Year: 2026
Type: dissertation
Abstract: This thesis investigates the operational viability of LLM-based agents for automated taxi request handling across voice and text channels in a real production environment.Four agents were implemented and deployed in live dispatch traffic at Geolink lda.: the WhatsApp Helper Agent (WAH), a structured multi-model pipeline using locally deployed Qwen3:30B and Gemma-4-31B-Heretic; the WhatsApp Conversational Agent (WAC), a fully conversational agent based on GPT 4.1-mini; the Napoleão voice expansion, a keyword-based voice detection layer built on a local Whisper model; and the Realtime Napoleão Agent (RTN), a low-latency voice agent built on OpenAI's Realtime GPT.The results show that architectural scope, more than raw model capability, determined production viability.The WAH agent reached a service completion rate of 65% and an average user satisfaction score of 4.79 out of 5, with consistent growth over a three-month deployment.The Napoleão voice expansion accounted for 14% of voice prompt interactions and demonstrably served an older demographic at low operational cost.The RTN agent achieved a 53% service rate in live testing.The WAC agent, equipped with thirteen tools across five functional groups, could not sustain reliable operation due to hallucinated tool calls, incorrect tool parameterisation and message loops, and its test was ended early.The central finding is that, with the models available at viable cost and latency, narrowing agent scope combined with tool-grounded commitments and explicit human-in-the-loop escalation outperforms full conversational autonomy in a production dispatch setting.The WAC and RTN architectures remain valid scaffolding for future models with stronger reasoning capability and lower cost.
Author: Miguel Quatorze Matias Lopes
Language: English
Topic: AI in Service Interactions
Subfield: Marketing
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7213573366