Baidu’s AI agents face a bigger test than adoption: changing how organisations work

Most enterprise AI agent projects still focus on general tasks, while many risk stalling without clear returns. Baidu is betting that agents connected to existing systems and specialist expertise can reshape business processes, and retain useful learning.

Enterprise AI agents have spread quickly, but their use remains concentrated in relatively general tasks. iResearch figures cited by 36Kr show that such uses made up 83.7% of enterprise agent projects between 2025 and the first quarter of 2026, while applications tackling specialised business processes accounted for 16.3%. A separate estimate from the research points to another hurdle: around 30% to 40% of projects could stall or be discontinued within 18 months if returns fall short. Together, the figures suggest that adoption alone is a poor measure of whether AI is changing how an organisation works. (This paragraph cites sources in prose; source map follows.)

At Baidu World 2026, Baidu founder Li Yanhong presented the next stage of AI development as the emergence of the “intelligent organisation”. As reported by Xinhua Finance, his proposed principles included rethinking boundaries between jobs, turning employees’ tacit know-how into shared organisational knowledge, and building feedback into business processes. The argument shifts attention from individual productivity tools towards how people and AI divide work, and whether each completed task can make the next one better informed.

That ambition faces a practical test: agents must work within the systems businesses already rely on. Customer renewals, for example, can involve checking sales opportunities in a CRM, reviewing orders in an ERP system, assessing performance data and seeking approval through office software. Baidu’s DuMate is designed to connect with existing applications and services, rather than require companies to move their operations into a separate environment. Baidu’s development platform documentation describes connections through MCP and other tools, while the existing systems continue to provide the underlying business information and access controls.

A demonstration described in the supplied coverage showed DuMate coordinating five applications and seven rounds of task execution to prepare a customer renewal proposal, assess risk and submit an approval request. Another example, reported by Aichaai, concerns Fujian Blue Ocean Blackstone, which said customised procurement agents reduced work that had involved two teams and at least ten people spending 100 hours a week to a process completed by two people in under an hour. That is a company-reported result, not an independently audited benchmark, and should not be assumed to translate directly to other businesses.

Connecting software is only part of the challenge: an agent may retrieve information without knowing how to interpret it in a specialist setting. Baidu’s approach is to combine its general agent infrastructure with expertise supplied by industry partners. IT Heat describes three partnership routes: linking partner tools and skills to DuMate, building shared workspaces, and developing joint products for particular sectors. In education, Century Tianhong’s Xiao Hong Assistant is positioned to handle teaching tasks such as lesson design and marking, while DuMate provides broader task coordination and office support. Wabei presents the arrangement as a shared workspace, but its report does not independently establish educational outcomes.

The path from a successful demonstration to dependable production use is more demanding still. The iResearch analysis reported by 36Kr identifies permissions, data quality, business rules, unusual cases, operating costs and maintenance as factors that can undermine deployment. Baidu’s product documentation describes a development environment for building, deploying and evaluating agents, with tools including workflows, knowledge bases and reusable Skills. The company also describes its Agent API as providing access to managed infrastructure for agent operation. These features may support development and oversight, but the wider evidence on project attrition underlines that technical capability alone does not establish lasting business value.

Baidu is also seeking to make specialist capabilities available through partners rather than building every industry tool itself. The company’s announcements include the Harness engine and Agent API, while Xinhua Finance reported that its model marketplace, Yuan Yuan Suo, is intended to help users find and use models, with verification and provider-rating features. The product documentation additionally describes evaluation for quality, hallucination and safety, alongside public- and private-cloud deployment options. These are platform provisions and announced aims; their effectiveness in particular organisations will depend on implementation and governance.

The longer-term proposition is that completed work should leave something useful behind. Baidu describes a cycle in which an agent retains relevant context, refines accumulated experience and turns proven methods into reusable Skills. Its product documentation supports the practical possibility of publishing and maintaining Skills for use by multiple agents. The distinction matters: stored information is not automatically organisational expertise, and a reusable procedure is valuable only if it works reliably when applied again. The iResearch findings reported by 36Kr similarly emphasise the need for ongoing calibration and for data, knowledge and agents to reinforce one another.

The measure of progress, then, is not simply how many agents an organisation has launched. It is whether they can operate within existing systems, draw on appropriate professional knowledge, complete work under clear controls and help preserve learning for future tasks. Baidu’s platform and partnership strategy is one attempt to address those requirements. Whether it delivers lasting change will depend on results in real operating conditions, and on organisations’ willingness to reshape processes and responsibilities alongside the technology.

Source Reference Map

Inspired by headline at: ^[1]

Sources by paragraph:

Source: Fuse Wire Services