Ioana energy billing support

​

Overview

​

Skills

• identification - Identifies customer accounts using Electel/Gaztel codes or Connection ID, calls check_context_metadata immediately, and sets the active account for the rest of the call
• billing_and_payments - Handles invoice and balance inquiries (amount due, overdue invoices, paid/unpaid status) and allows customers to report proof of payment (demoplata) for disputed invoices
• payment_methods - Explains how customers can pay (ATM, cash, bank transfer, internet banking, direct debit, etc.) strictly based on the approved payment-method knowledge base tool
• invoice_copy - Sends a copy of the customer’s invoice via SMS after confirming the active account and the SMS send action

​

Key Tools

​

Prompting Techniques

You are **IOANA**, a helpful and professional Voice AI Agent for **P-P-C Romania**, supporting **invoice and billing questions only**.

Your goal is to assist existing customers with their invoice according to the scope's description.

Ask **one question at a time**.

​

Lessons Learned

• Multi-skill architecture allows focused handling of different customer needs (identification, billing, payment methods, invoice copy)
• Parallel API calls in identification tool (identify_and_get_status) for faster authentication
• Comprehensive Romanian transcription context with domain-specific terms significantly improves accuracy
• Structured tool-based approach for each skill ensures consistent data retrieval and actions
• Dynamic, case-specific instructions instead of exhaustive prompt conditions: Rather than including all possible scenarios in the main prompt (e.g., “if invoice is paid, say X; if unpaid, say Y; if credit exists, say Z…”), tools generate targeted instructions dynamically based on the actual data state. This approach:
  • Reduces prompt complexity and token usage
  • Ensures the agent receives only relevant, actionable guidance for the current situation
  • Makes the system more maintainable as new cases can be handled in code logic rather than prompt updates
  Example from invoice_and_balance() tool:
  Copy

  // Generate minimal, clear English instructions based on actual data state
  const instructions: string[] = [];
  
  // Payment status instructions - dynamically generated based on actual payment state
  if (isInvoicePaid) {
    instructions.push(`CRITICAL: Invoice payment status: PAID (este achitata). 
      When stating payment status, use 'este achitata' (is paid), 
      NOT 'deja confirmata' (already confirmed).`);
  } else if (paymentStatus === "NOT_PAID") {
    instructions.push(`Payment status: Most recent invoice is not paid.`);
  }
  
  // Credit vs Invoice logic - only included when relevant
  if (shouldOnlyMentionCredit) {
    instructions.push(`CRITICAL: Current invoice amount is null/unknown, 
      but credit has value (${balance.credit_ron} RON). 
      When user asks about amount due, mention ONLY the credit amount.`);
  }
  
  // Invoice presentation - only included if invoice exists
  if (hasCurrent && !shouldOnlyMentionCredit) {
    instructions.push(`Most recent invoice: ${currentInvoice.amount_ron} RON, 
      due ${currentInvoice.due_date}. 
      Present this invoice separately when user asks about invoices.`);
  }
  
  result.agent_context = instructions.join(" ");
  
  This ensures the agent receives precise, contextual guidance rather than parsing through all possible conditions in the prompt.
• Separation of data retrieval from presentation: The identification tool (identify_and_get_status) fetches and normalizes all customer data from APIs, stores it in KV, but intentionally does not output any profile information to avoid the agent accidentally mentioning customer details prematurely. Only when the billing tool (invoice_and_balance) is called does it retrieve the relevant data from KV and generate contextual instructions for the agent to use. This separation ensures:
  • Data is fetched once during identification and reused efficiently
  • The agent only receives information when it’s contextually relevant to the conversation
  • Reduced risk of the agent mentioning customer details before they’re needed or appropriate Example pattern:
  Copy

  // In identify_and_get_status() - fetch and store, but don't output profile info
  await context.kv.set("account_info", {
    eneltel: normalizedCode,
    name: customerName,
    address: address,
    // ... all data stored
  });
  
  return {
    result: {
      status: "success",
      identified: true,
      // NO customer name or profile details in agent_instructions
      agent_instructions: "Check conversation history for user intent..."
    }
  };
  
  // Later, in invoice_and_balance() - retrieve and generate contextual instructions
  const accountInfo = await context.kv.get("account_info");
  const invoiceInfo = await context.kv.get("invoice_info");
  
  // Only generate instructions about data that's relevant NOW
  if (hasCurrent && !shouldOnlyMentionCredit) {
    instructions.push(`Most recent invoice: ${currentInvoice.amount_ron} RON...`);
  }
  

• Handling Romanian language nuances and pronunciation of bank names/acronyms (solved with diacritics rules for terms like “PPC” → “pipisii”, “BCR” → “becere”)
• Managing multiple invoices when customers report payment for several invoices at once (solved with parallel processing and matching logic)