LinkedIn's Self-Evolving Agentic Support System Boosts Self-Serve by 9 Points
LinkedIn has introduced an innovative customer support system that evolves autonomously, merging retrieval-augmented generation (RAG) with evolutionary auto-prompting and a modular evaluation framework aligned with production needs. This system is tailored for fast-paced enterprise settings where policies, product features, and knowledge bases are in constant flux, rendering traditional static assistants inefficient and costly. By implementing a closed-loop, versioned workflow for prompts, retrieval, and evaluation, it facilitates ongoing enhancements without the need to retrain foundational models. Results from offline simulations and ablations showed notable improvements over standard RAG and baseline agents, including fewer inaccuracies and better response completeness. In a two-week randomized A/B test involving LinkedIn's production support, the self-evolving workflow boosted QA self-service by 9.0 percentage points and increased cancellation self-service, demonstrating a major leap in AI-powered customer support.
Key facts
- LinkedIn's self-evolving agentic support system integrates RAG with evolutionary auto-prompting.
- The system uses a modular, production-aligned evaluation framework.
- It enables safe, continuous improvement without retraining foundation models.
- Offline simulations showed reduced hallucinations and improved response completeness.
- A two-week A/B test on production traffic increased QA self-serve by 9.0 percentage points.
- The system treats prompts, retrieval, and evaluation as a closed-loop, versioned workflow.
- Operational guardrails are included to ensure safety.
- The system is designed for rapidly changing enterprise environments.