Why AI and Design Thinking were made for each other.
Before a single pixel illuminates in a prototyping tool, Design Thinking asks the same question it has always asked: “What problem are we trying to solve?” Agentic AI hasn’t changed that question. It has changed who or what can be trusted to answer it, and the answer is more encouraging than you may expect.
My co-author, Scott Fegette and I have spent the past several months exploring the intersection of Design Thinking and AI in our new Apress book: Creative AI. Generative and Agentic AI for Design Thinking and the Creative Industries. It turns out Design Thinking and Agentic AI were made for each other when working through the Discover and Define phases of the Design Thinking process.
Using Design Thinking rigorously across much of the Research and Definition phase of the UX process forces the team to answer the core question: What problem are we trying to solve?
To explore how Design Thinking and AI intersect, we created a mythical campground where management decided they wanted a visitor app. Before involving Figma Make, we had to define the problem, question management’s assumptions, synthesize research, and arrive at a point of view worth building from. Sound familiar?
None of that is design work in the traditional sense. It’s the work that makes design possible, and a properly directed AI agent can carry almost all of it without skipping a single stage of the Design Thinking process.
That claim may sound like an argument for less intellectual rigour. It’s the opposite. Managed well, an agent doesn’t cut corners in discovery work; it removes the excuse.
The brief is not a design document, and neither is the PRD
A typical project starts with a client’s assumptions dressed up as facts. Park staff spend their peak season answering questions about parking and canoe availability. Management assumes an app will fix that. Maybe it will. Maybe the real problem lies elsewhere entirely. A Product Requirements Document captures what the client currently believes, not what’s actually true. A PRD is just a record of assumptions.
Analyzing those assumptions and writing the resulting document can take days. Handing the brief to an AI agent and asking it to generate a structured PRD covering business goals, assumed users, proposed features, technical constraints, and known risks takes about ten minutes. Don’t be mesmerized by those 10 minutes. The output isn’t the point. Your review is the point. Nothing in an Agentic Workflow proceeds without you, a Human In The Loop. You read it and balance it against what the client actually said. You correct anything the model invented or overstated, and only then does a human approve it. That correction step is where your judgment enters the process, and it’s non-negotiable. An agent will hand you a confident, well-organized draft that looks really impressive.. Confidence isn’t accuracy, and “seems about right” isn’t a solution.
Once approved, that PRD becomes the first entry in a living document-Design.md-, not a PDF that gets emailed around and immediately goes stale. Our solution was to format the document in Markdown, inside a project folder, alongside a second file- Architecture.md- that holds every technical constraint the build will need. Every subsequent decision gets appended to those files as the project moves forward. Nobody on the team, human or AI, starts from zero. Nothing gets assumed twice, and everyone works with the same knowledge base.
Empathize is where the agent earns its keep
This is the stage most people assume can’t be handed off. Empathy is a very human thing and is the one thing a language model can’t do. It doesn’t need to. The agent isn’t feeling anything on the visitors’ behalf. It’s doing something narrower and more useful. It is reading a research report at a level of attention most of us don’t have time for. It then surfaces what’s actually in it rather than what we expect to find.
The working sequence is deliberately constrained. First, ask the agent to identify patterns in the research without proposing solutions. Then ask it to challenge the client’s own assumptions against that same research, and to say plainly where the evidence is thin. In the park project, four assumptions in the original PRD didn’t survive that test. Visitors didn’t want a digital-first experience; they wanted something they could put away. The canoe problem wasn’t a booking problem; it was deeper than that. That’s not a small correction. It’s the difference between designing the right thing and designing the wrong thing very well.
The next step matters. We asked the agent to find the emotional layer sitting underneath each practical frustration. A family that can’t rent a promised canoe isn’t just inconvenienced. Someone has to explain the disappointment to a child. That reframe changes what the design has to do, and it came from a prompt, not a workshop with sticky notes. I’m not being flip about the sticky notes. We have all run plenty of those workshops. The key takeaway, a genuine empathic insight grounded in evidence, arrived faster and held up to more scrutiny than a wall plastered with sticky notes. Empathy, it turns out, was never the part of the process an AI couldn’t touch. It was just the part everyone assumed it couldn’t.
