DingTalk ONE: A 300-Day Flight and Landing — Inside the Rise and Retreat of an AI Office Product

Forum topic · ✨步子哥 · 2026-06-08

Summary

This long-form first-person retrospective chronicles the roughly 300-day lifecycle of DingTalk ONE, an AI-native workplace product launched at Alibaba's DingTalk under the codename '0' in 2025. Written by the last core product manager to remain on the project, the account details ONE's conception as a proactive 'things find people' information feed built on card-based units, its positioning conflicts between serving bosses versus employees and senders versus recipients, and design battles over read receipts and priority algorithms that favored rank over relevance. The post describes how launch-driven deadlines destroyed agile practices, how internal beta success created false confidence, and how a public launch peaked at 3 million DAU before next-day retention collapsed from 45% to 18% amid user backlash against algorithmic workplace surveillance. Systemic failures included dual-line management without arbitration, cross-department resource competition driven by group AI KPIs, and scattered focus across four fronts. The author concludes with lessons for China's SaaS+AI industry: B2B collaboration AI must not become a control tool, paid tiers should separate manager oversight from employee productivity, and AI products need small-scope MVP validation rather than launch-ceremony-driven mandates.

This is a translated and condensed summary of a Chinese forum post: a first-person account by the last core product manager on DingTalk's secret project '0' (later revealed as ONE), covering the product's full lifecycle from conception in June 2025 to wind-down.

Background

The author joined DingTalk in June 2025 after its controversial return of former lead Wuzhao (无招), who had reinstated long hours, fixed standups, shortened lunch breaks, single weekend days off, and mandatory Python exams.
The author's motivation: LLM services only work with sufficient data infrastructure and user context — and DingTalk uniquely holds organizational context. DingTalk risked elimination without AI; with AI, the enterprise messaging table could be reshuffled.
The hiring process included a 'family tree on DingTalk' assignment requiring a 6+ person active organization; the final interview was conducted by Wuzhao himself.

2025 Context and the Leader's Mirror

Mid-2025 saw the AI narrative shift from chat/summarize to agents executing tasks in real workflows (Google I/O 2025's AI Mode and Agent Development Kit, OpenAI's ChatGPT agent, Anthropic's Claude Sonnet 4.5 with computer use).
DingTalk had full-chain assets (messages, calendar, todos, meetings, docs, approvals) but carried years of product logic, permission systems, and technical debt.
Wuzhao's 2014 experience — standing on the sender's side, enforcing certainty through read receipts and strong push — shaped ONE's DNA. His return after failures at Laiwang and the HHO venture produced a fateful conviction to win again with AI.

Four Mixed Motivations (a Structural Flaw)

1. User: important work items should surface themselves, reducing searching and missed messages. 2. Product: DingTalk needed a new AI-era entry point proving it could reorganize work. 3. Organizational: a flagship battle to rally morale and rebrand after Wuzhao's return. 4. Commercial: consuming tokens and showcasing model capabilities in real scenarios.

Good products usually have one primary motivation; ONE's mixed seeds caused every later tradeoff.

Positioning Dilemmas

Boss vs. employee: paying users are managers (needing control and visibility); daily users are employees (needing boundaries and breathing room). The team shipped with a 'Schrödinger's user'.
Sender vs. recipient: DingTalk's genes favor senders with strong reach; ONE's vision of 'things finding people' conflicted with this at the root.

The Card Battles

ONE's form: everything as cards, each card ideally one complete work item with converged information, closed lifecycle, and anchored permissions.
Read receipts: management insisted on read tracking for paying enterprises; the compromise (paid = visible stats, free = hidden but still recorded) led users to call ONE 'the boss's surveillance assistant'.
Priority algorithm: initial weights — urgency 35%, related person's rank 20%, deadline 30%, interaction history 15%. Rank weighting pinned executive messages and buried peer collaboration; a reduced 18% still skewed feeds toward hierarchical notices.
Scope creep: business lines (approvals, meetings, docs) demanded card integration for KPI reasons; daily pushes ballooned from 10–20 in beta to 50+ after gray release. Information overload increased.

