In brief: On a life sciences program, our product design team cut a design cycle from four weeks to just over one by moving to parallel workstreams and building a design system that both designers and AI agents could use. The bigger lesson in AI product design was that speed brought us sooner to questions AI can’t answer on its own, such as whether the data is real and which experts actually know.
In our work with a large life sciences pharmaceutical company, the loudest and most consistent request was simple: go faster.
So, we did. Not by forcing speed, but by evolving how we worked, sprint by sprint, until the whole system changed shape. Over a handful of sprints, our product design team cut a design cycle from four weeks to just over one. We built a design system that both humans and AI could understand and use, moved our work directly into the environment where the real product gets built, and stood up a team of specialist AI agents to work alongside us.
Then came the most important lesson we learned about AI product design. It has almost nothing to do with speed.
Going faster with AI exposed bottlenecks that were already there
Acceleration has a side effect that rarely comes up. When you speed up the parts of the work that can go fast, you arrive much sooner at the parts that can’t. AI accelerates everything that lives in language, logic, and pixels. However, it can’t always speed up knowing whether the underlying data is real.
Every screen we design makes a promise. Our UI designs tell our users that this number exists, that this connection works, or that this field means what you think it means. Keeping that promise means doing the data readiness work: mapping the actual data, validating that it exists, and figuring out what still needs to be built. No amount of generative speed changes what is actually true in the system underneath.
Here is the analogy I kept coming back to. We built a bullet train and laid down a streamlined high-speed rail system. But a train still has to make stops to let people on and off. Our train can’t teleport passengers aboard while flying past a platform at full speed. It doesn’t matter how fast the train is; people, or in this case data, board and disembark at their own pace.
That is an honest assessment, not a complaint. The train is genuinely fast. The stops are genuinely necessary. Pretending that the stops don’t exist doesn’t make the journey shorter. You either derail the train or arrive at the final station with no passengers. What you can do, however, is make the stops themselves more efficient. Think: better platforms, clearer signage, or knowing who’s waiting before you pull into the station.
With AI, the design bottleneck moved from making to knowing
For years, the constraint on product design delivery was making things. The work of pushing pixels and designing screens was slow, so “making” was where both the value and the bottlenecks lived. AI has made “making” almost free.
But when making becomes free, the constraint doesn’t disappear. It moves upstream to knowing. Is this data real? Does it connect? What does this field actually mean? Those questions were always there. The speed brought us to them sooner.
And knowing doesn’t scale the way making now does. You can throw more people or more compute at making. You can’t throw more compute at institutional knowledge that lives in one person’s head. Knowledge has no API. It can’t be parallelized or generated on demand. It travels at the speed of the expert who holds it, and in a regulated industry like life sciences, that expert is often a single person with a demanding calendar that pulls them into multiple priorities and deadlines.
The real bottleneck sat in the truth underneath the design and in the small number of people who hold that truth. Speed turned out to be the easy part.
One finding before the rest of the series: speed came from the shape of the work
If I give away just one thing from what follows, it’s this. We doubled our design team, from three to six, and cut our design cycle fourfold. An industry expert might notice those numbers don’t add up because doubling our team doesn’t quarter our timeline.
Our speed didn’t come from people working faster. It came from how we changed the shape of the work, moving from sequential handoffs to parallel workstreams, and from building the things that made headcount matter less, specifically a design system that both our designers and our AI agents could build from without guessing.
That’s the pattern under everything in this series on AI product design. We got fast by building systems that made speed safe, then putting our people to work on the one thing systems can’t do: knowing what’s actually true. Over the next four articles, we’ll show how we accelerated, how we built the design system beneath it, how the AI agents work on top of it, and how we confronted the human bottlenecks that the speed exposed.
The design teams that get the most from AI over the next few years will know exactly when to slow down. They’ll also build the memory and the plan to make sure they only have to learn each truth once.
Next in the series: how we took a design cycle from four weeks to just over one.
Frequently asked questions
AI product design is the practice of shaping how people and AI work together inside a product. In traditional design, it is assumed that the product will behave the same way every time. This isn’t the case with AI products. Modern AI solutions generate responses, make recommendations, and increasingly act on a user or agent’s behalf. So the design work has shifted from designing screens to designing behaviors. This means deciding when and where AI should act on its own and when a person should provide judgement and/or approve. It also means designing trust and for the moments when AI gets things wrong. All this is measured so that we can actually improve outcomes that people care about. AI also changes how the work gets done. We collaborate with AI to research, prototype and build design systems faster. But speed is the smaller shift. The bigger change is that we as designers are now responsible for the experience of a products that can think.
AI makes the making part of product design fast and cheap. The constraint moves upstream to knowing whether the data behind each screen exists, connects, and means what the design says it means. Teams reach those questions much sooner than they used to.
When the fast parts of the work speed up, teams reach the slow parts sooner. In AI-accelerated design, the slow part is usually knowledge held by a small number of experts. In regulated industries such as life sciences, that expert is often a single person balancing several priorities and deadlines. That pressure shows most clearly when validating solutions. We can now produce ten workable concepts in the time it once took us to produce one, but each design still needs to be validated for accuracy, compliance, and is this the right solution.
In our experience, no, but the design roles are in transition. Specialist AI agents worked alongside our designers, and the speed came from parallel workstreams and a design system both could build from. What is changing is where our designers focus on their time now. Less is spent on pushing pixels and more went toward judgement, deciding which direction was right, confirming what is true in the data behind the product, and making calls AI can’t make alone.
