AI Assistant Configuration - Complete Guide

Google Chirp 3 HD offers the best quality/price ratio for Italian. Azure is great for professional voices. ElevenLabs is the absolute top for emotional naturalness but more expensive.

Depends on volume. For low usage (< 10,000 minutes/month) our prices are competitive. For high volumes, with your own API keys you pay the provider directly and avoid markup.

Yes, instant change. Select new TTS provider, choose the voice and save. Future conversations will immediately use the new configuration. No downtime.

Yes, Azure ASR is optimized for standard Italian and also recognizes strong regional accents well (Neapolitan, Sicilian, Venetian). Performance above 90% even with imperfect audio.

Total latency (ASR + LLM + TTS) typically 800ms-1.5s. Azure ASR ~ 200-300ms, Google TTS Chirp 3 ~ 150-250ms. The rest is LLM processing. Well below conversational naturalness threshold.

Studies show it depends on the sector: female voice more effective for healthcare, customer care, education. Male voice for finance, automotive, tech. Do A/B testing!

No if used correctly. The volume is pre-optimized not to interfere with the voice. Makes the conversation "warmer" and less artificial. Avoid only for technical support where maximum clarity is needed.

TTS providers (ElevenLabs, Azure Custom Neural Voice) support voice cloning, but requires external setup. You can use the cloned voice by providing the custom voice ID/name in the "TTS Voice" field.

Any common format (MP3, WAV, OGG). The system automatically converts to ALAW for phone compatibility. Max 50MB. After uploading you can adjust volume and listen to a preview.

Absolutely! Each assistant has independent voice configuration. You can have Sales_Assistant with a young female voice and Support_Assistant with a mature male voice.

GPT-4o-mini is 60% cheaper and 2x faster, great for simple FAQ, receptionist, screening. GPT-4o for complex conversations, consultative sales, technical troubleshooting. Start with mini, scale to 4o if needed.

Depends: Claude Sonnet excels in empathetic conversations, understanding long contexts, following complex instructions. GPT-4o is faster and costs less. For delicate customer care → Claude. For volume → GPT-4o.

Yes, but you must host Ollama on your own server with GPU and provide API endpoint. Models like Llama 3.1, Mistral work well. Pros: zero API costs. Cons: you must manage infrastructure.

Between 500-2000 words. Too short (< 300 words) = unpredictable behavior. Too long (> 3000) = AI loses focus. Include: role, tone, examples, edge cases. Quality over quantity.

ALWAYS start from a template! They are already optimized for telephone conversations. Personalize: company name, services, FAQ, tone. Reinventing the wheel leads to poor prompts.

Yes but better to use RAG! Prices in prompt = you must update manually each time. Prices in RAG knowledge base = update the file and it works everywhere. However, for 5-10 products, prompt is fine.

Clear instructions in the prompt: "If you don't know the answer, say 'I don't have this information, I'll transfer you to an operator.' DO NOT invent prices, hours, policies." + RAG for factual info + Intensive testing.

Yes, a lot! A long/confusing first message = disoriented user. Better short, clear, action-oriented: "Hi, I'm the [Company] assistant. I can help you with [X, Y, Z]. What do you need?"

Most are FREE: Google Calendar, SMTP Email, Webhook, Zapier, n8n, HTTP Request. Only specific enterprise plugins (e.g., Salesforce Premium) may have additional costs.

No technical limit, but we recommend max 5-7 tools per assistant. Too many tools = AI gets confused and performance degrades. Better specialized assistants with few targeted tools.

The AI receives an error and decides what to do: "I'm sorry, I can't connect to the calendar now. Can I take your details and have someone call you back?" Graceful fallback, no crash.

Yes! Use the "HTTP Request" or "Webhook" tool to call your own REST APIs. Define URL, parameters, headers. The AI calls your endpoint and processes the JSON response. Full documentation in the marketplace.

Absolutely! If you set the wrong time zone, Google Calendar appointments are off. Always set Europe/Rome for Italy. The get_current_date_time tool uses this time zone to give the correct time to the AI.

Depends on the plan: Start = 50 documents/5MB total, Boost = 500 documents/50MB, Scale = unlimited/500MB. One document = PDF, DOCX, TXT. The system automatically indexes and makes it searchable.

Yes! Every time the AI needs info, it automatically queries the knowledge base in ~200-400ms. Completely transparent to the user. Zero additional setup on your part.