Three Scenario Tabs and Their Distortions

Morning dashboard (8:00): lacked historical work-time data and cross-app permissions; users saw it as a reskinned todo list.
Evening review (18:30): token costs limited free users to 300-character summaries; only paying enterprises got full documents.
Discover page: originally optional industry-content recommendations; operations added default-on ads and SaaS promos (3–5 commercial cards daily), driving many users to disable ONE entirely.
Multi-device: PC integration required heavy refactoring (schedule stretched from 1 to 4 months), creating cross-device fragmentation.

User Rollout: From Greenhouse to Public Failure

Internal beta produced inflated positive data and a false sense of mission.
Head-customer gray release: managers loved it; frontline employees rebelled against total visibility and quantified monitoring — a resistance the team fatally misread as 'habit adjustment'.
Public launch: DAU peaked at ~3 million in week one; next-day retention crashed from 45% to 18%; 7-day retention fell below single digits; over 600,000+ users disabled card pushes within three opens. Reviews centered on surveillance, involution, oppression.
Post-mortem core abandonment reasons: elimination of all workplace gray-zone buffering, rank-based priority crushing, and noise exceeding value. Executives needed control but wouldn't use it daily; middle managers feared using it; frontline staff rejected it outright.

Process and Organizational Failures

The fixed August 25 launch date inverted agile practice: no requirement freezes, no MVP, full-scope full-stack delivery, military-cadence meetings, and zero mechanism to validate output. Negative signals were suppressed at every stage (beta risks downplayed, gray-release resistance dismissed, retention decline answered with operations pushes, public criticism with PR suppression).
Dual-line reporting (product vs. design center) with no arbitrator caused endless standoffs; cross-department developer staffing meant ONE was always second priority to existing roadmaps.
Business lines rushed integration to meet group-wide 2025 AI retrofit KPIs: 42 integration requests from May–July, only 8 validated by user research. Compute costs were borne solely by ONE.
Negative feedback was filtered upward; honor accrued to the top while layoff risk flowed down. 7 product staff and 5 frontend engineers left during the project.
Competitive dogma (copying Feishu AI and Office Copilot feature sets) and four simultaneous fronts (mobile, PC, hardware recording cards, government/enterprise private deployments) scattered resources; demo data was pre-staged, with sub-40% real-world availability for demoed features. Marketing spend dwarfed field research.

Long-Term Reflections

Right direction: proactive AI work feeds and 'things finding people' remain the industry's consensus evolution path; DingTalk built real capabilities in cross-module message capture, org-permission-bound LLMs, and multi-device AI card rendering.
Systemic causes of failure: DingTalk's sender-first, control-heavy DNA vs. an employee-relief promise; launch-KPI timelines crushing AI's natural iteration cadence; one product carrying brand, monetization, group KPIs, and user value simultaneously — none landing.
Industry lessons: B2B collaboration AI must not become a control instrument; design separate tiered systems (manager oversight as paid add-on, employee productivity tools as a lightweight independent version); avoid campaign-style launches — start from a minimal MVP wedge and expand after value validation.

> Product is never a simple blueprint-to-build exercise; it is the joint product of business, organization, human nature, technology, and industry cycles. Any product that violates ordinary people's goodwill and basic workplace breathing room will eventually be eliminated.

The swift of the title — a bird that can fly 300 days without landing — eventually must touch ground, as ONE did.

References

1. Teng Yaxin (Yousu). 置身钉内：钉钉ONE项目亲历与复盘[EB/OL]. 2026-06-05. 2. Google. Google I/O 2025: AI Mode and Agent Development Kit announcements[R]. 2025. 3. OpenAI. Introducing ChatGPT agent: tool use and task completion capabilities[R]. 2025-07. 4. Anthropic. Claude Sonnet 4.5: Advances in complex agents and computer use[R]. 2025-09. 5. Stack Overflow. Developer Survey and Monthly Question Volume Trends 2022-2025[R]. 2025.