The blank form is the worst part of building one. Not the fields, not the logic, not the styling. The moment where there is nothing on the canvas and you have to decide, from scratch, what a customer intake form for a physiotherapy practice ought to contain.
That is the part a language model is unusually good at, because the answer is not a matter of taste. It is a matter of convention, and conventions are exactly what a model trained on the public internet has seen a great many examples of. It has read more physiotherapy intake forms than any physiotherapist has.
What it is not good at is knowing your business. The gap between those two things is where all the useful thinking about AI form generation lives, and most of the writing on the subject skips straight past it in either direction: either the machine does everything, or it is a gimmick. Neither is true.
What generation is actually solving
The value is not typing speed. Anybody can add a field. The value is that you are handed a structure to react to, and reacting is a completely different cognitive task from inventing.
Give somebody a blank page and ask for a workshop booking form and they will stare at it. Give them a draft with twelve questions, four of which are wrong, and they will fix the four in about ninety seconds and immediately notice the two that are missing. The draft is doing the same job a rough sketch does in any other kind of design work.
This is why the quality bar for a generated form is lower than people assume. It does not need to be right. It needs to be close enough that correcting it is faster than starting, and that bar is easy to clear.
Reacting to a wrong draft is a different job from inventing a right one, and a much easier one.
Write the prompt like a brief, not a search
The single biggest difference in output quality comes from how much context you put in, and most people write three words when three sentences would transform the result.
A booking form is a search query. A booking form for a two-person dog grooming business, where people choose a size of dog, a service, and a preferred day, and we need to know if the dog has bitten anyone before, is a brief. The second produces something you can almost use. The first produces a generic template you have seen a hundred times.
The pieces worth including are what the business does, who is filling it in, what happens after submission, and any question that is unusual for your trade. That last one matters most, because the unusual questions are precisely the ones a generic model will not think of, and they are often the ones that make the form actually useful.
Say what you do not want
Generated forms have a consistent bias towards asking for more than necessary, because most of the forms in the world do too. If you say nothing, you will get name, email, phone, company, address and a message box, whether or not any of those are relevant.
Negative instructions work well and almost nobody uses them. Do not ask for a phone number. Keep it under six questions. No free text except the last one. Each of these is a single clause and each of them removes an entire category of correction you would otherwise do by hand.
It is also the fastest way to encode a house style. If every form you build follows the same rules, put those rules in the prompt every time and you will stop making the same three edits on every draft.
What it will reliably get wrong
There is a predictable list, and knowing it turns review from a vague once-over into a checklist you can run in a minute.
- Your specific option lists. It will guess service names, plan tiers and locations, and it will guess plausibly and wrongly.
- Required versus optional. It defaults to requiring more than it should, because most forms do.
- Anything legal. Consent wording, retention statements and terms links need a human, every time.
- Ordering for conversion. It writes forms in a logical order, which is not the same as the order that keeps people going.
- Which questions you already answer elsewhere. It does not know your booking system already collects the address.
Review in the order the person will see it
The instinct is to review a generated form field by field, top to bottom, checking each one is sensible. That catches wrong fields and misses wrong forms.
Read it instead as the person filling it in. Start at the first question and ask whether you would answer it. Then the second. You will notice things a field-by-field pass cannot surface: that three questions in a row are asking roughly the same thing, that the form asks for a budget before explaining what anything costs, that the tone shifts halfway through from friendly to bureaucratic.
This takes two minutes and it is the difference between a form that is technically correct and one that works. It is also a habit worth keeping for forms you wrote yourself.
Generate, then cut
A useful working pattern is to ask for slightly more than you want and then delete. It sounds wasteful and it is faster, because recognising an unnecessary question is easier than remembering a missing one.
Ask for a thorough version of the form. Read it. Delete everything you would not act on. What remains is usually shorter than what you would have written unaided, and it will include one or two questions you would never have thought to add.
