Overview
| Field | Value |
|---|---|
| Market | Greece - Package Tracking |
| Vertical | Logistics & Postal Services |
| Model | GPT Realtime |
| Language | Greek |
| Channels | Voice, Chat |
Challenge
ELTA, Greece’s national postal operator, faced a critical capacity crisis where only 60-70% of inbound calls were being answered during peak periods. Simple “Where is my package?” inquiries overwhelmed human agents, leading to high abandonment rates and customer frustration. The existing human-only support model, supplemented by basic IVR routing, couldn’t resolve issues or scale with demand. Adding complexity, tracking numbers spoken in Greek, often mixing Greek and Latin letters with unclear pronunciation, proved difficult to capture reliably through traditional automation. ELTA’s leadership sought a production-ready AI solution that could absorb peak demand without expanding headcount, while maintaining accuracy and customer trust.Solution
Wonderful deployed an AI voice agent specifically designed for courier tracking, featuring intelligent phonetic fallback logic, word-based letter recognition, and digit grouping to reliably capture spoken tracking codes. Within months of November 2025 go-live, the agent now handles 61K interactions monthly (33% of ELTA’s total call traffic) with an 86% success rate. The system provides 24/7 availability and scales peak capacity from 500 to 2,000 calls without additional agents.Skills
- Package Status - Core functionality for tracking packages and managing call flow
get_package_status- Retrieves current package status from tracking numberwelcome_message- Contextual greeting based on call center hoursend_call- Terminates call with goodbye messageend_with_feedback- Ends call while logging customer satisfactionforward_call- Escalates to human agent when neededcall_timeout- Handles call timeout
Key Tools
| Tool | Purpose |
|---|---|
| get_package_status(tracking_number) | Validates tracking number format, handles Greek-to-Latin conversion, runs validations and gived feedback to agent if any errors are detected, returns package status in TTS-friendly Greek; will iterate on a tracking number and will check multiple similar sounding tracking numbers |
| welcome_message() | Plays contextual greeting based on time/day, handles blacklisted numbers |
| forward_call() | Escalates to human agent (+302106073002) during business hours (Mon-Fri, 8am-6pm) – forwarding currently disabled upon client’s request |
| end_call() | Terminates call with goodbye message – 10% chance to prompt the customer for feedback and prompt the agent to proceed with end_with_feedback |
| end_with_feedback(customer_satisfied) | Ends call and logs satisfaction feedback |
Prompting Techniques
Critical Tracking Number Parsing for Greek Voice Input: The agent handles extensive Greek-to-Latin letter mapping for voice transcription, including Greek letter names, words, and phonetic equivalents.Lessons Learned
What worked- Extensive Greek phonetic mapping - Comprehensive rules for converting Greek letter names (alfa, vita, ita), Greek words, Greek names, and English phonetics in Greek (Mpi -> B, Si -> C) eliminated most transcription errors and resulted in a good feedback loop between the agent and the customer
- Edge case handling for common mis-transcriptions - Specific rules for “ITA” -> H, “ro” -> P, “eight” -> H at start, and common bad transcriptions like “I tatziou” -> “HQ” improved accuracy
- Digit grouping awareness - Supporting (groups of 3) and (groups of 2) matched how Greeks naturally dictate long numbers; prompting the customers to input numeric sequences in such sets also helped with STT
- Segment-by-segment validation - Having the tool run validations on the tracking number based on known heuristics and instructing the agent accordingly reduced customer frustration – this was one of the more important additions to the agent and greatly contributed to a high resolution rate
- Character spelling alphabet - Consistent word examples (G opos George, H opos Hotel) for both input and output created predictable interactions
- TTS-friendly digit output - Spelling digits as Greek words (tria, ena, tessera) instead of numbers (314) ensured correct pronunciation
- Iterating on commonly misheard characters - We found that certain characters are often misheard as other characters (for example D and G), when the tool had a tracking number that included such characters such as AD123, it would try and query both AD123 AND AG123; this greatly mitigated the impact of transcription errors
- Country code defaulting - Most tracking numbers end in “GR” but transcription often mangled it; proactive suggestion of “GR” when country code validation fails improved success rate
- Long rigid inputs are problematic - Even with a relatively small character error rate, a single character being off will result in failure
- Greek letters that look like latin letters - For example Greek rho (P) sounds like English R but maps to Latin P
- “ITA” vs “ita” confusion - The Greek letter name “ita” (H) was often transcribed as “ITA” and misinterpreted; required specific edge case rule
- Gendered language consistency - Greek requires grammatical gender agreement; agent must maintain consistent gender in pronouns and verbs even if customer uses opposite gender
- Filler sound filtering - Greek thinking sounds (“Eee”, “aaamm”, “Err”) were being interpreted as input
Builder Kit
Everything below is a ready-to-go package for quickly building a package tracking demo agent using the Wonderful Agent Builder. Hand these materials to the builder and you’ll have a working agent in minutes.Agent Builder Prompt
Copy-paste this prompt into the Agent Builder to generate a package tracking agent:API Reference
Endpoint
Request Payload
| Parameter | Type | Description |
|---|---|---|
number | string | The 13-character tracking number |
s | integer | Always 0 |
Response Format
result array is ordered most recent first.