AI‑Driven CRM: Intelligent Opportunity Distribution and Sales Voice Assistant at 58.com

Since Alan Turing introduced the concept of machines capable of thought, AI has rapidly evolved, and at the 2021 Global Machine Learning Conference organized by Boolan, 58.com’s TEG AI Lab presented the talk "AI+CRM Improves Enterprise Efficiency and Performance".

58.com’s lifestyle service platform connects millions of C‑end users with B‑end merchants across real‑estate, automotive, recruitment and local services. Sales teams sell membership packages to merchants, and CRM (Customer Relationship Management) systems are essential for managing leads ("opportunities"). In June 2020, the AI Lab took over CRM intelligence, applying personalized recommendation, search, voice, NLP and dialogue technologies to the local services line.

To improve efficiency, the traditional sales workflow was restructured into a "Michigan" model, splitting teams into an Opportunity Group (screening leads) and a Sales Group (deep follow‑up). The Opportunity Group focuses on generating transferable leads, while the Sales Group concentrates on converting them, greatly increasing overall team efficiency.

The intelligent opportunity distribution system treats the CRM as an information‑distribution engine. It consists of a data layer (collecting opportunity, call and sales behavior data), a core algorithm layer (recall strategies and ranking models), and application layers for personalized recommendation and search. Models such as XGBoost, MMOE and PLE are used to predict metrics like call‑to‑transfer rate (CTCVR) and conversion rate (CVR).

Two major model versions were deployed:

V1 : Used a call‑to‑order scoring model for recall and XGBoost for ranking, achieving a 32% increase in CTCVR and 13% in CVR.

V2 : Added FM and DNN twin‑tower recall, and MMOE/PLE ranking, resulting in a 62% lift in CTCVR and 53% in CVR over V1.

In the "one‑click claim" scenario, a new workflow randomly splits traffic between the original and the personalized recommendation logic for AB testing, ensuring unbiased evaluation.

Beyond recommendation, a sales voice assistant was built. The assistant uses speech recognition, NLP intent detection (both single‑turn and whole‑session), and a knowledge base generated via semi‑automatic Q&A extraction to conduct outbound calls, identify transferable leads, and emulate human sales dialogues.

Intent detection models include TextCNN, LSTM, DSSM, BERT and lightweight pre‑trained models (SPTM). The system also incorporates an intelligent speech quality inspection module that transcribes call recordings and applies NLP to automatically assess compliance, reducing manual review effort.

Overall, the AI‑enhanced CRM has become a core revenue engine for 58.com, with ongoing efforts to expand AI applications such as sales script learning, dialogue coaching robots, and further optimization of the intelligent distribution and voice assistant pipelines.