How Baidu’s “Sheng Suan” Turns Agents from Outsiders into Business‑Savvy Assistants

AI‑generated code can handle over 90% of routine programming, yet enterprises remain hesitant to let agents make supply‑chain replenishment or store‑location decisions because agents typically achieve only 80%‑90% accuracy on read‑only queries and lack the confidence to control core production systems.

According to Liu Bin, general manager of Baidu Intelligent Cloud Data Platform, the limitation stems from agents being “smart outsiders” that understand language but not the intricate entity relationships and decision logic embedded in enterprise processes. Conventional Retrieval‑Augmented Generation (RAG) merely feeds documents into a vector store; while the model can retrieve text, it cannot grasp complex rules such as “who should be ordered from and how much when stock runs low.” Consequently, agents are confined to read‑only operations like report lookup, and enterprises retreat when write‑operations such as order placement or inventory transfer are required.

Baidu’s newly announced enterprise data‑intelligence platform “Sheng Suan” tackles this bottleneck through two complementary pillars: context engineering and system engineering.

Context engineering introduces a “business ontology” composed of three layers:

Business knowledge : structural representation of orders, products, warehouses, suppliers, etc., forming a “business map” that tells the agent the relationships among entities.

Business logic : formalization of expert decision rules (e.g., when a stockout occurs, which supplier to order from, how many units to replenish), converting tacit experience into executable logic maps.

Business execution : API integration that enables the agent not only to view data but also to perform actions such as placing orders or processing returns, thereby breaking the read‑only barrier.

Liu Bin likens the ontology to an onboarding training for a fresh graduate, providing the company’s regulations and procedures so the agent knows “who to contact when something goes wrong.”

System engineering ensures that granting agents write access does not jeopardize production stability. Baidu builds a sandbox‑like operating‑system layer that offers:

Isolation so agents can explore without affecting live systems.

Full‑link audit that records every decision step for traceability.

Forward simulation to predict business impact before execution.

Back‑trace capability to retrieve the rationale behind a decision.

These under‑the‑surface capabilities, often overlooked in hype‑driven discussions, become the decisive factor for enterprise adoption.

Three real‑world deployments illustrate the impact:

A large manufacturing client reduced material‑shortage decision time from months to minutes and eliminated missed shortages entirely.

A retail chain achieved minute‑level performance‑analysis, unified evaluation standards across heterogeneous stores, and attained over 95% root‑cause accuracy, pinpointing issues such as “high‑performer turnover.”

A new‑energy manufacturer raised multimodal document recognition accuracy above 95% and accelerated processing cycles to the hour level.

All three cases share a common outcome: agents moved from peripheral Q&A to core production workflows.

In Baidu’s view, the next frontier for large models is not larger scale but richer, structured business knowledge. The “Sheng Suan” approach answers the pivotal question: when AI becomes smart enough, how can it evolve from a “clever outsider” to an “insider that understands the business”? The answer lies in structuring business context rather than merely expanding model size.