SaaS product, AI survey builder
Formlark
Describe a survey in a sentence and get one back in about thirty seconds. Generating the draft is the easy half. The design problem is everything after it: making review feel like authorship, and making the finished survey worth answering.
- Client
- Formlark
- Sector
- SaaS, AI survey platform
- Scope
- Product design, design system
- Role
- Nibin Nelson
01 The problem
Thirty seconds, then the hard part
An AI that writes your survey in thirty seconds creates a problem it cannot solve: nobody ships a draft they have not read.
So every product like this has the same trap. Review is mandatory, and every step of review spends the time the generation just saved. Push people through a wizard to check the output and the thirty seconds becomes ten minutes. Skip the review and they send out questions they never actually agreed with.
Branching logic is where survey builders traditionally fall apart. Conditional routing is genuinely complex, and the usual answer is to expose it as a table of rules, which is readable to the person who wrote it and to nobody else, including that same person three weeks later.
Underneath sits the thing most survey tools quietly ignore. There are two users with opposite interests: the builder, who wants power and options, and the respondent, who wants to be finished. Completion rate is decided entirely on the respondent's side, which is the side the builder never sees.
And the reach requirements make every layout decision harder. Over fifty languages including right-to-left, plus a WCAG 2.1 AA commitment, means each surface has to survive direction flip, heavy text expansion, keyboard-only use and a screen reader, not just look correct in English.
02 The approach
Make review feel like authorship
The generated survey arrives as a finished, editable document rather than a wizard to click through. Nothing stands between the prompt and the thing itself. Reading it is the review, and because every question is editable in place, agreeing with the draft costs nothing while changing it costs one click.
Refinement happens on the question, never in a modal. Rewriting, tone changes and logic suggestions appear against the line they affect, so the user stays in the document and keeps the sense that the survey is theirs. The AI is positioned as a fast first draft, which is what it is good at, rather than an oracle to argue with.
The empty state accepts a sentence, a link or a PDF without asking anyone to choose a mode first. One field, three kinds of input, no decision before the user has seen the product do anything.
Branching logic became a readable flow instead of a rules table. Conditions are shown where they happen, attached to the question that triggers them, so the path a respondent takes can be followed by eye. Complexity that cannot be removed can at least be made visible.
The respondent experience was designed first, on the argument that it decides whether any of the builder's work pays off. One question at a time, progress always visible, and a form that stays quick on a phone with a poor connection.
Accessibility went into the component library rather than a checklist at the end, so every generated survey inherits keyboard support, focus order, contrast and screen reader labelling by default. Layout uses logical properties throughout, which is what makes right-to-left a setting rather than a rebuild.
Results lead with the chart. The AI written narrative sits underneath as a way in for someone skimming, never as a replacement for the numbers, because a summary you cannot check is not evidence.
- Prompt to draft generation flow
- In-place question editing and rewriting
- Branching logic as a visible flow
- Respondent survey experience
- Results and analytics views
- Component library to WCAG 2.1 AA
- Multilingual and right-to-left layout system
- Marketing site and pricing page
03 The outcome
A draft you are happy to send
The thirty second promise survives contact with the review step.
Because the draft arrives as a document rather than a wizard, reading it is the review, and most users publish after changing two or three lines. The work the AI does is visible, and so is the part the user did, which is what makes the result feel like theirs.
Accessibility and right-to-left support sit in the components, so every survey built on the platform inherits them without the person building it having to know they exist. For a tool whose whole point is that anyone can make a good survey quickly, that is the part that scales.
Next case study