Describe the survey. Then edit the draft.

A parking consultation takes an afternoon to build and a minute to describe. So describe it:

What comes back is a draft along these lines:

• A multiple choice asking whether they are responding as a resident, a business, or both.
• Checkboxes for how they usually travel into the town centre.
• A section break for the proposal itself.
• A linear scale from 1 to 5 for support, with labels at both ends.
• A paragraph question asking what the zone would change for them.
• A section break for businesses, with a number question about deliveries per week.
• A closing section break titled "About you", with an address question for street and postcode, the postcode checked.

Demographic questions land at the end and arrive marked optional, and the drafter is told to ask one thing per question in neutral wording. Those are house rules baked into the drafter, not something you have to ask for — and like everything else in the draft, you can overrule them.

That is a survey you can argue with, which is the point. You will rewrite the support question because your legal team has a preferred form of words. You will add the rule that hides the business section from residents, because a draft is questions, options, scales and section breaks — not conditional logic, scoring or translations. Those you add once the shape is right.

Every kind the drafter can produce is one the builder already supports, so nothing arrives that you cannot then edit by hand. The full set is on the question types page, and the drag-and-drop editing that follows is described on the form builder page.

Your prompt, and nothing else

The only thing that leaves the platform is the sentence you typed. Not your other surveys, not your response data, not your uploaded files. The drafter runs before a survey exists, so there is nothing about respondents for it to send.

Respondents are not in the loop at all

Most of the awkward questions about AI in a survey tool are questions about respondents' data: which sub-processor sees it, what the lawful basis is, whether consent was informed, whether it trains someone else's model. Those questions do not arise here, because no respondent data is involved at any point. The UK GDPR page sets out how respondent data is handled, and the sub-processors page lists every third party that touches it.

Do not paste personal data into the prompt

The prompt is text you write, and it is sent as written. You do not need anyone's name, address or case reference to describe a survey, so do not put one in. Worth saying out loud rather than leaving it implied.

The word "AI" is currently attached to a great many things. Here is the boundary, stated plainly.

• It does not analyse responses. AI analysis of survey responses is not built. The reporting you get is counts, charts, completion and drop-off, worked out from your own data — see analytics and reporting.
• It does not read your data. It has no access to responses, partial responses, uploaded files or exports. It runs before the survey is sent.
• It does not make decisions about individuals. It writes question text before the survey is sent, so there is no profiling of a respondent and no automated decision about an identified person. It is also optional: every survey can still be built by hand. How the platform as a whole handles this is set out on the UK GDPR page.