AI Agents — Automate Conversations with AI

Introduction

Add AI-powered agents to your Answe widgets. AI agents can respond to visitor questions automatically, recommend products, and handle conversations when you're not available. You bring your own API key from one of three AI providers — OpenAI (ChatGPT), Anthropic (Claude), or Google (Gemini) — and Answe handles the rest.

How It Works

Adding an AI Agent

1. Create a question with answer options
2. Click an option and add the LLM Agent action
3. Configure the agent in the Settings tab:
   • System prompt — defines the agent's role and tone
   • Prompt — the actual question sent to the AI, with placeholders that pull in the visitor's form fields
   • AI Model — pick ChatGPT, Claude, or Gemini
   • Datasource (PDF) — optionally upload a PDF the AI can reference
   • PDF processing — choose how the PDF is sent to the AI (see below)
4. Visitors who select that option get AI-powered responses

Setting up your API key

Before the AI agent can run, you need to add your API key for the provider you chose:

1. Go to Account → AI Provider API Keys
2. Paste your key in the matching field:
   • OpenAI API key for ChatGPT (starts with sk-…)
   • Anthropic (Claude) API key for Claude (starts with sk-ant-…)
   • Google Gemini API key for Gemini (starts with AIza…)
3. Click Save API keys

You only need a key for the providers you actually use. Your keys are stored on your account and never shared. The page only ever shows a masked version (••••xxxx) — the raw key is never sent back to your browser after you save it.

If you select an AI Model in a widget but haven't saved that provider's key yet, the widget reply will say "…API key is not set. Open Account → AI Provider API Keys and paste your … key, then submit again." — no silent failures.

What AI Agents Can Do

• Answer questions — Respond to visitor queries about your products, services, or policies
• Make recommendations — Suggest products based on visitor preferences
• Reference documents — Upload a PDF (menu, brochure, price list, FAQ) and the AI uses it as context
• Qualify leads — Ask follow-up questions to understand visitor needs
• Provide support — Handle common support requests automatically
• Follow rules — Combine with conditional logic to route conversations

Combining with Other Features

AI agents work alongside other Answe features:

• After an AI response, show a form to collect contact info
• Let AI recommend products, then link to the shopping cart
• Use AI to determine what the visitor needs, then route to the right next question

PDF Datasource

Upload a PDF in the AI agent's Settings tab and the AI will use it as context for every conversation. The upload happens once; the file is stored privately on your account.

How the PDF reaches the AI

The PDF processing setting controls how the PDF is sent to the model:

• Auto (recommended) — Sends the PDF as a real document when the model supports it (Claude, ChatGPT, Gemini all do today). The model can see layout, tables, and embedded images. Falls back to extracted text automatically if the model rejects the document for any reason.
• Native — Force document mode. Best fidelity for menus with photos, scanned documents, charts, or anything where layout matters.
• Text only — Extract the text first and inline it into the prompt. Fastest and cheapest, but loses tables / images / visual structure.

Auto is right for almost every case. Pick Text only when you want to minimise token cost and you know your PDF is plain text. Pick Native when layout fidelity matters and you've seen Auto fall back to text-only on your file.

Reference the PDF in your prompt

Use {action.datasource} in your prompt template — Answe substitutes it with the PDF's content automatically:

You are a hotel concierge. Use the document to answer.
Visitor question: {form.question}
Reference document: {action.datasource}

If your prompt references a form field that doesn't exist (e.g. {form.preference} when your form actually has a field named question), the widget will reply with a clear "Configuration issue" message naming the missing field — instead of letting the AI hallucinate around the empty placeholder.

Viewing Your Usage

Go to Account → Usage (last 30 days) to see what your AI agents have done across all your widgets:

• Top-line cards — total calls, AI tokens used, estimated cost
• By action type — counts and OK/failed split for every action type (LLM agent, API call, show message, etc.)
• AI usage by provider × model — tokens and estimated cost broken down per AI model
• Recent failures — expand any row to see what went wrong, including the prompt that was sent and the model's reply
• Recent llm_agent calls — last 10 AI calls (any status). Expand a row to inspect the system prompt, the prompt sent to the model, and the model's reply. Useful when the AI gives an answer that's "almost right" — you can see exactly what context it received.

The estimated cost is based on each provider's published per-token pricing and is best-effort — slightly off when providers change their prices. Use it as a "your usage so far" guide, not as a billing total.

The Usage card only shows your own activity. Visitors never see this data.

Key Benefits

• 24/7 availability — AI responds instantly, any time of day
• Bring your own key — Use your own OpenAI, Anthropic, or Google account; pay only for what you use, directly to the provider
• Three AI providers, one widget — Switch between ChatGPT, Claude, and Gemini per question; no code changes
• PDF-aware — Upload a brochure or menu and the AI uses it as context
• Transparent usage — See tokens, cost, and the actual prompt + reply for every AI call
• Safe-fail messages — Missing keys and prompt typos surface as readable chat replies, not silent errors

Getting Started

1. Get an API key from your chosen AI provider (OpenAI, Anthropic, or Google AI Studio)
2. Save it on Account → AI Provider API Keys
3. Create a question with an option
4. Add the LLM Agent action to the option
5. Configure the system prompt, prompt, AI Model, and (optionally) upload a PDF
6. Test the conversation flow in your widget
7. Embed and go live
8. Watch usage build up under Account → Usage

Related

• Chat Inbox — Review AI conversations
• Forms & Surveys — Collect data after AI interaction
• Online Store — AI-powered product recommendations
• Template Gallery — AI-ready templates
• Analytics & Insights — Visitor-side analytics (impressions, conversions); the Usage card on Account is separate, owner-side AI cost tracking