At DoorDash, if a designer modifies an interface in the code, the change goes through the same GitHub approval pipeline that engineering uses. Before it ships, there’s also a design review. Shali Nguyen describes that workflow in AI in Design 2026, a report from the investment firms Designer Fund and Foundation Capital. The case shows what happens when more people can implement a change: someone still has to decide whether it’s ready to go out.
As of November 9, 2026, the strongest evidence of time savings comes from tasks where the outcome was already defined. Company case studies show agents at work on prototypes, components, and code. To know whether a team is saving time across the whole process, you’d need to track those changes from problem definition through user validation and product maintenance.
What is agentic design?
Here we’re talking about systems that receive a goal, act on design tools or code, observe the result, and step in again. The degree of autonomy varies by tool and by task. The term “agent” also comes up when a product lets a user delegate a purchase, or when another system operates its interface. Each case raises different questions.
| Use | What the system does | What’s worth evaluating |
|---|---|---|
| Generative assistance | Produces text, images, or variants | What can be used as-is and how much editing work remains |
| Agentic production | Modifies components, code, or prototypes and adjusts the result | How many tasks get completed and how much work review and fixes require |
| Products that incorporate agents | Let the user delegate tasks | What the user understands, what permissions they grant, when they confirm, and how they regain control |
| Agent-operated interfaces | Expose content and actions to other systems | What those systems can access and what operations they can execute |
This article focuses on agentic production. Figma Make, Figma’s prototyping tool, is a different product from the platform’s agent and from coding agents. The trial we’re examining measured Make.
Who’s already using agents to design?
In its third-quarter 2026 earnings report, Figma said that, as of the end of October 2026, 65% of paid accounts contributing more than $10,000 a year in recurring revenue used its agent weekly. The figure describes a segment of the platform’s customers, defined by what each account pays Figma. Weekly frequency indicates that those accounts keep coming back to the tool. The release measures usage and leaves open the question of how much work they’re resolving with it.
In the same release, Figma said usage of its MCP server tripled compared to the previous quarter and that write operations in Figma quadrupled. MCP is a protocol that connects models with tools and data. The company published how much that activity grew. To know the scale at which it’s used and with what result would require the absolute volume and the task-completion rate.
Meanwhile, Designer Fund and Foundation Capital surveyed 906 designers from more than 60 countries in March 2026. According to the chapter on design practice, 50% said they had shipped AI-generated code to production at least once. The share was 68% among those working at early-stage companies and 33% among those at publicly traded companies. In the chapter on tools, 62% cited inconsistent or unreliable results as their main difficulty.
According to the report’s methodology, the survey circulated through social media, newsletters, and communities. 60% of respondents had more than ten years of experience, and the authors acknowledge the sample may have skewed toward people more inclined to use AI. The responses are useful for understanding what those designers did. The gap between startups and public companies leaves open a question: how much of it is explained by permissions, risk, and approval processes at each type of company. The report includes more than 25 conversations with professionals and has commercial partners such as Anthropic and Framer. Its authors say they designed and analyzed the survey independently of those partners.
How much time did Figma Make save in the trial?
On tasks that were completed, participants with access to Figma Make took roughly 20% less time. The benefits were greater for product managers and, among designers, varied by task. The finding comes from a preprint by four Figma researchers published in September 2026, ahead of any peer review.
The research firm MeasuringU recruited and moderated the 100 participants. Half were product designers and half were product managers. All had at least two years of experience and used Figma at least once a month. They worked on the same interface, in the same order, with a reference image for each exercise. First they had to switch the interface to dark mode, then add an option to the help and settings menu, and finally build an interactive comments panel.
One group could use Make and edit by hand. The other worked without AI. Moderators checked how closely the result matched the reference. In the comments panel exercise, access to Make raised the completion rate by 19.9 percentage points. That figure matters on its own, and it also changes how the time savings should be read. When a tool lets more people finish, the average among those who completed the task can end up comparing different groups.
Participants knew in advance what they had to build. That makes it possible to study how long it takes someone to execute a defined change. To evaluate a product’s design work, you’d also need to look at how the team chooses which problem to solve, what it does when there’s disagreement, and how it maintains the interface after launch. In the trial, quality was assessed mainly through participants’ perceptions and how closely the result matched the reference. A panel that looks like the one in the image says little about whether a person can use it with a screen reader or recover from an error.
The corporate summary from August and the September preprint use different statistical specifications and publish different aggregate figures. Within the preprint itself, the discussion section and a table disagree when presenting an aggregate result for designers. Hence the “roughly 20%.” A more precise figure would require clarification from the authors.
How do teams bring those changes to production?
In the chapter on tools in AI in Design 2026, Katie Dill describes two internal tools at Stripe. ProtoDash incorporates the company’s design system, and Dante integrates with content work. When components and rules are already available, the team can reuse them for each request. Stripe is a partner on the report. Its account explains how it organized that work, and it leaves open the question of measuring the economic return and long-term quality.
In the report’s chapter on design practice, David Stinnette says his team at Samsara built prototypes directly in the product’s code using customer data. That let them observe states and edge cases in an environment closer to what the customer uses. At Watershed, according to Hannah Hudson, every designer submitted proposed changes to production and relied on engineers to review the code and learn from it. DoorDash kept engineering approval and added a design review on top. These accounts show how these teams divided the work. Evaluating the effect on the product would require data on which changes shipped and what resulted from them.
