Guide8 min readUpdated
AI Workflow for Designers: What to Automate, What to Keep
An AI workflow for designers, stage by stage: what AI can draft in a UI/UX project, which decisions stay with a person, and how the client benefits.
By Tithi Mondal, UI/UX Designer
Suppose a founder in Bardhaman sends 9 voice notes on Monday about a grocery ordering app. The founder wants screens by the end of the month. An AI tool could produce 30 screens by Monday evening. The screens would look finished. Nobody would have asked whether local shoppers pay by UPI or cash on delivery, or whether the shop owner can update stock from a basic Android phone.
An AI workflow for designers is a fixed plan that says which steps of a design project an AI tool may draft and which decisions stay with a person. This guide maps such a plan onto the five stages I follow on every project: Understand, Map the flow, Wireframe, Design the interface, Prototype and hand off. For each stage I describe what AI can draft, what I keep, and what the client gains.
Checked in November 2026. Tools change quickly, so confirm details on the project's own page.
What is the rule behind this AI workflow for designers?
The rule is simple: AI drafts, a person decides. This is often called human in the loop design, which means a person reviews and approves every AI output before the output moves forward.
Before I would give any task to an AI tool, the task must pass three questions:
- Is the data safe to share? Client material under a non-disclosure agreement does not go into an online tool.
- Can I check the output quickly? A summary can be checked against the source. A claim about "what users want" cannot be checked without users.
- Who answers if the output is wrong? If the answer is "the designer", the designer must make the final call.
| Stage | AI can draft | A person keeps |
|---|---|---|
| Understand | Summaries, question lists, competitor notes | Conversations with the client and users, the goal |
| Map the flow | Edge cases, alternative paths | The order of steps, what to leave out |
| Wireframe | Layout variations, placeholder text | Priority on each screen, final wording |
| Design the interface | Mood images, placeholder visuals | The visual system, consistency, accessibility |
| Prototype and hand off | Automated checks, first draft of notes | Testing with people, final approval |
The general limits of AI design tools are covered in what AI design tools can and cannot do. This guide stays with the method.
Stage 1, Understand: what can AI speed up?
In the Understand stage, AI can turn raw material into organised notes, but AI cannot do the understanding. The stage ends when the designer can state the business goal and the main user in one sentence each.
AI can draft: a one-page summary of a long brief, a transcript of voice notes, a list of open questions for the first call, and a table of features seen in competitor apps.
A person keeps: the call with the founder, the conversations with real users, and the decision about what the product is for. A language model has never stood in a kirana shop at 8 pm. A model will describe a "typical user" from patterns in its training text, and that description is a guess.
One example of a tool here: Ollama is an open source program that, according to the project's documentation, runs AI models locally on your own computer, which suits confidential briefs.
Client benefit: the first call starts with sharper questions, so fewer surprises appear in week 3. If you have not spoken to users yet, my guide to user research for startups shows a small way to begin.
Stage 2, Map the flow: where does AI help?
In the flow stage, AI helps most as a checker that asks "what if". A user flow is the path a user takes, screen by screen, to complete one task.
AI can draft: a list of edge cases for a flow I have already drawn. For a seller onboarding flow, the list might include: the OTP does not arrive, the document photo is blurred, the seller closes the app halfway, the network drops during upload. On a seller app I designed for a marketplace, cases like these in onboarding and verification needed as much design attention as the main path.
A person keeps: the order of the steps, and the decision about what to remove. A model tends to add steps, because more steps look complete. A good flow usually comes from removing steps. A designer weighs each field against the business need: does the app truly need the PAN number before the first order?
Client benefit: fewer missing screens are discovered during development, when changes cost more. My guide on what a user flow is explains how to read the diagram you will receive.
Stage 3, Wireframe: what can AI draft?
In the wireframe stage, AI can produce quick layout options and realistic placeholder text. A wireframe is a plain, grey sketch of a screen that shows what goes where, before colours and style.
AI can draft: 3 or 4 arrangements of the same content for comparison, sample product names and prices in rupees, first versions of button labels and error messages, and a rough Bengali or Hindi version of the text for checking length.
A person keeps: priority. On a home screen for a grocery app, only one thing can sit at the top: search, repeat last order, or today's offers. That choice depends on the business and on how customers actually shop. The final wording also stays with a person. A native speaker must read every Bengali and Hindi string, because machine text in Indian languages often sounds stiff or wrong.
