Half of designers are now using an AI model every single week to build interfaces — and the person running the company behind the industry's biggest design tool still says the output "feels incremental." Figma CEO Dylan Field told investors this month that AI-generated visuals don't yet adhere to a design system and the experiential details aren't there. Meanwhile, Dev Mode, Code Connect and Figma's MCP server now let an AI agent read a design file and write production code directly, skipping the handoff step that used to be a designer's job to manage. And 95% of live websites still fail basic accessibility standards, a gap AI-assisted plaintiffs are now suing over at a record pace. The UX job isn't disappearing. It's splitting into "prompt it" and "fix it, make it hold together, and make it work for everyone." This article covers what AI now does well, what it still can't touch, and six moves to land on the valuable side of that split. What actually changed in 2026 AI tools are now a default part of the design process, not an experiment. UX Tools' Spring 2026 State of Prototyping survey, run by Tommy Geoco across 1,478 designers and builders in 18 regions, found 50.8% use Claude every week and 48.2% use ChatGPT every week, with Figma Make close behind at 34.8% weekly usage. Nearly 44% of respondents now spend more than half their building time working with AI-generated code. The tools moved from novelty to habit in about a year. What hasn't moved: quality at the system level. At the Goldman Sachs Communacopia + Technology Conference on September 8, Field was candid about where AI design still falls short. "For design, it feels… incremental," he told the audience. "The visuals are not quite there, let alone the experiential details, UX… adhering to a design system." That's not a competitor talking down AI — that's the CEO of the company building the AI tools, describing their own product's ceiling. He also said usage of Figma Make "went up" only after the company cut credit pricing by roughly half, which tells you adoption was gated by cost as much as by capability. At the same time, the handoff that used to define a UX designer's relationship with engineering is being automated out from under it. Dev Mode, Code Connect and the Figma MCP server let an AI coding agent open a design file, read its components and tokens, and write HTML, CSS, Tailwind or SwiftUI straight into a codebase. Bidirectional too: an agent can propose a change back into the design file. The old redline-and-handoff ritual is disappearing. What replaces it is a direct line between design intent and shipped code — and whoever doesn't understand that line gets routed around. 6 moves to stay ahead 1. Let AI own the first draft, never the final review First Draft and Figma Make can turn a text prompt into a wireframe or a working prototype in minutes. That's a genuine gift for exploration — ten directions before lunch instead of one direction by Friday. The mistake is treating that output as done. LogRocket's 2026 review of Figma's AI capabilities is blunt about it: outputs still need human review for "accessibility, semantics, and production readiness" before they touch a real product. Do this: use AI generation for divergence — the widest possible spread of first ideas — and do every convergence pass yourself. Never ship a Make or First Draft screen without a named human sign-off, the same way no agency ships unreviewed AI copy. 2. Become the design-system guardian If Figma's own CEO says AI output doesn't yet adhere to a design system, that's the single clearest signal of where a designer's judgment is worth the most right now. AI can generate a screen that looks plausible in isolation and breaks the moment it sits next to three hundred other screens built to a shared token set. Do this: position yourself explicitly as the owner of design-system integrity — component libraries, token governance, naming conventions, and the rule that nothing ships without matching the system. This is a service line clients will pay a retainer for once they've shipped one AI-generated screen that didn't match anything else in the product. 3. Make accessibility your billable specialty Ninety-five percent of websites still fail basic WCAG compliance tests, and digital accessibility lawsuits are climbing back up after a two-year dip — projected near 5,000 filings in a single year, a 20% jump, according to Seyfarth Shaw's tracking reported by Accessibility.Works. Forty percent of those filings are now pro se, often drafted with AI help identifying violations, which means the cost of finding a violation has collapsed even as the cost of having one hasn't. Accessibility overlay plugins don't fix this: sites with an overlay installed accounted for 22.6% of all filings in the first half of 2025. Do this: offer a standalone WCAG audit and remediation service, separate from general design work. AI can generate an interface fast; it cannot certify that interface is usable by someone with a screen reader, and it does not stand behind that certification when a demand letter arrives. A human who can is now selling something genuinely scarce. 