Every new leap in AI coding tools restarts the same panic in product teams: if agents can already write software, is AI design automation next—and should designers prepare to learn a different job? The fear is understandable. Tools now generate screens, variations, and even motion. The instinct is to assume design is about to get replaced the same way implementation is getting accelerated.

To dive deep into this, I sat down with Loredana Crisan, a leading authority in product design as Chief Design Officer at Figma. She joined Figma at the end of September after almost a decade at Meta, where she led messaging for Facebook and Instagram DMs and a consumer GenAI product for the superintelligence lab. She has been a designer, and she has also led teams across product, engineering, and data science. Her read is clear: AI is transforming how we work, and the companies that build our tools have a responsibility to help us enter that era without giving away control.

---

Why AI Design Automation Can’t Replace Taste

Design will not get the same full-automation moment as code, because design is the brand, the point of view, and the part users actually experience.

Code and design are not the same kind of work, so they should not be automated the same way. Before agents, developers spent a lot of time on abstractions and on the language the code was written in. Users never experienced that layer. They experienced the design. That is why it is easier to let agents replace something people never really saw.

Design is harder to automate because it is supposed to represent your brand, your point of view, and who you are as a builder. In that sense, code is disappearing as a thing people stare at all day—you will hear developers say they have not looked at code in a long time—while it is becoming a creative medium. Designers are already playing with code to invent new experiences. That is not code as implementation. It is code with a point of view: creating with it, not just shipping it.

A lot of code got automated away. At the same time, that automation elevated code as a material designers can actually play with. The same moment is not fully possible for design. Human imagination is too large a part of the job for it to be automated the way implementation is being automated.

Writing is a useful parallel. AI has been able to produce creative writing for a while. It is still far from matching human imagination, and people generally do not want to consume content that feels generated. They want something that feels authentic, directed at them by a person who cares. Design has the same properties. Intention cannot be fully outsourced. Some tasks can be automated. The point of view cannot.

Code as a Design Material, Not a Handoff

Use AI as scaffolding for your imagination: let it generate options and grunt work, then keep your hands in the material for the parts you cannot describe in a prompt.

If some of the job can be automated, the useful question is which parts—and who stays in control. Figma’s agent, released about a week and a half ago, one day before Config, is built to help with design tasks you might want to explore rather than to take the file away from you. It can create variations so you can look at more options. Loredana is explicit that looking at many ideas helps creativity. It can take on work many of us enjoy less, like setting up components. It can give feedback. It can look at a design through the lens of a different person: a synthetic persona of a user, or a simulation of what a CEO might say. It can organize layers. In practice, it can help with basically anything a human could do on the canvas.

Having that power is not the same as giving it away. The agent is there the entire time as scaffolding for your imagination. Motion, introduced at Config two weeks ago, is the clearest example. The agent can set up motion for your layers. You still go in and control the keyframes and the easing until the animation feels right. That last part is experiential. You cannot describe it. You cannot prompt “I want this ease curve that will feel like that” and get a coherent result. A lot of design work is discovery. You do not yet know what you want, so you cannot only play with prompts.

The same logic applies to code as a design material. Figma also brought code layers onto the canvas, so designers can play with code, data, and interactivity in the file they already know. For Config demos about helping someone discover events, a few people created about 20 directions in around an hour—from a scroll, to a wheel, to an S-curve that made it into the keynote. Doing that in code, on the canvas, with the process designers already use, was revelatory for the team.

That matters because a split workflow kills flow. When Figma Make shipped a year or two ago, Design lived on one canvas and Make on another. They could communicate, but they were still two places. Code layers start to collapse that into one canvas with both opportunities. Keeping people in the flow is not a nice-to-have. Every extra transition between states or file types is a wall. Hit that wall often enough, and you simply stop. The direction is to bring more materials onto the canvas without making it overwhelming. Because the agent is a simple prompt box, it invites you to ask for anything—and the canvas has to be able to operate on any material.

Shaders are the extreme case. They are painful to manipulate by hand and require a lot of technical knowledge. With a prompt, you can iterate on them and drop an advanced material into a canvas that is still easy to use. The canvas could always hold that complexity. What was missing was the code and the math. The agent scaffolds that work, and it does not just dump a finished shader on you. It gives you parameters to play with. A shader is an indescribable thing. You are trying to replicate a material, but what feels right, what fits the brand, and what fits the moment is something only you know. You still need tools that let you express that.

Product Design for Non-Deterministic Systems

The designer role expands into more materials and into non-deterministic UX—while taste, empathy, and a grounded point of view become the scarce skills.