Yes for specific cases: long code input (e.g., case number 18 digits), sensitive confirmations (press 1 to confirm charge), users who prefer keypad, fallback when voice is incomprehensible.

Absolutely! Best practice: RAG for static info (price lists, policies, FAQ), Tools for dynamic data (calendar availability, order status). The AI autonomously decides which to use.

The system intelligently chunks it into fragments of ~500 words and indexes them. When the AI searches, it retrieves only the 3-5 most relevant chunks, not the entire document. Efficient even on huge documents.

Only if you enable "Enable Hangup". Otherwise it waits for the user to hang up. With hangup enabled, the AI evaluates goal achieved + user satisfied = polite closure. Saves costs.

Yes, in the prompt write: "If user asks to speak with a real person, transfer IMMEDIATELY to extension 100". With "Enable Redirect" active, the AI performs immediate transfer.

Yes, if you enable "Conversation Saving" with valid emails. Transcripts are encrypted, stored in EU, accessible only to you. Remember to inform users about recording for transparency.

Hangup URL sends to ONE endpoint. If you want multiple: use n8n/Zapier/Make as middleware. AI4CALL → n8n → n8n distributes to CRM + Google Sheets + Slack + etc.

Fields that the AI must extract from the conversation: customer_email, reported_issue, priority, etc. Define name, type, description. AI populates them and sends to webhook. Automatic CRM integration.

No, the webhook is completely optional. If not configured, the assistant works normally without pre-injected data. It is an advanced feature for those wanting maximum personalization and performance.

Not directly, but you can inject data in callerInfo that influences behavior. For example, a field "customerType": "VIP" can be used by the LLM to adapt tone. For complete prompt change, use the "Conditional Prompts" feature of AI4CALL.

Yes, in your webhook analyze the number field and respond enable: false for any pattern. For example, block all numbers not starting with +39 to accept only Italian calls.

Absolutely yes. You can simply log the call to your system and always respond enable: true without callerInfo. Useful for centralized tracking.

No, currently callerInfo supports only primitive values (string, number, boolean). For lists (e.g., "last 3 orders"), serialize into string: "lastOrders": "ORD-123, ORD-456, ORD-789".

Yes, but beware of timeout. If calling a slow CRM, set a local timeout lower than AI4CALL's (e.g., 3s) and use fallback in case of timeout. Consider cache to reduce dependency on external APIs.

Yes, each AI4CALL assistant can have its own configured webhook endpoint. Useful if you have assistants for different departments (sales, support) that query different databases.

VAD ON = more natural conversation, AI waits for user to finish speaking. VAD OFF = faster responses but risk of interruptions. For customer care → ON. For fast IVR menus → OFF.

No, it's optimal for Italian. 500ms allows natural pauses between sentences. If users speak very fast (internal call center) you can go down to 300-400ms. For elderly/slow users → 700-800ms.

Depends on the use case. 10s is fine for quick FAQ. For consultative sales where user needs to look for documents → increase to 30-45s. The AI warns before closing anyway.

Yes! In "Allowed number types" deselect "International". All calls with non-Italian prefix will be automatically rejected. Useful against international spam.

No, they are optional. They are for advanced statistics (e.g., "How many calls from mobile vs landline?") or to apply differentiated logic (e.g., "From mobile transfer to department A, from landline to department B").

The button is active only if an AI4CALL number (demo or purchased) has already been assigned to the assistant. Go to "AI4CALL NUMBERS" → "Routing" and assign one.

No. The system uses the number saved in your user profile. If missing, on the first click a modal will ask you to enter it; afterwards it will be reused.

The cost depends on the called number's country and is deducted from the AI4CALL credit. Rates are visible on the "Profiles" page.

Yes. Calls will pass through your operator without AI4CALL charges.

With AI4CALL trunk they must be in international format with "+", e.g., +390200000. With customer trunk use the format required by your operator.

Maximum 50 numbers per file. Accepted formats are CSV or Excel (.xls/.xlsx) with one number per line.

No. The system calls one contact at a time; when it ends (answer, busy, no answer) it automatically moves to the next.

With AI4CALL trunk you can only select from assigned AI4CALL numbers. With customer trunk manually enter the CLI you want to display.

Yes. Click "STOP CAMPAIGN" in the management panel; the ongoing call will be completed, but no new ones will start.