4. Learn the Dev Mode and MCP handoff, don't wait to be routed around it When an AI agent can read your Figma file and write the component straight into a live codebase, a designer who only hands off flat images or specs becomes a bottleneck an engineering team will quietly work around. A designer who understands Code Connect, how tokens map to code, and what an agent needs from a file to generate clean output becomes the person both sides still need. Do this: set up Code Connect on your next project even if no one asked for it, and learn to structure files — named layers, real components, consistent tokens — so an AI agent reading them produces usable code on the first pass. That fluency is now part of the job description, not an optional extra. 5. Own the experiential polish AI still can't reach Field's own phrase for what's missing is "the experiential details" — the motion, the timing, the feel of an interaction, the empty state that doesn't feel like an error, the edge case a generic prompt never considers because it was never told the business actually has one. These are the details a user notices without being able to name, and they're exactly what a fast AI draft skips. Do this: build a portfolio section specifically around interaction and motion decisions — before/after micro-interactions, error and empty states, onboarding flows tuned to a real audience. Make the case in client conversations that this layer is what separates a product people tolerate from one they enjoy. 6. Price for judgment, not for screens When a founder can generate a passable first draft in Figma Make for a fraction of what it cost a year ago, per-screen or per-deliverable pricing is a losing frame — you're competing with a tool your own client has access to. What a tool can't price is accountability: the design-system call, the accessibility sign-off, the user-research finding that changes the roadmap. Do this: repackage services around outcomes and ownership — a design-system audit and governance retainer, a WCAG certification package, a research-and-validation sprint with a documented recommendation. Quote the judgment, not the pixel count. If you're a business owner: how to hire for this Ask for Red flag A stated process for reviewing every AI-generated screen against the design system A portfolio of only AI-generated mockups with no shipped, working product A WCAG accessibility plan, named explicitly, not "we'll add an overlay" "Accessibility isn't really our thing" or reliance on an overlay plugin Familiarity with Dev Mode, Code Connect, or your team's handoff tooling Deliverables as flat images only, no components or tokens Evidence of user research or testing behind past decisions Speed promised with no mention of validation A pricing model tied to outcomes, not a per-screen rate card Pricing that assumes every screen takes as long as it did in 2023 The failure mode to avoid: a founder who generates an entire app's worth of screens with AI, ships it without a system pass or an accessibility check, and discovers both problems at once — an inconsistent product and a compliance letter — months apart. The winning setup is a designer who uses AI to move fast on exploration and is personally accountable for everything AI can't yet verify. Finding the AI-era UX designer: Giggrabbers Whether you need a first product built from scratch, an existing app brought into line with a real design system, or a WCAG audit before a launch, Giggrabbers connects businesses with vetted freelance UI/UX designers. Its AI-powered tools help scope the brief and generate an early wireframe to react to, and its built-in crowdfunding lets you fund a fuller redesign without waiting on a budget cycle. For designers, it's where the six moves above become billable work: founders arrive having already seen what an AI draft can do, and needing exactly the system thinking, accessibility accountability, and polish it can't. The bottom line AI didn't replace UX designers; it made the first draft free and left everything after the first draft more valuable than before. Generating a plausible screen is now nearly instant. Making four hundred of them hold together, making the product usable by someone AI never thought to test with, and making the small moments feel right — that's the scarce layer now, and it's the one machine output still can't sign its name to. Designers who make the six moves above, and the businesses that hire for them, are positioned exactly where 2026's tools stop. Sources Investing.com — Figma at Goldman Sachs Communacopia + Technology Conference 2026 transcript (September 8, 2026), remarks by CEO Dylan Field LogRocket Blog — Figma AI in 2026: everything it can do, and what it still can't UX Tools, Tommy Geoco — State of Prototyping, Spring 2026 (1,478 respondents, March 14–April 6, 2026) Accessibility.Works, David Gibson — AI Is Fueling a New Wave of Accessibility Lawsuits (January 9, 2026), citing Seyfarth Shaw ADA Title III filing data