Define is where you write the sentence, and the agent tries to break it
A Point of View statement is intentionally short. In simple terms, it consists of a user, a need, and an insight. Writing it is your job, not the agent’s. Handing that job to AI produces something generic, because the model has no knowledge of the client’s specific visitors. But once you’ve written it, the agent is genuinely useful as an adversary to pressure-test your POV statement. Ask it directly: “Is the need specific enough to design from?” “Is the insight actually non-obvious, or does it merely restate the research?” “Would this statement help a designer choose between two competing approaches?” This is the text sent to Gemini:
Here is a Point of View statement for the families with children visitor type for the Rabbit Blanket Lake visitor app: ‘Families with children visiting Rabbit Blanket Lake need to know — before they arrive and before they commit — whether the experiences they have promised their children are actually available, because the emotional cost of arrival-day failure isn’t measured in wasted time. It is measured in their children’s disappointment and their own guilt.’ Is the need specific enough to design from? Is the insight genuinely non-obvious or does it simply restate the research? And would this statement help a designer choose between two competing design approaches?
That last question matters most. A POV statement that can’t help you choose between two directions isn’t a statement. It is decoration. Statements that survive that pressure test become decisions in the living document. Statements that don’t survive get rewritten and tested again. That loop- write, challenge, revise- is exactly what Design Thinking has always asked for. The agent just runs the loop faster and doesn’t tire of arguing with you.
What doesn’t change, and why that matters more than what does
None of this replaces the Double Diamond. It compresses the first diamond, the half of the process devoted to finding the right problem, without cutting a single stage out of it. You still empathize. You still define. You still loop back when a discovery upends an earlier assumption, because that’s the whole point of the process being circular rather than linear. What’s different is the ratio of time spent generating options versus time spent evaluating them. An agent can produce a dozen framings of a problem statement in the time it takes to type. Your value shifts to deciding which framing survives contact with the evidence, which risk actually matters, and which assumption is worth killing before it costs anyone real money.
There’s a failure mode worth naming. These agents rewrite entire documents on update, and sections you didn’t mention can quietly disappear, replaced with placeholder text. I watched a carefully built technical constraints section vanish because an update prompt only referenced design decisions. The fix isn’t complicated. Check both files after every update, and never assume the AI preserved what it didn’t touch. Trust but verify, every time. The agent is not the project manager. You are.
Where the line actually sits.
The work that precedes design, defining the problem, synthesizing research, building empathy, writing and pressure-testing a point of view, is investigative. It is that first diamond. It has a right answer, or at least a defensible one, and that answer exists somewhere in evidence you can query. Design itself is different. The moment you start choosing a colour palette, a layout, a visual language, you’re no longer investigating what’s true. You’re deciding what should exist. That’s a judgment call with no ground truth to check it against, and it’s where a designer’s taste, experience, and accountability to the people who’ll live with the result actually belong.
Hand the first diamond to an agent, but you, the Human In The Loop, make the decisions. Design Thinking was never really about who makes decisions at each stage. It was about refusing to let a plausible-looking answer stand in for a tested one. An agent that pressure-tests your assumptions faster than you can is doing exactly what the process has always demanded of you.
It’s just doing it before lunch instead of before the client meeting three weeks from now.
Further Reading:
● “A Practitioner’s Journal on Navigating UX in the Age of AI,” UX Collective
● “Next-Gen Agentic AI in UX Design: Evolving the Double-Diamond Process,” UXmatters
● “The Double Diamond,” Design Council
● “Empathy Mapping: The First Step in Design Thinking,” Nielsen Norman Group
● “Using AI to Write a Product Requirements Document (PRD),” ChatPRD
● AGENTS.md — the open standard for giving AI agents persistent project context
● Creative AI. Generative and Agentic AI for Design Thinking and the Creative Industries.
Nothing has changed other than how we talk about it was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.
