In late summer 2026, three giants folded Agents toward an “office entry” in the same window. Tencent’s WorkBuddy shipped human–AI co-editing; Alibaba merged QoderWork, Wukong, and MuleRun into Qwen Office for public beta; ByteDance folded Feishu product into Doubao and GTM into Volcano Engine, with Doubao Enterprise riding Feishu’s B2B reach.

TMTPost’s “产业家” piece nails the substrate: not only smarter models, but scenes, ecosystems, and a commercial system that turns Tokens into sellable products. 36Kr’s Zhidx timeline pins the org moves to the same weeks. Skip funding gossip. Tear down two things: how the three products differ, and seat subscription vs Token burn.

Mechanism 1: three product paths, three default doors

Product	Public shape	Scene assets behind it	Recent moves (public)
WorkBuddy	Desktop/multi-end office agent; NL over local files, docs, browser	Tencent Docs and collab stack; shared cloud/security	Standalone app on three mobile OSes in July; “dual writing” on Jul 30 inside Word/Excel/PPT/Markdown
Qwen Office	Org-facing one-stop AI productivity; state a goal → plan → deliver files	DingTalk org, permissions, approval; desktop + cloud orchestration heritage	Public beta Aug 3; three Agents no longer sold alone; DingTalk CLI/APIs for Agents
Doubao Enterprise	Productivity assistant tied to Feishu suites; sales push image/video gen and browser agents for dashboards	Feishu enterprise scenes + Doubao models + Volcano B2B sales	Feishu product under Doubao; GTM under Volcano “Creativity Service Platform”; enterprise beta with some Feishu accounts

WorkBuddy’s breakout story is individuals first: Analysys monthly visits from ~8.85M in March to ~20.97M in June; Tencent’s quarterly language calls it among the widest-used efficiency AI agents by DAU in China. The product bet is “open and work”—multi-model, Skills, local tools—for non-coders. Dual writing pulls the agent out of a side chat and into the file you are already editing.

File-native means the collaboration object is document state, not chat history. Human edits a cell, AI edits a cell—conflicts are visible, rollback looks like office software. For individuals and small teams, that cuts paste-back friction; for enterprises, it hits permissions immediately—can AI edit controlled docs, and do edits enter audit logs? WorkBuddy’s public heat proves use threshold, not production readiness.

Qwen Office bets on an org door: not a C-end toy, an enterprise productivity platform. Desktop execution (ex-QoderWork), cloud workflows (ex-MuleRun), and org/permissions (ex-Wukong/DingTalk) are supposed to live in one product. Coverage via Huxiu and others says near-term KPIs de-emphasize MAU and prioritize enterprise scenes—classic trust and effect before scale.

Merging three Agents into one brand costs learning curve and positioning clarity: users must grasp “goal → plan → deliver files,” not three point tools. The gain is reusable org context—approval flows, dept permissions, existing DingTalk identity—without rewiring each Agent. In public beta, watch human takeover on long-task failure, least-privilege across systems, and whether deliverables land in enterprise knowledge stores instead of personal download folders.

Doubao Enterprise looks most like selling model capability through existing B2B tentacles: 产业家 describes Feishu assistant promos bundling Doubao Enterprise with Feishu Navigator; sales pitch Seedream, Seedance, and browser-agent dashboards to OPCs. Zhidx cites >90% of newly signed Feishu accounts also buying AI modules—demand forcing “collab tools must grow AI.” Org-wise Feishu yields and Doubao rises: the story shifts from “AI inside Feishu” to “Doubao Enterprise with Feishu as the scene.”

Sales bundling ships fast: the customer already buys Feishu; AI rides the contract. The weakness is expectation management after deep use—demo image/browser agents are not the same slide as production permissions, data residency, and long-task cost. A 90% attach rate proves “willing to buy,” not “willing to hand over core workflows.”

Counter-intuitive: the war looks like three chat apps. The real fork is which system the first screen talks to by default—local files and Tencent Docs, DingTalk identity/permissions, or Feishu + Volcano’s sales list. Model gaps are repeatedly called hard to widen in coverage; scene context is where paths diverge. Pick the wrong door and the agent performs in a vacuum without identity or file context.

Mechanism 2: seat price vs Token price — different ledgers

Office Agents are not yet a clean large business—36Kr’s “躺姐指数” is blunt: personal-side heat proves the use threshold; enterprise budget, permissions, and long-task cost decide scale. Public pricing language already splits.

Seat / suite logic (classic SaaS): sell seats monthly; pack AI into the suite. Feishu/DingTalk buyers already know this; WorkBuddy’s personal side often grows on free quota and suite buzz, then must land on org budgets. Seats are easy to buy; the risk is long Agent runs whose compute blows past seat gross margin—AWS illustrative math cited in that piece shows multi-tool turns lifting inference cost several- to ten-fold versus a single tool call. Fixed seats make the vendor eat volatility.

Token / outcome logic (产业家’s “Token economy”): giants do not only want to be pure Token resellers; they want Tokens processed into products and ecosystem burn inside their perimeter. Sales reorg means cloud + collab reps can sell across the stack; KPIs and price cards drift toward Token consumption. Doubao Enterprise under Volcano sales and Qwen Office inside Alibaba Token Hub (ATH) sit on that ledger. For customers, the bill may shift from “how many seats” to “how many tasks, how many Tokens, whether to top up.”

The two ledgers read differently to procurement. Seats fit budget tables: heads × price, finance can sign. Tokens track real cost: heavy users pay more—but feel like a cloud bill, with numbers clear only at month-end and business teams happy in demos, fighting at reconciliation. Mix (seat gate + Token pack) is the common compromise: seats for access, Tokens for usage. If the contract is silent on long-task caps, overage policy, and pause-before-overdraft, the mix becomes a surprise bill too.

Product shape forces quote shape:

• WorkBuddy can keep winning habit with free quota; enterprise needs permissions, audit, managed agents—or seats sell while production stays off-limits.
• Qwen Office, if it stays org-native, likely mixes seats/private cloud with Token packs: DingTalk base for trust, Tokens for usage.
• Doubao Enterprise already sells capability lists in sales scripts and can ship suite-includes-AI first, then surface Token bills in deep use—Feishu’s high AI attach rate is a bundling greenhouse.

Founders should not only chase DAU. Personal noise builds entry mindshare; a copyable business needs production-workflow share, human takeover rate, expansion renewals, and templatized delivery (the investor checklist in 躺姐指数). At the product layer, who finishes an interruptible job first; at the commercial layer, who writes long-task cost into a price card customers can read.

Vertical Agent / services opportunities sit in the fork: giants want ecosystem Token burn; if domain context becomes distributable Skills / expert packs, you can hang on the entry; if you only ship a generic chat shell, the entry taxes your pricing power. Over-depending on one entry for distribution also leaves you exposed when share or default models change.

Sources

• What decides the AI office entry war (TMTPost / 产业家)
• Alibaba, ByteDance, Tencent push AI office (36Kr / Zhidx)
• Why office Agents are not yet a big business (36Kr)