Ant Group rolls Qianwen Office out to every employee
Two of China's biggest tech names tied their office stacks together today, and the way they did it is the story: not a cloud deal, but a private install inside Ant's own network.
Ant Group has adopted Alibaba's Qianwen Office as the agent base for its entire workforce, with the product running on Kubernetes inside Ant's intranet and the group's own models available alongside it. Ant, which employs tens of thousands of people across payments, fintech and health services, says the arrangement keeps data inside its domain, logs auditable and conclusions verifiable — the compliance shape that has kept most large Chinese enterprises away from public AI assistants, and the reason a company holding payment and medical data can put an agent in front of every employee at all. Qianwen Office will also be wired into Ant's own agent stack: its digital-employee programme, the internal Ant Ding messaging and project management services, and the group's AI security controls.
Ant's platform technology group president Luo Ji framed the joint build as a reproducible template for complex enterprise office scenarios, which is roughly what it is — a reference architecture other large buyers can copy rather than a product Ant bought off the shelf. The competitive read is less flattering to the "office app" framing. Qianwen Office is a month old with more than 30 million users, over half of them on enterprise accounts, and Alibaba's last move in this fight was adding rival models to the same product — Alibaba's Qianwen Office adds GLM-5.3 and DeepSeek V4 Pro. Whoever owns the workspace owns the routing decision, and Ant just handed that layer to Alibaba.
China Telecom's in-house general agent, TeleAgent, took third place in IDC's enterprise agent evaluation, scoring full marks on task performance and the highest cost-efficiency score in the test. The evaluation ran close to a hundred undisclosed tasks — cleaning 3,755 rows of messy data, distilling an 11-page deck out of roughly 30,000 words of source material — and then verified system state and the delivered files instead of grading the written answer. TeleAgent has picked up nearly 1.2 million users since its July launch, with a context window past 400,000 tokens, about 30,000 lines of self-written Go at its core, and a routing layer that splits work between light, mid-tier and flagship models to cut inference cost by roughly 40% on the company's figures.
The interesting part is the shape of the win: IDC's own April survey found 48.5% of enterprises already evaluating, piloting or using general agents, and 96.9% of those deployments touching office automation. A state-owned carrier shipping a consumer-paced agent product — desktop trial in April, public release in July, 56,000 employee-contributed skills since — says the enterprise agent market is now being fought on harness engineering and unit cost, not on which frontier model sits underneath.
Graphics researcher Xin Tong has joined AI 3D company Meshy as chief scientist after 25 years at Microsoft Research Asia, where he led its internet graphics group. Tong's published work carries more than 21,000 citations and fed into Xbox development APIs, Windows' 3D printing driver and Direct3D tooling; at Meshy he owns long-term research strategy while the company pushes a thesis it calls "AI for Fun" — building worlds people can walk into rather than assets they download. Meshy says it has 12 million users and customers at half of the world's ten most valuable companies, and raised close to $400 million in July at a post-money valuation above 10 billion yuan.
The hire is the tell on where this is heading. Meshy's roadmap is not better triangles but a replacement for the triangle-based rendering pipeline, with a coding agent assembling world skeletons and video models painting the output — which is why a career spent on physical rendering and material capture is suddenly a candidate-selection asset for a generative video and 3D company.
What to watch: whether Ant's private-deployment template gets copied by other Chinese enterprises, and whether TeleAgent's cost lead survives a second IDC round with Western agents in the same test.
Enterprise AI adoption in China is being decided by deployment topology, not model quality — does that pattern hold in your company? Tell us in the comments.
