AI for UX · Step 1 · 7 min read
AI tools landscape for designers
A beginner's map of the AI tools designers meet in late 2026: the five kinds of tool, which one fits which design task, the data and security questions to ask, and a simple way to decide whether a tool is worth keeping.
There is no "best AI tool". There are tasks you do every week, data you are allowed to share, and tools that fit some of those tasks. This note gives you a map of the kinds of tool that exist, what each is good for, and the questions to ask before you put real work into one.
What you'll be able to do after reading:
- Recognise the five kinds of AI tool designers use, and what each is for
- Match a design task (a prototype, a synthesis, a stakeholder deck) to the right kind of tool
- Ask the data, security and cost questions that matter before using a tool for real work
- Run a small, fair trial of a tool instead of deciding by hype
The map reflects the market as of November 2026. Tools change every few months: treat names as examples and check the vendor's page before you recommend anything.
1. Five kinds of AI tool
Most AI tools for design fall into one of five groups. Knowing the group tells you more than the brand name.
Edits work where it already lives, with your files and design system.
Figma agent, FigJam AI, Miro AITurns a description or a design into a clickable app.
Figma Make, Claude Design, v0, Lovable, StitchWrite, summarise, analyse and explain anything.
Claude, ChatGPT, Gemini, Microsoft 365 CopilotChange a code base step by step, run it, check it.
Claude Code, Codex, CursorDo one job inside a bigger product.
Dovetail, Maze, image and video generatorsPick by stage and fidelity, not by which tool is newest
How the groups connect
The groups are blending. A design tool now has an agent; a general assistant can design (Claude Design is part of every Claude conversation); a coding agent can read and write Figma files through MCP, a standard way for AI tools to connect to other apps. When tools connect like this, the question "which tool?" becomes "where does the source of truth live?": in the design file, or in the code.
2. Which tool for which task
| Task | Start with | Why | Watch out for |
|---|---|---|---|
| Explore layouts within your design system | AI inside your design tool | Uses your real components | Output still needs a design review |
| Clickable prototype for a usability test | Prompt-to-prototype | Fast, realistic interactions | Usually off-system unless you give it your design system |
| Prototype with real coded components | Coding agent + Figma connection | Closest to what will ship | Needs a code base and someone to review the code |
| Research plan, interview guide, survey | General assistant | Good at structure and wording | Check every claim about methods and sources |
| Synthesis of interviews | Research tool or approved assistant | Saves hours of tagging | Shallow themes, invented quotes; see AI in UX research |
| Workshop clustering | FigJam AI, Miro AI | Sorts and summarises stickies | Clusters are a draft; the team decides the meaning |
| Stakeholder deck or one-pager | Claude Design, Microsoft 365 Copilot | Layout and summary in minutes | Numbers and claims must be checked by you |
| UX copy and translation | General assistant, design-tool text tools | Many options quickly | Tone drift; product terms translated wrongly |
| Specs for engineers and agents | General assistant + a template | Structure, completeness | See Design specs for AI |
3. Vocabulary you'll meet
| Term | What it means |
|---|---|
| Seat | The kind of licence a person has in a tool (in Figma: Full, Dev, Collab, View). It decides what AI features you can use |
| AI credits | A monthly allowance of AI use. Each action costs credits; when they run out, AI stops or you pay more |
| Content training | Whether the vendor may use your files to improve its models. Often a setting, sometimes on by default |
| Data processing agreement (DPA) | The contract that says how a vendor handles your organisation's data |
| Data residency | Where your data is stored and processed (for example, inside the EU) |
| Connector / MCP | A link that lets an AI tool read or write another app (Figma, Jira, Notion, Slack…) |
| Agent | An AI that carries out several steps on its own, not just one answer. See AI agent |
| Research preview / beta | Early versions: features change, limits apply, and data terms may differ from the final product |
4. Questions to ask before you use a tool for real work
You don't need to be a lawyer. You need to ask a few questions, and stop if the answers are unclear.
Data
- Is this tool approved by my organisation? If you're freelance or a student: whose data am I about to paste?
- Will my files be used for training? Can I turn that off, and who can?
- Where is the data stored, and for how long?
Quality
- Does it work with our design system (components, tokens), or does it invent its own look?
- Can I check its output easily, or do I have to trust it? See AI design review process.
Cost and exit
- What does it cost in seats and credits, and what happens when the credits run out?
- Can I export my work if we stop using it?
Security flags from 2026
Two real events show why "install only from official sources" matters:
- Tampered extensions. On 29 October 2026, Cursor reported that two tampered versions of the Claude Code extension had been in its marketplace for about 11 hours; they installed a backdoor and were not made by Anthropic (notice). Extensions, plugins and connectors can read your files: install them only from the official publisher.
- Owners change. Cursor was acquired by SpaceX in August 2026 (announcement). A change of owner can change data terms. Re-read them when it happens.
5. Deciding whether a tool is worth keeping
Nielsen Norman Group's PROVE framework (August 2026) is a simple way to judge a tool for a single task (article):
Does it solve a problem I actually have?
What happens if it is wrong, and would I notice?
Is the result good enough to use, after review?
Am I really faster, counting checking and fixing?
Is it pleasant and sustainable to work with?
A fair trial in four steps
- 1Pick one real task
Something you do often.
e.g. writing test tasks - 2Do it without
Note the time and the result.
- 3Do it with
On a similar case, counting review and fixes.
- 4Score and decide
PROVE: keep, drop or try again later.
6. Try it (no code needed)
- List the tasks you did last week (10–15 is enough).
- Mark each one: routine (similar every time), craft (needs your judgement), or sensitive (personal or confidential data).
- For each routine task, write which kind of tool from section 1 could help.
- Pick one, run the four-step trial from section 5, and write two sentences about what you learned.
Reflection questions: Which of your tasks would you not hand to AI, even if it were fast? Why? Where does the source of truth for your work live: in design files or in code?
7. If you lead a team
- Keep a short list of approved tools and what data each one may receive. Update it every quarter.
- Run pilots for 4–6 weeks on a real task, measured with PROVE, before you buy seats for everyone.
- Check whether a tool could be used to monitor people's work. In many countries that needs legal review and employee representatives. See AI governance and data protection.
- See also Guiding designers through AI adoption.
Checklist
Learn more
- NN/g: PROVE framework · Same Conclusion, Less Variation: synthetic users vs human respondents
- Standards: Design Tokens Format 2025.10, the first stable version (a W3C Community Group report, not a W3C standard) · DESIGN.md, a format for describing a visual identity to coding agents · AGENTS.md
- Specs: GitHub Spec Kit · Kiro specs
Books, courses and videos to go deeper
Related
How a model learns: train one on your laptop · Figma AI · Claude Design · Generating design with Claude Code · AI in UX research · AI governance and data protection · MCP · Synthetic users