· 4 min read · AI, UX, Process
Can AI automate the UX process?

I tried to automate the UX process with AI, from a project overview all the way to user flow diagrams. I knew the idea was flawed. I wanted to see how far it could go anyway.
What I built
It is an eight-step flow. Each step has a prompt. The prompt reads the documents from earlier steps and writes a new document. Human input is optional, and the goal is to keep it as small as possible.
The steps are:
- Project overview
- Stakeholder insights
- UX strategy
- Competitive analysis
- User research
- Sitemap and content strategy
- Use cases and user scenarios
- User flow diagrams
Some steps feed extra information into later ones, and the paths join at the end. The whole thing is in a Figma file. There is also a project page.
What worked and what didn't
The AI is good at summarizing briefs. It finds patterns, makes first frameworks, structures content by common practice, and draws the usual user flows.
It does not understand the details of a company. It struggles with emotion and body language. It cannot check how real users behave. It cannot empathize or run a meaningful interview. So full automation did not work.
The idea is not new
Splitting a task into prompts that pass output along is called prompt chaining. Wu, Terry and Cai described "AI Chains" in 2021. In their 20-person study, chaining improved results and made the system feel more transparent and controllable. LangChain built tooling around the same idea, including making several model calls in sequence to break a task into steps. And Wei et al. showed in 2022 that showing a model its intermediate reasoning steps improves results on reasoning tasks. My prompts ask for a reasoning path too.
One benefit of a chain is that you can read every document in the middle. When step 6 looks wrong, you can check steps 1 to 5.
What others found
The design tools of 2024 are not great. In April, Nielsen Norman Group found that few design-specific AI tools meaningfully improve UX work. The tools they tested gave generic layouts. Designers mostly used text tools like ChatGPT for brainstorming and copy. NN/g also made a short video in May saying the same. Figma showed AI layout generation at Config in June, but pulled the feature a few weeks later because it produced designs that resembled existing apps. Uizard has offered text-to-UI since 2023. These tools try to automate the pretty end of the process. My flow starts at the thinking end.
On research, NN/g tested "synthetic users" in June. They found the answers shallow and too positive. Synthetic users said they had finished courses that real learners abandoned. Their advice is to treat such output as hypotheses to test. Their line: "UX without real-user research isn't UX."
Research on personas is a bit kinder. A 2024 paper, Persona-L, found that UX designers felt the tool built empathy, but they wanted more transparency about the data behind the personas.
My view
My step 5, user research, is the weakest one. It synthesizes what is likely true about an audience. That is not research. It is a good place to start a conversation and a poor place to end one. The NN/g work and the persona paper fit my own finding: the model cannot validate real behavior. You only learn that by testing with real people.
This is my opinion: the useful part of the flow is the paperwork. Briefs, strategy drafts, sitemaps and first user flows are things designers often start from a blank page. A chain can give a first draft in minutes. A person should then fix it with real knowledge of the client and the users.
So I see AI as an assistant that augments the designer. It does not replace them.
OpenAI released its o1 reasoning model on September 12. It is designed to spend more time computing before it answers. I think models like it will increase what can be automated here, mostly in the middle steps that need multi-step logic. I expect the real-user part to stay human.
Sources
- Figma file of the whole process - the eight-step flow, by me
- UX Automation project page - my write-up of the experiment
- AI UX-Design Tools Are Not Ready for Primetime (NN/g, 2024-04-12) - tests of AI design tools and how designers used them
- AI Isn't Ready for UX Design (NN/g video, 2024-05-13) - the "as of Spring 2024" verdict
- Synthetic Users: If, When, and How to Use AI-Generated "Research" (NN/g, 2024-06-21) - why AI-made users cannot replace real ones
- Usability Testing 101 (NN/g) - why testing with real users is the way to validate a design (published 2019)
- Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (Wei et al., 2022-01-28) - intermediate reasoning steps improve results
- AI Chains (Wu, Terry, Cai, 2021-10-04) - chaining prompts, with a 20-person study
- LangChain Expression Language (LangChain, 2023-08-01) - chaining LLM calls as an engineering pattern
- Persona-L has Entered the Chat (Sun et al., 2024-09-23) - LLM personas evaluated with six UX designers
- Config 2024 in review (Figma, 2024-06-26) - Figma's first AI features
- Inside Figma: a retrospective on Make Designs (Figma, 2024-07-17) - why the feature was pulled
- Uizard announces Autodesigner (Uizard, 2023-03-07) - text-to-UI tools
- OpenAI releases o1 model with human-like reasoning (Fortune, 2024-09-12) - the o1 release and its reasoning approach
Comments