good looks like, and the discipline to use AI on the parts that don't need taste — are going to have a very good decade.