Teams preparing to put agents into business operations should resist opening a new AI project in every department. At Baidu World on November 12, Robin Li promoted the idea of an intelligent organization. For deployment owners, his account of Baidu's mistakes offers a more immediate lesson: enthusiasm for AI does not guarantee well-executed work.
Decentralized experiments created maintenance problems
In Yicai's reporting from the event, Li described engineers fine-tuning open models for their own business needs. Early results were useful, but subsequent foundation-model improvements left those efforts behind. Baidu then assigned training to a dedicated model team rather than individual business units.
Agent development brought another version of the problem. Departments pursued fashionable applications, spreading resources across too many directions. Li said monthly reviews pushed teams to narrow their commitments.
Define the job before choosing the software
Baidu's announcement sets out three principles: organize tasks across job boundaries, make employee know-how reusable, and improve workflows through measurable feedback. Deployment teams still have to translate those principles into specific tasks and responsibilities.
Our recommendation is to start with a tightly bounded task. Consider a hypothetical customer-service drafting pilot: give the agent a customer's question and approved service documentation. Permit retrieval and drafting, but not sending replies or changing orders. For acceptance, every policy statement must be traceable to the documentation, with no unauthorized refund or compensation promises. Missing evidence, conflicting documents or requests to change an order should trigger a handoff to a designated support employee. Written into the task specification, these checks are easier to apply than an instruction to answer customers accurately.
Track more than model charges. Data integration, workflow maintenance, human review and rework after failures belong in the pilot record too.
Human review is a valid deployment choice
Microsoft's earlier framework for human-agent teams offers a useful cross-check. It asks how many agents a task needs, how many people must guide them and when people should remain accountable for consequences. Where reliable acceptance checks are missing, keeping AI in a drafting or advisory role is a reasonable alternative to granting execution rights.
At the next pilot review, ask for a record of completed work and exceptions, not just a count of agents. Establish whether the job actually ended successfully before funding a broader rollout.