The test for deletion is simple and worth applying ruthlessly: if the answer to this question would not change anything you do, it should not be on the form. Most forms carry two or three questions that fail that test and have done for years.
Iterate in conversation, not in the editor
When a draft is close but not right, the temptation is to switch to the builder and fix it by hand. Often the faster move is to say what is wrong and regenerate, because structural problems are structural and fixing them field by field is slow.
Make it shorter and move the contact details to the end. Split this into two steps. Replace the free-text box with a set of choices. Each of those is a sentence, and each would be several minutes of dragging.
Switch to hand editing once the shape is right and the remaining work is specifics: your actual service names, your actual prices, the exact wording of a consent line. That is work only you can do, and it is the point where the machine stops helping.
The honest limits
A generated form cannot know what your business needs to operate. If your fitters need to know whether there is parking, no model will guess that, and the form will look complete while being useless on the day.
It also cannot know your history. The question you added three years ago because of one disastrous job is exactly the kind of hard-won specific that never appears in a generic draft, and it is frequently the most valuable question on the form.
Treat the output as a competent first draft from somebody who knows forms in general and your business not at all. That is precisely what it is, and it is a genuinely useful colleague to have as long as you do not mistake them for an expert in your trade.
Where the time actually goes
If you time yourself building a form from scratch, very little of it is spent adding fields. Most of it is spent deciding, and a fair amount is spent on a specific kind of stalling that happens when the decision feels consequential and the page is empty.
Generation removes the stall rather than the work. You still make every decision, but you make them as edits rather than as inventions, and edits arrive with much less friction. People who have built a lot of forms describe the difference as going from an hour to about ten minutes, and almost all of the saving is in the first five.
That is a smaller claim than the marketing around AI usually makes, and it is a real one. The ten minutes is also more pleasant than the hour, which matters more than it sounds when the task is one you have been putting off.
Use it for the forms you were not going to build
The most interesting effect is not on forms you already have. It is on the ones you decided were not worth the effort.
Most small businesses carry a mental list of things that would be better as a form and currently happen over email: the post-job feedback you keep meaning to collect, the availability check, the supplier onboarding, the event sign-up for the thing in November. Each one has been sitting there because forty minutes of setup was never quite worth it.
At five minutes, that calculation changes, and the change compounds. Every one of those becomes structured data you can read rather than a thread somebody has to go and find.
Keeping the drafts honest
One practical warning. Because generated forms look finished, they get published with less scrutiny than hand-built ones, and the mistakes that survive tend to be the quiet sort: a required field nobody meant to require, a dropdown missing the option half your customers need, a question that reads as rude in your market and neutral elsewhere.
Fill it in yourself once, properly, as though you were a customer, before it goes live. Not a glance at the builder preview. An actual submission, on a phone, with real answers. It takes three minutes and it catches nearly everything.
Then check the submission looks right at the other end, because a form that collects perfectly and lands in a shape nobody can read has solved the wrong half of the problem.
Templates and generation solve different halves
It is tempting to treat a template gallery and a generate button as competing answers to the same problem. They are not, and knowing which one you want saves a lot of fiddling.
A template is somebody else's finished decision. It is the right choice when your situation is genuinely standard, because a good template has been used a thousand times and the rough edges have been sanded off by people complaining. Generation is the right choice when your situation is specific, because that specificity is the whole thing a template cannot contain.
The strongest workflow uses both. Start from a template if one nearly fits, then describe the differences and let the model apply them, which is faster than editing by hand and far faster than starting from nothing. The failure mode is picking a template that nearly fits and then living with the parts that do not, which is how businesses end up with a question about company size on a form only consumers ever see.
What to try first
Pick a form you have been meaning to build rather than one you already have. There is less to lose and you will see the effect more clearly.
Write three sentences of context, one sentence of constraints, and generate. Read it as a customer. Delete anything you would not act on. Add the one question only you would know to ask. Publish it and fill it in yourself on your phone.
If the whole thing takes more than a quarter of an hour, the prompt was too short. That is nearly always the fix.