This is an unusually good moment for design, because you can command more materials and bring more of an idea to life. Shipping a product used to require many specialists in many tools. AI can now scaffold some of those areas so an idea can travel further toward production. Designers will keep expanding how they think, and that requires more fluency across mediums. The tools themselves are less of a barrier: agents can help you execute even if you do not yet know how to control something precisely.

You still have to be in the material. That is how you develop a way of seeing, taste, a worldview, and the empathy that makes a product actually connect. These tools give almost anyone superpowers. AI on its own tends to converge toward the mean. It is a pattern-recognition system trained on a lot of data, and RLHF pushes it toward the safe side. Designers are the ones who keep expanding the thinking—and expanding culture. The move is to embrace the new materials, play, and have fun, while keeping your point of view deeply grounded and expressing it in everything you ship.

The other shift is non-deterministic UX. Almost any experience will soon have a human way to use it and an agentic way to consume it. You might browse a website and read it yourself, or you might ask your agent to browse it and take an action. Designers are responsible for both. As agentic experiences become common, verification becomes part of how you specify what the experience should be. That looks like caring about datasets and evals in a considered way. Those skills were mostly outside the design discipline until now. Some of them did not exist. They are going to matter more.

That is already showing up in how design teams work. A lot of Figma’s designers ship code. An AI survey they released showed that the number of designers who ship code has generally doubled in the last year. Internally, that exploration becomes roadmap: they build Figma in Figma, so the directions designers start playing with become product. They have also gone deep on evals and datasets while post-training proprietary models so the agent can design in Figma, and while making sure third-party agents produce good results. The bar is quality. AI should not just design for you. Design is for people. When an agent scaffolds an action, it still has to do that action well.

Adoption inside a team is not uniform, and that is fine. Some people explore ahead. Some people move slower. Both behaviors have good reasons. What became essential is sharing. On the canvas, you can see other people’s AI conversations. A conversation with an agent is not only an output. It is an instruction. When designers open each other’s files and see how someone used an agent for critique, feedback, or variations, that becomes shared knowledge. The same need produced shareable tools: if you want to repeat an action, and you want teammates to run it too, the agent can help you create it. That is how generative plugins became a product—an avenue to share vibe-coded tools with each other.

Exploration still needs a coherence bar. AI makes it easy for everyone to wander in a different direction. The canvas is how you pull the team back into a co-creation space so the product still makes sense. Share prompting techniques. Share tools. Bring the work into the space where the team actually operates. Give people access to materials and encourage play. Then be strict about selection. You can explore a lot. You should not ship everything just because you can. The end user still walks into one product, and that product has to hold together.

If you want something to try immediately, Loredana’s advice is practical. Try the different tools. Learn the techniques and the combination that actually lets you express what you want. Then go further into non-determinism: play with model outputs, define what “good” is, and learn how you would create a skill in an agent loop so the agent produces the result you intend. That is a new form of design, and it will keep growing. The non-negotiable is simpler: do not ignore your point of view, and do not give it up. That is what makes you special. Lean into it with the new tools.

---

Final Thoughts

AI design automation will take grunt work, generate options, and scaffold materials we could not previously touch. It will not replace the part of the job that is taste, brand, and a human point of view. Keep your hands in the material. Design for both the person browsing and the agent acting. Share what you learn so the team stays coherent. And keep putting soul into the work.

Loredana closed with a line from the early days of techno, when some people argued the computer meant it was not music. A well-known European musician answered: do not blame the computer if your music has no soul. If it has none, it is because nobody put soul into it. The same is true here. People still need to feel that someone cared. Otherwise the world just gets a bit more gray.

If this framing is useful on your team, share how you are keeping taste in the loop—or connect with me on LinkedIn. I am always up for a sharp conversation on digital product innovation.

---

FAQs

Will AI fully automate product design the way it is automating code?

No. Code was a layer users never experienced, so agents can replace more of it. Design carries brand, point of view, and the experience itself, so human imagination remains too central to automate away in the same way.

What should a design agent do versus what should you still do by hand?

Let the agent scaffold variations, component setup, feedback, personas, layer organization, motion setup, and even shaders with playable parameters. Keep control of the parts you cannot describe yet—keyframes, easing, brand fit, and the feeling you only discover by working in the material.

How is the product design role changing in the AI era?

Designers need fluency across more mediums, including code as a design material, plus new skills in non-deterministic UX such as evals, datasets, and verification for agentic experiences. Taste and a grounded point of view become more important, not less, because AI on its own converges toward the mean.