The context problem: Why AI design tools need more from pharma before they can give more back
by Ed Hart 27 May 26Every few years, a new tool arrives and the design industry briefly loses its mind.
Photoshop. Sketch. Figma. Each one shifted what was possible. Each one prompted the same cycle of excitement, adaptation, and eventual integration into normal practice.
The AI design tools arriving now feel different. Not incrementally better at the same thing, but genuinely disruptive to the relationship between an idea and something real. We have been deep in testing across a range of these tools, including Claude Design and Figma's AI integrations, and the distance between concept and working, testable prototype has collapsed in a way that still surprises us.
But here is what the hype misses, particularly in pharma. These tools will produce output immediately. What they produce depends almost entirely on what you give them to work with.
AI's favourite food is context
An AI design tool given no context will produce the average of everything it has ever seen. Plausible, functional, and completely generic. A nothing burger with your logo on it.
That is a manageable problem in a startup. Small team, single product, relatively contained brand. In that context, AI tools can be extraordinary.
Pharma is not a startup. The organisations we work with operate across multiple brands, multiple therapeutic areas, multiple markets with regulatory frameworks that differ by country. They communicate with oncologists making high-stakes clinical decisions and with patients who have just received a diagnosis that has turned their world upside down. The visual and tonal signals that serve those audiences are not the same. The wrong choice, made by a machine with no context, is not a minor aesthetic issue.
If you are adopting AI production tools and you do not have a structured design system connected to them, you are essentially asking the machine to guess. Your brand logic. Your tone. The emotional register appropriate for a cancer screening context versus an HCP portal. None of that exists for the machine unless you have built it in.
What a design system actually does for AI
Most design systems were built for humans. Visual references, Figma libraries, PDF guidelines. Useful for a designer who can read between the lines. Not sufficient for a machine that needs explicit, tokenised rules to produce anything consistent.
A machine-ready design system is different. It encodes not just the colour and typography logic, but the harder understanding that takes years to accumulate. It knows that a red status chip, standard in almost every component library, would be exactly the wrong choice in a cancer screening interface. It knows what visual register is appropriate when communicating with a patient who is frightened, versus an HCP who wants clinical precision. It knows the market-specific regulatory constraints that shape what can and cannot be shown.
That understanding has to be built deliberately. It does not emerge from a Figma library.
We have spent years building this kind of design system for pharma and healthcare clients. We have always built them for humans to learn from. Now we build them for machines too. The principle is the same: give the system enough context to produce something that genuinely serves the people using it. The difference is what that context needs to look like for a machine to use it.
What changes when the prototype is real
Once that contextual foundation is in place, the speed gains become commercially meaningful.
We are now building browser-native, coded prototypes as a primary deliverable rather than as a late-stage output. Real interactions. Real components. Real brand logic integrated from the start. The kind of thing you can put in front of an HCP in a research session this week, iterate on the findings by next week, and hand to an engineering team the week after as working code they can integrate directly. Not rebuild from scratch. Integrate.
There is one implication of this shift that we think is genuinely under appreciated in pharma specifically. Early regulatory conversations no longer have to happen on wireframes or static visuals. They can happen on a fully functioning prototype. Stakeholders and reviewers are responding to the actual experience, not an approximation of it. That changes the quality of the feedback. It changes the confidence going into the next stage.
Build the right thing. Build the thing right. Both have always been hard in pharma, and slow. The tools available now make both meaningfully faster, but only if the context that gives them direction has been built properly first.
The tools are ready. The question is whether you are.
We are not yet at a point where we would claim to have everything figured out, and we are sceptical of anyone who says they do. What we are confident in is the question that matters: not whether to adopt AI design tools, but what you are feeding them.
If your design system is a PDF someone made in 2019, the machine cannot use it. If you do not have one at all, you are generating output without foundation.
The opportunity is real. So is the work required to take it seriously.
Article written by Ed Hart, Director of Design at Graphite.
