After ONE: DingTalk's AI Organizational Experiment and Its Lasting Legacy

Published: June 7, 2026
Reading time: ~3 minutes

---

What was claimed?

In April 2025 DingTalk’s returning CEO Wu Zhao announced ONE, a single AI‑driven hub that would replace the fragmented mess of group chats, meeting invites, approval flows, task boards, document libraries and calendars. The public launch in August 2025 was accompanied by a bold promise:

• Every employee would open DingTalk each morning and see a curated list of the most important actions – essentially turning “tasks finding people” into a reality.
• ONE would be DingTalk’s flagship AI product, positioning the platform alongside the wave of large‑language‑model (LLM) services emerging worldwide.
• The project would showcase Wu’s vision for a post‑return transformation, proving that DingTalk could evolve from a traditional workflow tool into an AI‑first workspace.

The marketing deck claimed 3 million daily active users (DAU) within weeks, and highlighted a suite of generative features – automatic meeting summarisation, smart approval routing, and a “daily briefing” generated from internal data streams.

---

What actually happened?

An internal document titled Inside DingTalk (circulated widely among Chinese enterprise‑software engineers) provides a more nuanced picture.

1. A mixed‑purpose product

ONE tried to be both a universal daily portal for every user and a deep‑dive solution for specific pain points (e.g., contract‑review assistance, compliance‑check bots). Those goals are at odds:

Goal	Typical metric	Result
Universal portal	DAU, average session length	3 M DAU reported, but average session < 2 min; most users only opened the “briefing” and closed the app.
Deep‑dive solution	Feature‑specific adoption, revenue per seat	Only the smart‑approval bot saw > 10 % adoption; other modules remained in beta with negligible usage.

The data suggests that users appreciated a quick glance at priorities but did not trust the AI to replace their existing tools for more complex tasks.

2. Organizational friction

DingTalk’s DNA is management‑oriented: read receipts, DING notifications, attendance logs, and hierarchical approval flows are baked into the product to give managers visibility and control. ONE’s AI layer accessed the same data streams to generate recommendations, which raised two practical concerns:

• Surveillance perception – employees noticed that the AI‑generated briefings often highlighted overdue tasks flagged by managers, making the assistant feel like a monitoring device rather than a helper.
• Decision‑making ambiguity – when the AI suggested a different approver or reordered a meeting, managers questioned the authority of the recommendation, leading to frequent overrides.

The internal memo notes that the product team spent four months negotiating with the corporate‑services division to define “acceptable AI nudges,” a process that slowed rollout and diluted the original vision.

3. Commercial performance

ONE was bundled into DingTalk’s existing subscription tiers, with a modest $2 per‑seat premium for the AI features. Revenue impact was limited:

• Upgrade conversion: 5 % of existing customers added the premium, well below the 15 % target.
• Churn: Small‑to‑mid‑size firms reported higher churn after ONE’s launch, citing “feature bloat” and “privacy concerns.”

The product’s cost‑to‑serve also rose sharply because the backend relied on a custom‑trained LLM (named DingGPT‑1.0) hosted on Alibaba Cloud GPU clusters, consuming roughly 1.2 kW per 1 k queries, a figure that made the service financially unsustainable at scale.

---

What are the limitations revealed?

1. The “assistant vs. manager” paradox

Enterprise AI sits at the intersection of individual productivity and organizational governance. ONE demonstrated that the same data that makes an AI useful for personal assistance also empowers it to enforce managerial policies. Without a clear governance model, users perceive the AI as an extension of surveillance rather than a neutral helper.

2. One‑size‑fits‑all UI does not scale

A single dashboard that tries to surface all work for all roles inevitably becomes noisy. The briefings were useful for knowledge workers but added little value for field staff who rely on location‑based check‑ins. The internal memo recommends role‑specific AI skins instead of a monolithic view.

3. Model ownership vs. platform stability

DingGPT‑1.0 was trained on DingTalk’s internal chat logs, meeting recordings, and HR records. While this gave the model contextual awareness, it also introduced data‑privacy liabilities and made model updates a legal bottleneck. The engineering team spent more time on compliance reviews than on model improvement, slowing the iteration cycle.

4. Commercial pressure on a research‑heavy product

Charging a modest premium while running a custom LLM on expensive GPU hardware created a negative unit economics situation. Competing products (e.g., Microsoft Teams + Copilot, Google Workspace AI) rely on economies of scale from their massive cloud infrastructures, putting DingTalk at a cost disadvantage.

---

What does this mean for the next wave of enterprise AI?

The ONE experiment suggests that future success will hinge less on raw model size and more on integration strategy:

1. Clear separation of assistance and oversight – design AI nudges that are explicitly opt‑in for personal productivity, while keeping managerial dashboards separate and transparent.
2. Modular AI components – rather than a single portal, expose a marketplace of role‑tailored agents (e.g., “Sales‑assistant”, “Compliance‑checker”) that can be plugged into existing workflows.
3. Hybrid model deployment – combine a lightweight, on‑premise inference engine for privacy‑sensitive data with cloud‑hosted LLMs for generative tasks, reducing cost and regulatory friction.
4. Metrics that reflect enterprise value – move beyond DAU to measure time‑to‑decision, approval‑cycle reduction, and policy‑compliance uplift, which align better with corporate ROI expectations.

If DingTalk can re‑architect ONE along those lines, the platform may still salvage the investment. More broadly, the episode serves as a cautionary case for any vendor trying to graft a consumer‑style AI assistant onto a management‑centric product without rethinking the underlying organizational dynamics.

---

For the full internal memo, see the translated excerpt on the Pandaily tech forum.