# Lead forms and guardrails > The three ways a form reaches a visitor, and the rules the assistant follows while it asks. Source: https://welcomeai.dev/docs/lead-forms Category: Widgets Updated: 2026-09-07 --- ## The three layouts A lead form is a set of fields plus a decision about how the visitor is asked for them. You build one at **Leads -> Forms**, and the layout is the **Layout** setting under Submission. - **Inline.** Every field at once, inside the chat panel. The right choice for two or three fields. - **Step by step.** One question per screen with a progress indicator. Better for a longer form, because a visitor who sees eight boxes at once usually closes the panel. - **In the chat.** No form is shown at all. The assistant asks for the details in conversation, one or two at a time, while it answers the visitor's own questions. The stored lead is identical whichever you pick. A name asked in conversation lands in the same two columns a submitted form fills, a dropdown answer is still checked against its options, and a hidden field still takes the value you configured. ## What is different about In the chat There is no submit button, so there is no moment where everything arrives at once. Each answer is saved **as it is given**. Somebody who leaves after two questions leaves a real lead with two real answers on it, marked **Partial** in the leads list with a **Chat** trigger. A partial lead becomes a normal one by itself the moment the last **required** field arrives. An optional question left unanswered does not hold it back. Two consequences worth knowing: - The assistant never asks twice for something it already has, including a week later in a new conversation. It is told what is already saved on every turn, so this does not depend on the transcript still being in its context. - Your **Consent text** is spoken rather than shown. A conversation has nowhere to put a checkbox, so the assistant says the sentence before it asks for contact details. ## Guardrails Guardrails are the rules the assistant follows while it answers. They live in two places and the second one only exists to disagree with the first: - **Settings -> Guardrails** is the workspace rule. Every knowledge base and every lead form runs under it unless it says otherwise. - **The Guardrails card on a form** shows exactly what that form will do, and can replace the workspace rule for that form alone. A new form always starts under the workspace rule and picks up any later change made there. Turning on **Give this form its own rules** takes a full copy and stops following: from then on, changing the workspace rule does not change this form. The replacement is the whole policy, not a patch. If the workspace blocks a topic and this form does not, the topic is no longer blocked here. That is deliberate -- a half-inherited policy is one nobody can read off the screen. ### Two rules you cannot switch off Every assistant in every workspace follows these, and no setting removes them: - **It answers only from your knowledge base.** If the material does not contain the answer, it says so and offers what the visitor can do next. It does not fall back on general knowledge, and it does not invent a price, a date or a policy. - **It never shares a credential it finds.** If an API key, a token, a password, a cookie or a private key is sitting in one of your documents, it is removed from the material before the assistant reads it -- so there is no version of the question that gets it back out. The assistant is also told not to explain how to exploit a weakness in your systems. The second one is worth reading twice: it is enforced, not requested. That is the difference between an instruction a clever visitor can argue with and a value that is not there to be repeated. ### What you can change | Setting | What it does | | --- | --- | | Decline questions that are not about your business | The assistant politely refuses anything unrelated and offers to help with something related. | | What it says when it declines | Your own sentence. Leave it empty and the assistant phrases it in the visitor's language. | | Topics it must never discuss | A list. The assistant refuses these even when your documents cover them. | | Every claim needs a source | If nothing retrieved supports an answer, it says the information is not available rather than answering. | | Longest answer | A word cap it keeps to unless the visitor asks for detail. | ### Check the rules before you publish Open **See the instructions the assistant is given** on the Guardrails card. That is not a summary -- it is the literal text the model receives, in order. Read it, then press **I have read these rules**. The warning comes back whenever the rules change, including a change made at the workspace level that this form inherits. A tick that refers to text nobody has read is worse than no tick at all. ## Attaching a form to a widget A form does nothing until a widget uses it. Open the widget, find **Lead form**, and pick it. One form can serve several widgets; deleting a form that a widget still uses is refused, and the message names the widgets. ## Troubleshooting **The assistant is not asking for anything.** Check the widget is pointed at this form, and that the form's Layout is In the chat. A form in Inline or Step by step layout is only shown when the assistant decides to show it, or when a visitor uses a quick action. **It asked for something the visitor already gave.** Look at the lead row: if the answer is not stored, it was rejected by validation -- a dropdown answer outside its options, or an address that is not one. The assistant is told to ask again for exactly those. **It answered something off-topic.** The workspace rule may be off, or this form may have replaced it. The Guardrails card on the form says which, in one badge at the top.