Anita Kirkovska says that, during Vellum’s website migration, the marketing team started producing pages using coding agents. The challenge was reproducing a design precisely and maintaining consistency across pages. It’s easy to imagine how that problem shows up when several people can publish on their own. One decision is which tool to use to build and launch an AI-assisted site. Another is who decides when a page is allowed to deviate from the shared components.
What happens when the team is still defining the problem?
In a study presented at CHI 2025 with 280 designers, Shin and colleagues compared three ways of using language models against a manual process. Under the study’s conditions, the models did not improve the quality of problem reframing. The researchers also observed that the gap widened between more and less experienced professionals, and that less experienced participants felt they had less control over the process. A team can reach a functional interface sooner and still be working from a flawed diagnosis.
What the team explores before choosing a direction also matters. In a 2023 experiment published at CHI 2024, Wadinambiarachchi and colleagues gave 60 participants twenty minutes to sketch avatars. One group worked unassisted, another used Google Images, and the third used Midjourney V4. Those who used Midjourney tended to stick closer to the initial example and produced less varied, less original proposals. The sample had little professional experience, and the result is specific to that tool and that exercise. In a current project, you can test for the same risk: if every variant starts from the same idea, ask for an alternative that changes one of the solution’s assumptions.
Tanner Kohler of Nielsen Norman Group draws a similar distinction in his analysis of agentic development workflows, published in October 2026. A team can give the agent design rules and verify that the code compiles, uses the intended components, and follows the spec. But to know whether a person understands the interface, completes the task, or recovers from an error, you need to watch what happens when they actually try to use it.
What changes when the agent is part of the product?
In a qualitative study by Nielsen Norman Group, six Chinese participants tested Qwen’s ordering and booking features. When asked to buy tea, five of them started with traditional apps instead. During the study, participants had trouble finding some features, delivery addresses showed up where they didn’t expect them, and information about final prices or luggage was missing. Some abandoned the process.
With six participants, the study’s value lies in the problems it surfaces during use. For example, before an agent places an order, a person should be able to understand which address it will use, how much the purchase costs, and at what point they can stop it. These are product decisions the team has to make regardless of how quickly it can generate the interface.
How to evaluate an agentic design pilot
| What might explain the improvement | What we saw | What’s still unmeasured |
|---|---|---|
| The agent speeds up well-defined changes | The Make trial recorded savings on narrow tasks, and more participants finished the third exercise | Total time to production, including review and fixes |
| Company components and rules make the work easier | Stripe built its design system into its internal tools | How much that context contributes to the result |
| Part of the effort shifts from producing to reviewing | DoorDash added a design review, and Vellum flagged visual consistency problems | How much effort the team gains or loses overall |
The Make trial offers the most direct support for the first explanation. The other two emerge from how the companies cited organize their work. With the data available, it’s still unclear what the effect is on a team’s budget, or on adoption and employment in markets like Argentina and Spain.
Pick a recurring task with an outcome the team can verify. Compare how it’s resolved with and without an agent, from preparation through publication. Track how long it takes to generate the first version, review it, and fix it. The key measure is how long it takes the team to get a change it can approve.
Check whether the change respects the design system and reuses the intended components. Test accessibility and watch what happens when a person tries to complete the task or fix an error. Define who can approve and revert the change, and make sure technical permissions reflect those responsibilities. If the problem is still open, set aside time to revisit the diagnosis before committing to a solution.
For a design studio, that means budgeting for research, components, and maintenance alongside production. For a branded media team maintaining its own digital experience, it means measuring both publishing speed and the ongoing work of keeping things consistent as pages are added.
What would change our conclusion?
Our hypothesis for the next six months is that some teams will be able to ship more changes if the agent works with the organization’s components and rules and someone is responsible for reviewing the result. When the problem is still open, generating faster can also leave more fixing to do.
On May 9, 2027, we’d revisit this hypothesis with data from complete projects. We’d be looking for savings that hold up after accounting for reviews and post-launch changes, along with sustained outcomes for users. If the savings reach production and accessibility, comprehension, and maintainability hold up, there would be more reason to expand what gets delegated. If regressions and rework increase instead, then a faster first version only explains part of the story.
Frequently asked questions
What is agentic design?
It’s the use of systems that receive a goal, modify a design, prototype, or code, and adjust their work based on the outcome. Autonomy depends on the tool and the task.
Do UX design AI agents save 20% of the time?
In the Figma Make trial with 100 participants, savings were around 20% among completed tasks. These were three interface changes with visual references. Estimating savings for a full project would require accounting for everything from problem definition through maintenance.
What does it mean that 65% used Figma’s agent?
As of the end of October 2026, 65% of accounts contributing more than $10,000 a year in recurring revenue to Figma used its agent every week. The figure measures usage frequency within that segment.
Does AI help reframe design problems?
In Shin and colleagues’ study of 280 designers, the models tested did not improve reframing quality. The gap between more and less experienced professionals widened under the study’s conditions.
How do I evaluate an AI prototyping pilot?
Compare the same task with and without an agent and measure how long it takes until the team can approve the change. Include review and fixes. Test whether the interface follows the design system, is accessible, and lets a person complete the task and recover from an error.