Two AI Workflows, One Default You Didn't Choose

AI is pulling design tooling in two directions at once. One camp keeps the canvas and teaches the machine your rules. The other throws out the canvas and lets you talk your way to a screen. Both are moving fast. And while your team argues about which to use, a lawsuit is asking a colder question: who agreed to let any of this train on your work?

Let me catch you up on what actually changed and what to bring to your next review.

The canvas or the conversation

Here is the fork. In the old flow, Figma sits at the center: you research, sketch, wireframe, polish, prototype, and control every layer by hand. In the new flow, you describe what you want and the tool builds a first version you refine by talking to it. Same goal, opposite starting point. One gives you control over every pixel. The other gives you speed and a rough draft in minutes.

This is not really a tool choice. It is a workflow choice, and it changes what your designers spend their day doing. Pick the canvas and they keep their hands on the craft. Pick the conversation and they become editors of a machine's first pass. Most teams will run both, but you owe your team a clear answer on which one owns which stage of the work.

Feeding the machine your own rules

The teams getting real value are not asking AI to design from scratch. They are teaching it how their company already designs. Yu-Ching Lin fed years of a company's UX guideline into an LLM and turned it into a wiki you can talk to. A question that used to cost ten minutes of digging now takes three seconds, with a reference link so you can check the answer.

Then it did something nobody planned. It started making judgments. Ask it how to notify a user after a bulk delete, and it walks the tree: a toast with Undo if the action reverses, a modal for a second confirm if it does not. First-draft prototypes now land at 70 to 80 percent usable, and work that took two weeks compresses into two days.

Same idea, smaller scale, with reusable "skills." Chetan Singh got tired of retyping his standards every project, so he built two Markdown files that hand the AI his tokens, naming rules, and audit steps up front. The payoff was consistency, not speed. Work came back already matching his standards instead of needing correction.

The rules are only as good as the person who keeps them

Here is the part that should shape your headcount thinking. Feeding AI your guideline does not shrink the designer's job. It moves it up. Lin lists what stays human: defining the real problem before the tree starts, inventing new interaction patterns, sitting across from a confused user, and taking the room's political heat when engineering says a change costs two more sprints.

And someone has to keep the guideline alive. New components appear, old rules start to contradict each other, the design language shifts. The stronger the AI gets, the more that maintainer matters, because everything downstream now runs on their work. If your guideline is a stale PDF, this whole approach breaks. That maintainer is a role, not a favor someone does on Fridays.

The setting you didn't flip

While you sort out workflows, a harder question landed in court. In November 2025, a founder sued Figma over consent, claiming the company switched on model training over design files by default after years of promising otherwise. Figma disputes it and says its training targets general patterns, not customer content. Nothing is decided. But the design decision is already public, and it is not really about the law.

The pattern rhymes. Zoom in 2023, Slack in 2024, Adobe in 2024. A policy lands with little notice, enrollment is the default, someone outside the company finds it, then comes the walk-back. Zoom and Adobe reversed inside two weeks. Figma walked nothing back, because there was no drafting error to fix. The default was announced, defended, and allowed to take effect.

What makes it sting is the split. Figma's own explainer says training defaults to on for Starter and Professional, off for Organization and Enterprise, because the enterprise contracts carry negotiated terms. Two defaults along a payment line. The customers with a procurement team got protected. The freelancers under the same client NDA got enrolled.

The deep cut

You ship defaults to your own users every quarter, and this is the test your users will one day run on you. Neeman offers four questions: consent, symmetry, disclosure, exit. Symmetry counts more than the other three combined, because a single default for everyone can be a philosophy, but two defaults split by who pays is a statement about whose work you think is worth protecting.

The concrete move before your next renewal: open every AI tool your team uses, find the training setting, and look at its state. If it is on and nobody turned it on, the vendor decided for you. Write down what you find for each tool, then decide which ones you actually trust with client files. Twenty minutes now beats a surprise in discovery later.

Three questions for your team

1. For each stage of our process, which workflow owns it: the Figma canvas or the conversation with AI, and who decides when a tool crosses from draft to shippable?
2. Do we have a living, machine-readable version of our design guideline, and who owns keeping it current now that AI depends on it?
3. What is the training default on every tool we use with client work, and can we tell a client with a straight face that their files are not being trained on?