Every AI conference claims its enterprise sessions are about real deployment. Few hand the microphone to the people who had to make a daily 40-minute reporting chore disappear.
At Baidu World 2026 on 12 November at the National Convention Centre in Beijing, the afternoon open course did something closer to a curriculum than a product launch: more than 40 sessions, each 30 minutes of walkthrough plus audience questions, streamed for six hours from the exhibition floor and online, drawing over 450,000 participants on-site and online combined (Beijing News, 16 November 2026).
Key takeaways
- The open course attracted over 450,000 online and offline participants across 40-plus AI classes, taught by outside experts, developers and Baidu's own product and engineering staff (Beijing News, 16 November 2026).
- The conference's theme, "The Next Act: Intelligent Organisations" (下一幕·智能组织), framed the course: Baidu founder Robin Li argued that if AI is the new electricity, the intelligent organisation is the new factory floor where productivity actually happens.
- A recurring operations agent chained data extraction, validation, chart generation and push delivery, turning a daily report that took 40 minutes of staff time into an automatic delivery.
- The most-scrutinised content was methodological: three sessions dissected the forward-deployed engineer (FDE, 前置部署工程师) role and ontology-based system design, with an energy-industry case laying out a four-step path — scenario discovery, system connection, application development, asset reuse (场景发现—系统连接—应用开发—资产复用).
- Baidu also put its OCR stack on stage: PaddleOCR, one of China's widely used open-source projects, plus a Wenxin Unlimited OCR 2.0 release claiming thousand-page document parsing at state-of-the-art accuracy on limited inference resources.
The course structure is the message
The sessions were organised into three rooms: AI WHY for research and trend decoding, AI HOW for tools and real project teardowns, and AI WOW for finished, cross-domain applications. The sequence — from "why" through "how" to "what shipped" — maps onto the actual difficulty curve enterprises report: understanding models is now the easy part; connecting them to systems and budgets is the hard part.
Instructors were mixed by design. Outside experts and developers taught alongside Baidu's senior product and R&D staff, which kept at least some sessions from becoming vendor demos. Organisers said the three rooms ran full for all six hours.
The cases worth stealing
Four deployment patterns surfaced repeatedly:
- The daily-report agent. An operations agent that chains four steps — pulling data, validating it, generating charts, pushing the result — turned a report that consumed 40 minutes of someone's day into an automatic delivery. The value is not the AI; it is the elimination of a repetitive, low-judgement task that previously depended on one person's diligence.
- The customer-service agent that scaled. A service agent described as having moved from pilot to multi-site rollout, now deployed at major telecom operators. The pilot-to-scale transition is where most enterprise agents stall, so the case drew questions about what changed between the two stages.
- The ontology trick. Baidu's "Shengsuan" platform uses a business ontology to move agents from answering questions to understanding the domain — the difference between a chatbot over your data and a system that knows what a valid order looks like.
- Coding agents as teams. The Multica tool organises coding agents into multi-role teams that run requirement development and bug triage end to end.
FDEs and the four-step path
The most-watched thread was methodological. Three separate sessions took apart the forward-deployed engineer role — engineers embedded at the customer site to find, wire and harden AI use cases — and the ontology concept behind them. Two real projects anchored the theory: an AI-supported precision-diagnosis system and a restaurant-scheduling system, presented as a deliberately "large and small" pairing to show the same method at different scales.
The energy-industry case generalised the method into four steps: discover the scenario worth automating, connect the existing systems, develop the application, then convert what was built into a reusable asset for the next scenario. That last step — asset reuse — is where deployments compound or stall, and the session walked through five delivery cases to show the chain.
Adjacent sessions extended the pattern: digital-human deployment across industries plus an "AI-native organisation" staffing model where every role builds with AI; a medical AI system (the Harness engine) built on a data-closed-loop, model-iteration and effect-optimisation cycle; and a supply-chain operations agent built from scratch, whose hardest problem turned out to be trust — users wouldn't act on its outputs until the validation layer was visible.
The honest read on the 450,000
Attendance figures for open courses include online viewers who may have watched minutes, not hours, and "40-plus classes" spans sessions of very different depth. Baidu selected the cases, and every case taught is a case that worked; the course format structurally hides the projects that stalled. Robin Li's "intelligent organisation" framing is a corporate thesis — useful, but it is the vendor's own framing of where value accrues, and it happens to sell the platform and services that build such organisations.
Still, the format choice itself carries information. A vendor that opens its conference with teardown-style classes, names the failure modes (the supply-chain agent's trust problem, the pilot-to-scale gap), and teaches a reusable method rather than only shipping products is responding to a buyer audience that has moved past demos. The four-step path is not proprietary knowledge, and readers can apply it without buying anything Baidu sells.
Honest limitations
This article draws on the Beijing News report of the open course and related coverage of the conference; figures for attendance and course counts are organiser-stated. Case metrics — the 40-minute daily report, rollout claims for telecom customers, the thousand-page OCR benchmark — are presented by the vendor without independent verification or published baselines. The FDE methodology is Baidu's articulation of a role that other vendors (and systems integrators) define differently. No information was available on how many course attendees went on to deploy anything, so the course's real-world effect is unmeasured.
What readers can do now
- Run the four-step audit on your own organisation: list which AI use cases have completed scenario discovery and system connection, and note that most stall before the asset-reuse step — that list is your roadmap.
- If you are staffing AI delivery, study the FDE role as a job description rather than a buzzword: engineers who sit with operators, wire systems together, and document what becomes reusable.
- When a vendor shows an agent case, ask the two questions this course surfaced: what broke between pilot and scale, and what validation layer made users trust the output enough to act on it.
