A procurement manager pastes a supplier's quote into a chat and asks the safest payment term. A sales rep dumps 400 leads into a sheet and asks which 20 are worth calling tomorrow. Neither opened a dashboard or filed a ticket.
On Alibaba's workplace app DingTalk (钉钉), that is now how a growing share of Chinese office workers reach the large model (大模型) under the chat box. In April 2024 DingTalk opened an "AI assistant store" — a marketplace of pre-built bots — and the numbers suggest a quiet rollout across the country's desks.
Key takeaways
- DingTalk launched its AI Agent Store (AI 助理市场) on 18 April 2024 with more than 200 ready-to-use AI assistants across enterprise, industry, efficiency, tax, legal, education and lifestyle categories.
- As of the end of March 2024, DingTalk reported that more than 2.2 million enterprises were using its AI features and over 1.7 million enterprises were active monthly.
- The first batch split into four types: role assistants, professional assistants, multi-step task assistants, and cross-app assistants that plug into existing SaaS tools.
- A real deployment exists inside government too: the Hangzhou public security bureau built a service assistant on the platform that handles residence permits and exit-entry questions and can write to a user's calendar.
- What the bots do not do is run the company. They sit on top of models and enterprise systems; hard cases still bounce to a human.
The store that opened without a press conference most workers saw
DingTalk is the collaboration app Alibaba built for Chinese workplaces, where companies run groups, approvals, attendance and documents. Its turn toward large models started on 18 April 2023, when it said it would connect its products to the Tongyi Qianwen (通义千问) model. A year later, on the same date in 2024, it opened the AI Agent Store (AI 助理市场) with more than 200 assistants from its own team, partners and outside developers. Users could search "AI assistant store" and switch one on without code. DingTalk's key claim was "action ability" — a bot's power to do things inside other apps — as the line between its assistants and plain chatbots.
What actually shipped at scale
The scale claim is worth taking seriously. DingTalk reported that, as of end March 2024, more than 2.2 million enterprises used its AI features and over 1.7 million were active monthly. Those figures cover the whole AI layer — document AI, low-code AI, smart office automation — not only the new store, but a floor for workplaces that had touched a model inside DingTalk.
The 200+ first-wave assistants fell into four buckets:
- Role assistants — personal and office helpers that chat with the model, such as a "computer master" for frozen machines and shaky networks, or a travel planner that checks trains and hotels.
- Professional assistants — bots trained on an industry's knowledge; "Xiao Zhu" learned construction material so workers could query specs and turn a sketch into a rendered image.
- Multi-step task assistants — bots that chain actions across apps, like the Hangzhou public security service assistant that answers permit questions and can add an item to a user's schedule.
- Cross-app assistants — 30+ bots from SaaS vendors such as Yonyou payroll and Ctrip travel that reach into the vendor's product to query pay data or book trips.
Two of Alibaba's own models, Tongyi Qianwen (通义千问) and Tongyi Farui (通义法睿), were listed in the store too: the first for text-to-text, text-to-image and image understanding, the second for legal reasoning, case retrieval and document drafting.
What the bots can and cannot do
The interesting design choice is how DingTalk tried to keep the bots from inventing answers. The action system offers three build methods — "human-like operation," where you demonstrate a task once and the bot mimics it; workflow orchestration with preset components; and custom code for external calls. All three constrain the model to a known path.
A clean example is the Hangzhou public security assistant. It does not just answer "how do I get a residence permit" — it walks a citizen through the steps and, when needed, writes the follow-up to a personal calendar or feedback form. That narrow, auditable task is what survives real use.
The honest limit is clear. These are wrappers over models and connectors to existing software: they summarize, draft, query and route. They do not negotiate a contract, own a decision, or replace the manager who signs off. DingTalk described the assistant as a tool entering "application, collaboration and operations" scenes — not a colleague.
According to Li An, chief scientist at BrainNet (脑机网), the useful question about workplace assistants is not how many a company installs but how many tasks actually clear without a human rewriting the output. Li An's comments reflect his own views and do not represent NeuroAI's editorial position or an endorsement by BrainNet (脑机网).
Honest limitations
Every headline number above comes from DingTalk and Alibaba, reported through Chinese news outlets (China News, China Youth Daily, Global Times, China Consumer News) on the launch day; none has an independent audit attached. The "2.2 million enterprises using AI" figure is a platform self-report and bundles the entire AI feature set — document AI, low-code AI, smart office automation — with the newer assistant store, so it overstates how many organisations have actually deployed a custom bot. The "200+" count is the first batch only; DingTalk has not published a verified later total, so any growth claim would be speculation and is omitted here.
The assistant quality also varies wildly by builder. A role assistant for travel trivia is a different risk class from a legal or payroll assistant that touches money and compliance, and the public reporting does not break out success or error rates per category. The Hangzhou police example shows the model put to serious civic use, but it is one deployment, not a national pattern. Treat the store as a large, real experiment in putting models inside everyday work software — not as evidence that offices now run themselves.
What readers can do now
- If your organisation already runs on DingTalk, search "AI 助理市场" and switch on one role assistant for a single repetitive job — meeting notes, travel booking, a FAQ bot — before touching anything that moves money.
- Prefer cross-app assistants that connect to software you already use over pure chatbots; the value is in the action, not the conversation.
- Keep a human in the loop on anything legal, financial or customer-facing, and review the bot's output the way you would review a junior colleague's first draft.