Overview
The common swift (*Apus apus*) is one of the world's fastest birds in level flight. It eats, sleeps, and even mates in the air, migrating from Africa through India and Southeast Asia to breed in northern China—and can stay airborne for over 300 consecutive days. In June 2025, a core product manager joined DingTalk for a confidential project codenamed “0”—later revealed as ONE. He was the last core PD to join and the last person remaining to send the project off. After 300+ days “in flight” at DingTalk, the time came to land and leave.
This article is a firsthand record of how an AI product that once stood at the strategic center of the AI wave shrank from grand ideal to retreat: how ideals were translated into goals, goals into requirements, requirements deformed through the organization, and finally pushed into real users' hands. As the original says: “The art of calendaring lies in constants, not in variable motion.” DingTalk's high-intensity pace let participants live through the full birth, life, and death of a large-user-base AI product in ten months—capturing both the “variable motion” and the recurring “constants”: structural tensions that look accidental but are in fact inevitable.
Key points
- Joining DingTalk: After Wuzhao's April 2025 return, DingTalk introduced longer work hours, fixed morning/evening meetings, shortened lunch breaks, single-day weekends, and mandatory Python exams; net headcount outflow followed. The author joined after rapid interviews (final round by Wuzhao), including the “family tree on DingTalk” assignment—building a 6+ member active organization. He arrived with three hopes: context, proactive service, apprenticeship. DingTalk seemed like a promised land.
- The 2025 context: AI narrative shifted from chat/summarize/search to tool use, task completion, and real workflows (Google I/O 2025 agents, ChatGPT agent, Claude Sonnet 4.5). DingTalk had organizational relationships, messages, calendars, todos, meetings, docs, approvals—a natural home for workplace AI—but also years of product logic, permissions, and tech debt. Wuzhao's return carried old wounds and old victories; the DING/read-receipt legacy of fighting for sender-side certainty flowed into ONE.
- Fourfold motivation: ONE mixed user motivation (surfacing important matters), product motivation (an AI-era new entry point), organizational motivation (a battle flag for morale), and commercial motivation (token consumption scenarios). Multiple motivations planted every later problem: serve users or serve the launch event?
- Positioning dilemma: Boss vs. employee—paying users are managers who want control; daily users are employees who want relief. Sender vs. receiver—DingTalk's genes stand with senders (strong reach, read receipts), while ONE's vision promised “work finding people.” The product was mis-calibrated from day one on “who am I speaking for.”
- The card war: Design ideals—one card = one complete work item; information convergence, lifecycle closure, permission anchoring—hit three walls: read-status rules (enterprise paid tiers kept read tracking; users called it “AI as the boss's surveillance assistant”), a priority algorithm retaining an 18% weight on job rank (grassroots feeds flooded by superiors' notices), and uncontrolled card scope creep (42 integration requests in May–July 2025, only 8 backed by user research; daily pushes grew from 10–20 to 50+).
- Three-tab scenarios: Morning dashboard degraded into a restyled todo list (no historical work-time data available); evening review limited free users to 300-word summaries due to token cost; the Discover page, originally a value-added opt-in feed, was forced open-by-default with 3–5 daily promotional cards—users disabled ALL ONE notifications to block ads. PC adaptation slipped from 1 to 4 months, weakening cross-device credibility.
- User exposure in three stages: Internal beta and head-customer gray releases produced false confidence—managers loved it, grassroots employees strongly resisted (quantified surveillance, no breathing room); the team misread this as “habit adaptation,” the fatal misjudgment. Public launch: DAU peaked at 3M in week one, then collapsed—next-day retention fell from 45% to 18%, 7-day retention below single digits, 60%+ users closed pushes within three opens, and 1M+ users shut the entry within four weeks amid reviews citing surveillance, involution, oppression.
- The brutal verdict: Bosses need control but don't use it often; mid-levels want efficiency but dare not use it; grassroots reject it absolutely. A product whose payers use it rarely, core users fear, and mass users refuse cannot survive. AI turned management from “people managing people” (with leniency and room for communication) into “algorithms managing people” (absolute standards, quantification, transparency, pressure).
- Agile's death: The fixed August 25 launch date made all schedules subordinate to the event—no requirement freeze, no MVP, no single-point validation, no course correction from negative data (internal beta risks suppressed, gray-release resistance ignored, retention drops met with operations pushes, backlash met with PR control). Extreme busyness masked strategic confusion.
- Organizational decay: Dual-line control (product vs. design with no final arbiter), borrowed cross-department developers whose own KPIs came first, business lines rushing in to meet group AI-conversion KPIs while ONE bore all compute costs, filtered upward information that mistook managers' needs for universal pain points, and misaligned rewards—7 product staff and 5 front-end engineers left during the project.
- War-time costs: Doctrine-bound benchmarking (Feishu AI, Office Copilot) with demo-only capabilities (sub-40% daily real-world usability), four simultaneous battle lines splitting scarce compute and talent, marketing spend dwarfing user research, and token budgets slashed post-launch, cutting free-tier AI capabilities.
- Long-term lessons: The direction—proactive AI work feeds, “work finds people”—was right, and ONE built real underlying capabilities (cross-module message capture, org-permission-bound LLMs, multi-end AI card rendering). But: sender-first control genes conflict irreconcilably with employee relief under current B2B payment logic; launch-KPI deadlines deform long-cycle AI products; a product cannot simultaneously serve brand, monetization, group KPIs, and user efficiency. The author proposes layered B2B AI design (management controls as paid add-ons, employee tools as lightweight independent versions) and MVP-style, small-entry-point validation over event-driven grand launches.
Closing
A product is never a simple act of executing a blueprint; it is the joint product of business, organization, human nature, technology, and industry cycles. Any top-level vision detached from real user psychology, organizational realities, and technical feasibility will remain suspended on launch-event slides. Products that violate ordinary people's goodwill and the basic sense of workplace breathing room will eventually be eliminated; shortcut-driven crunch can prop up data and spectacle short-term, but the market always collects repayment with interest. The swift cannot fly forever—it must eventually land, just as ONE rose on the wind and slowly contracted. If these missteps become a cautionary reference for those who follow, that is the full meaning of these 300 days of flight.
------
References
1. Teng Yaxin (Yousu). “Inside DingTalk: Firsthand Experience and Review of the DingTalk ONE Project” [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.