Internet service providers receive customer complaints through instant-messaging platforms, yet manual handling can delay responses and create message backlogs. This study develops and evaluates a Telegram chatbot for classifying Indonesian ISP complaints submitted as text or voice notes. Its primary contribution is a lightweight multimodal architecture integrating Telegram, Whisper-based transcription, TF-IDF, and Logistic Regression for six operational intents; analysis of speech-recognition error propagation is treated as a secondary exploratory contribution. The corpus contained 1,500 messages: 500 historical complaints collected from January to April 2026 and 1,000 simulated messages created to balance the classes. The combined corpus was partitioned using an 80:20 split, yielding 1,200 training messages and 300 test messages. Because simulated instances were not explicitly restricted to training, the results represent internal test performance rather than conclusive evidence of real-world generalization. Quantitative evaluation on the 300-message test set produced 91.00% accuracy and a macro-average F1-score of approximately 0.92. Router Error showed the lowest precision because of overlap with semantically related complaint categories. Functional verification comprised eight predefined black-box scenarios, all of which passed, whereas the voice pathway was assessed only through an illustrative case in which transcription errors removed discriminative terms and reduced classification confidence. The findings support the technical feasibility of lightweight multimodal complaint triage while emphasizing the need for real-only external testing, systematic audio evaluation using Word Error Rate and downstream intent metrics, confidence-based escalation, and operational validation before large-scale deployment.