What's the concrete criterion behind this "narrow works, broad doesn't" dividing line?
The criterion isn't whether a task sounds simple or complex — it's how concrete the input and output are, and how easy it is to verify right from wrong. Renaming a layer takes the existing layer structure as input and produces a new name as output; a user can tell at a glance whether it's right. But generating an entire wireframe takes a vague text description as input and produces a complete screen involving information hierarchy, Visual Hierarchy, and interaction logic — the question "is this design right" doesn't even have a single clear-cut answer; it requires design judgment to evaluate.
In other words, a task's "verifiability" is the real dividing line, not the size of the task.
Why does Figma's "First Draft" feature only produce minor variations no matter how the prompt is adjusted?
Another limitation NN/g's report points to is that prompt character limits (500 characters, for example) simply aren't enough to describe the full context a complex screen needs. That means even if a user works hard to refine the wording of a prompt, the actual volume of information that can reach the model is capped by that limit — details get forced out, and the model can only fill in the missing information based on what it learned during training about "what kind of generic screen this type of description usually corresponds to," which naturally tends toward conservative, generic output with a weaker information hierarchy.
This is a separate issue from how capable the model itself is — no matter how strong the model's capabilities, insufficient information on the input side structurally caps the diversity and precision of the output.
What specifically does "no AI tool effectively supports design systems" mean, and how does it affect teams?
A Design System's core value is consistency — the same button component appearing on any screen should follow the same style rules, and changing it once should propagate to everywhere it's used. Most current AI generation tools are optimized for a single generation task (generate this one screen, this one component), not for "making sure this generated result stays consistent with the rest of the existing design system" — these are two different technical problems, and good generation quality doesn't automatically mean compliance with system rules.
The real-world impact on teams: if your team already has a mature design system, content generated by AI tools will almost certainly need an extra manual "align with the design system" check step — you can't assume generated output will automatically follow existing rules. This is also why the off-brand risk and AI Design Hallucination issues mentioned earlier tend to get especially amplified at teams that have already adopted a design system.
Now that I know this dividing line, how should I actually adjust how I use AI design tools?
The concrete approach is to first inventory which of your team's everyday tasks are "clearly bounded, easy to verify right from wrong" — renaming things, generating placeholder text or images, searching existing assets for similar components — and prioritize handing those to AI, since that's the time savings you can relatively reliably count on.
For broad-scope tasks like "generate a complete screen," it's worth positioning AI output as a brainstorming starting point rather than a finished product you can ship directly — use it to quickly generate a few directions for the team to discuss, rather than expecting it to hand you a launch-ready design in one pass. At the same time, no matter what gets generated, build in a manual check step against Design System rules, since no current tool automatically guarantees that. None of these three adjustments require switching tools or waiting for a technology upgrade — they're usage changes you can implement immediately with the tools you already have.
Almost every design tool on the market is promoting its own AI features, but a 2026 hands-on evaluation of several mainstream AI design tools by the user experience research firm Nielsen Norman Group (NN/g) found a clear dividing line: narrow-scope, single-task AI features generally work well, while broad-scope features that try to generate an entire screen in one go generally fall short of expectations. That dividing line is worth taking seriously, because it isn't a problem with any one tool — it's a consistent pattern observed across tools.
In NN/g's testing, the features that performed well share a common trait: clearly bounded tasks with concrete inputs and outputs. Figma's layer-renaming feature, for instance, "completely eliminates the tedious task of renaming layers." A text-generation feature (like Figma's "Rewrite this") helps designers who aren't primarily writers produce a usable draft. An asset-discovery feature like "Find more like" helps teams quickly locate similar components across a large pool of design files. The color-palette tool Khroma Color uses pattern recognition to build custom palettes. AI image generation (Midjourney, for instance) used to produce placeholder visuals in prototypes also proved quite practical. The shared logic behind all of these: AI is handling a small task with a clearly bounded scope that's easy to verify as right or wrong, so users can readily judge whether the result works.
By contrast, features that attempt to generate an entire screen or prototype in one shot consistently underperformed their marketing. NN/g's report states plainly that wireframe and prototype generators "do not meet expectations," and in practice are useful at best for the "ideation" stage, or as a starting point for newer designers — not for producing a usable final result directly. A more specific example is Figma's "First Draft" feature — testing found it tends to produce "generic" designs with weak information hierarchy and Visual Hierarchy, and no matter how the prompt wording was refined, the output showed only "minor variations," unable to produce meaningfully different results for different needs. The report also points to a more fundamental limitation: no current AI tool "effectively supports design systems," a significant constraint for professional design teams, since a Design System is exactly the foundation for team collaboration and consistency.
The report mentions a technical detail that's easy to overlook: many tools' prompt character limits (500 characters, for example) simply aren't enough to describe the full context a complex screen needs — which components should appear, how they interact with each other, how styling differs across states. That volume of information easily exceeds the character cap, forcing users to omit detail, and those omitted details are part of why the generated output "looks generic." This is a completely different situation from narrow-scope tasks: renaming a layer doesn't need much context, but generating an entire screen does, and a tool's input constraints often fail to keep pace with the volume of information the task actually requires.
NN/g's overall conclusion is fairly measured — the job-displacement risk designers face is "nowhere near" as high as previously feared, and AI design tools are only "marginally better" than a year ago. The implication behind this conclusion is that AI design tools' value right now is concentrated in assisting and accelerating single tasks, not replacing a designer's overall judgment of a screen or systematic thinking. Positioning AI's role as "handling small, tedious, repetitive tasks" rather than "directly producing the final design" is currently the usage pattern best supported by the test results.
If your team is evaluating whether to adopt an AI design tool, or already has one and feels the results fall short, this test gives a concrete direction to check: first confirm whether the task you expect AI to handle is "narrow scope, clearly bounded" or "broad scope, produce a complete screen in one go." If it's the former, existing tools can likely genuinely save you time. If it's the latter, the gap you're seeing is a widespread pattern, not a sign your team configured something wrong or picked the wrong tool — adjusting expectations and treating AI output as a starting point rather than a final product is more practical than continuing to search for "the one tool that finally gets it right in one shot."