AI + CRM: Improving Enterprise Performance and Efficiency

In Q4 2020, 58.com launched the Yellow Pages CRM Opportunity Intelligent Allocation project, deploying a machine‑learning model that increased transferred opportunities for the direct‑sales team by 31.8% and 60‑second effective call opportunities for the tele‑sales team by 62%.

The project introduced a new "Michigan" workflow that splits sales staff into an opportunity‑generation group and a sales group, allowing each group to focus on a specific stage of the sales funnel and improving overall efficiency.

CRM opportunities are categorized as new or historical, with various pools (public sea, temporary, private, dormant) that control visibility and lifecycle. Functional modules such as new‑opportunity allocation, merchant‑sale matching, one‑click claim, and transferred‑opportunity allocation support the workflow.

Initially, the system used a scoring model that ranked opportunities by predicted conversion rate, but this approach suffered from indirect optimization goals and lack of salesperson features.

The new solution defines clear optimization targets (CTCVR, CVR) for each stage, focusing on increasing transferred opportunities and 60‑second effective calls. It treats the problem as an information‑distribution task similar to recommendation systems.

Machine‑learning models (XGBoost, later MMoE) were added to the one‑click claim search pipeline, first re‑ranking a larger candidate set retrieved from Elasticsearch, then returning the top‑K results. Features include opportunity attributes, salesperson attributes, behavior data, and voice‑text features.

AB‑tests showed that in the Michigan group, CTCVR improved by 31.8% and CVR by 13%; in the tele‑sales scenario, CTCVR rose by 62% and CVR by 41.2%. The models have now been rolled out to 100% of traffic.

Further enhancements added multi‑path recall (FM, DNN twin‑tower models) and additional scoring dimensions (transfer rate, 30‑second effective call rate). The MMoE multi‑task model further improved CTCVR by 11.9% and CVR by 10.8% over XGBoost.

Experimental design required random traffic splitting due to limited salesperson numbers and varying experience levels.

The article concludes that defining proper ML optimization goals is crucial, and future work will extend models to other CRM modules, incorporate recommendation features, and explore voice‑based quality inspection, speech‑to‑text, and conversational assistants.

Personal reflections emphasize open collaboration across AI, CRM, and business teams, and the importance of long‑term technical investment.

Author: Zhan Kunlin, AI Lab lead at 58.com, senior algorithm architect with experience in recommendation systems, now focusing on AI applications such as intelligent客服, voice robots, and the CRM opportunity allocation system.