Client benefit: the client sees more options in the first review and reacts to real-looking content, not meaningless filler text.
Stage 4, Design the interface: what must stay with a person?
In the interface stage, the visual system must stay with a person: colours, type, spacing, components and their rules. AI can supply raw visual material around that system.
AI can draft: mood images to agree on a direction, placeholder photos and illustrations, and variations of a single element such as a banner background.
One example of a tool here: ComfyUI is an open source program that, according to the project's documentation, lets you build image generation workflows by connecting nodes. Each image model has its own licence, which must be read before any business use.
A person keeps: consistency across the whole product. A product with 50 screens needs the same button, the same spacing and the same error style on every screen. Image and screen generators produce each result separately and do not hold that discipline. A person also keeps accessibility decisions, such as contrast and touch target size, and every final brand asset. A logo or app icon should be drawn by hand as a clean vector file.
Client benefit: the direction is agreed earlier, with pictures and not adjectives. That saves revision rounds. My projects include two to three revision rounds, and I prefer that clients spend them on real decisions.
Stage 5, Prototype and hand off: what can a machine check?
In the last stage, a machine can check measurable rules, and a person must check whether people can use the product. A prototype is a clickable model of the app that behaves like the real thing.
AI and automation can draft: a first version of hand off notes for developers, a list of screens and states, and automated checks once pages are built in code.
Two examples of tools here: Lighthouse is an open source tool from Google that, according to the project's documentation, audits web pages for performance, accessibility and more. axe-core is an accessibility testing engine from Deque that, according to the project's documentation, finds only part of the accessibility issues automatically and marks others for manual review. Both tools test built web pages, not design files.
A person keeps: usability testing with 5 or so real people, the review of the built screens against the design, and the final approval. A simulated user is text generated by a language model. A simulated user can point to weak spots quickly, but it is not evidence that real people will understand the screen or want the product.
Client benefit: developers receive clearer files and ask fewer questions. My Figma developer handoff checklist lists what a complete hand off contains.
How does the client benefit, and what are the limits?
The client benefits through faster first drafts, more options at each review and wider checking, while one named person stays responsible. The client does not benefit if AI is used to skip the thinking.
A founder or small team can apply this method this week in four steps:
- Write your own two-column list: "AI may draft" and "a person decides". Use the table above as a start.
- Ask your designer, in writing, which AI tools will touch your files and what data goes into them.
- For one upcoming task, let an AI tool draft, then have a person mark every error. Count the errors honestly.
- Keep the tasks where checking was quick. Drop the tasks where checking took longer than doing the work.
The honest limits:
- Made-up answers. Language models state invented facts with confidence. Every draft needs a check against a source.
- Privacy. Online tools process your text on outside servers. Read their data policies, and keep confidential material out.
- Bias. Models repeat patterns in their training data, which may not match users in tier-2 and tier-3 cities.
- Cost. Paid tools have fees, and local tools need capable hardware and setup time. Check the current plans.
- Sameness. Products drafted by the same tools from the same prompts tend to look alike. Distinct design needs a point of view.
- Time. AI shortens drafting, not deciding. An app or website design with me still usually takes 2 to 4 weeks, because feedback and testing take the time they take.
The full sequence of my process, without the AI angle, is in my article on the mobile app design process. If you would like your product designed with this method, with a person answerable for every screen, my UI/UX design service explains how a project starts.
Sources
Frequently asked questions
Which parts of UI/UX design can AI do?
AI can draft summaries, question lists, edge cases, placeholder text, layout variations, mood images and automated checks. A person must still talk to users, choose the flow, set the visual system and approve the final screens.
What does human in the loop mean in design?
Human in the loop means a person reviews and approves every AI output before the output moves to the next step. The AI tool drafts, and the designer decides and stays responsible.
Does using AI make a design project cheaper or faster?
AI shortens drafting, so first versions and options arrive sooner. Decisions, feedback and testing with people still take time, so the saving depends on the project.
Should I ask my designer which AI tools they use?
Yes. Ask in writing which tools will touch your files and what data goes into them, especially if your product is unreleased or covered by a non-disclosure agreement.
Working on something like this?
Tell Tithi what you are planning. You get a clear scope, a fixed price and a timeline before any work begins. I reply to every enquiry within 24 hours.