Building forms with AI

Building forms with AI

The AI builder is the heart of AmazingForms. This guide explains how to get the best out of it.

Writing a good prompt

The AI builder understands plain English. You do not need special syntax. But better prompts produce better forms.

Include:

• What the form is for (the business context)
• Who fills it in
• What fields you need (or a general description of what data you want)
• What should happen after submission

Example — vague prompt:

A contact form

Example — good prompt:

A quote request form for my roofing business. Collect the customer's name,
email, phone number, home address, and photos of the problem area.
After submission: score how urgent the job looks, text me if it scores above 70,
and send the customer an email promising a callback within 24 hours.

The second prompt produces a 5-field form with a file upload, a workflow that calls an AI scoring step, a conditional SMS alert, and a personalised email reply.

What happens under the hood

AmazingForms uses a NL to readable DSL to validated schema pipeline. This is not an LLM writing raw PHP — it is safer and far more reliable.

1. Natural language in — your prompt goes to a small reasoning model
2. DSL generated — the model writes a readable intermediate language (structured YAML-like instructions)
3. Validated — the DSL is checked against a schema of whitelisted primitives; invalid instructions are rejected, not executed
4. Schema out — the form schema and workflow definition are written to the database

This means:

• The AI cannot execute arbitrary code
• Errors are caught before they reach your site
• You can inspect and edit the output in the visual editor

The agentic workflow layer

After your form is built, every submission triggers an AI agent. The agent:

1. Reads the submission data
2. Reasons about it (qualification, urgency, sentiment, completeness)
3. Executes the actions you described: replies, alerts, CRM updates, integrations

The agent uses hierarchical tool-calling: it first decides what kind of action to take, then fetches the configuration for that action. This keeps inference costs low — it never loads the entire plugin surface into context for a simple reply.

Example agent trace

For the roofing quote form:

Agent: Read submission — John Smith, 14 Oak Lane, roof photos attached
Agent: Score urgency — photos show active leak, date mentioned "this week" — 91/100
Agent: Route — Hot (score above 70)
Agent: Send SMS to owner — "Hot lead: John Smith, 14 Oak Lane, 91/100 urgency"
Agent: Generate reply — personalised callback promise in owner's tone
Agent: Send reply to John Smith
Agent: Create CRM contact — tagged roofing, hot-lead, pipeline stage: New

All of this runs automatically, within seconds of submission.

BYOK vs managed credits

AmazingForms supports two AI modes:

Mode	How it works	Best for
Managed credits	AmazingForms proxies calls to a small, cheap model	Most freelancers — lowest friction
BYOK	You supply your own API key (OpenAI, Anthropic, etc.)	High volume, frontier models, data privacy

You can set per-form and per-workflow-step model preferences. Per-step routing means a lead qualification step can use a cheap fast model, while a nuanced reply generation step can use a stronger one.

Setting your API key

In AmazingForms → Settings → AI:

1. Select BYOK
2. Enter your API key
3. Select default provider and model
4. Save

Your key is stored encrypted in wp_options and is never transmitted to AmazingForms servers.

Editing AI output

After the AI builds a form, you land in the visual editor. Everything is editable:

• Rename, reorder, delete fields
• Change field types (text to textarea, select to radio)
• Edit validation rules
• Add or remove workflow steps
• Change the AI reply template

The visual editor writes back to the same schema format — the AI and the editor speak the same language.

Tips and limits

• Prompt length: up to ~2,000 characters. Longer is fine but rarely helps.
• Photos / files: always use the phrase “upload” or “attach” — the AI will add a file field.
• Conditional logic: say “only show X if Y is answered” — the AI handles it.
• Multi-step: say “split this into 3 steps” or “make it a wizard” — it will paginate the form.
• Reply tone: paste 2-3 sentences of your own writing and say “reply in this tone” for better personalisation.