DingTalk Wangshu Plan Takes on Feishu: Which AI Office Tool Actually Delivers?

Two Months Inside the Great China Collaboration AI Showdown

Let me set the scene for you. It’s a Tuesday morning in March, I’m sitting in a glass-walled meeting room in Shanghai with my VP of Operations, our IT director, and a slightly suspicious procurement person who has a spreadsheet open and a laser pointer. We have sixty days to decide whether our 400-person company is going to roll out DingTalk or Feishu as our primary collaboration platform, and the kicker is that the AI features in both have been blowing up in the press, and neither of us knows which one will actually survive contact with our actual employees. That is the moment, friends, when a tech blog gets written. I spent the next eight weeks doing the kind of hands-on, messy, sometimes-furious evaluation that vendor whitepapers will never tell you about. This is what I found.

First, a bit of context. We are a mid-sized SaaS company selling B2B logistics software. We have engineers in Hangzhou, sales people in Shenzhen, customer success folks scattered across the country, and a remote-heavy product team. Our existing stack was a Frankenstein monster of WeChat Work groups, email, a half-broken Asana instance, and enough Notion pages to fill a small library. We needed to consolidate, and more importantly we needed to stop being the company where the new hire asks “where do I find the Q4 OKR document” and gets eleven different answers from eleven different people. Both DingTalk and Feishu are pitched as the all-in-one solution. The question was, with AI, which one actually delivered.

Let me start with DingTalk, because chronologically that’s where we started, and because it gave me a strong first impression that I had to slowly walk back over the next several weeks. DingTalk’s AI suite is built on top of Alibaba’s Tongyi Qianwen model family, which now includes the Qwen 2.5 and Qwen 3 series, and it shows up in the product as a little blue dot in the corner of basically every screen. Click it, and you get what DingTalk calls the AI Assistant, sometimes labeled as 钉钉AI助理. Out of the box, it can do the things you would expect: summarize a thread, draft a reply, generate a meeting agenda, pull data from your workspace, and answer questions like “what did we decide about the API rate limit last week.” The killer feature in our first demo was the meeting experience. DingTalk can join a video call as an actual participant, transcribe in real time, distinguish between speakers with reasonable accuracy, and produce a structured summary at the end with action items, decisions, and a follow-up email draft. I watched a thirty-minute all-hands become a four-paragraph brief in about eight seconds, and a Slack-style message thread that was starting to feel like a Choose Your Own Adventure novel collapse into a clean bullet list. I was sold. I was ready to fly the DingTalk flag.

Then the second week happened. We rolled DingTalk out to the customer success team first, because they have the highest meeting density and the most painful documentation burden. The good news is that the AI meeting features held up. The transcription accuracy in Mandarin was around 95 percent in quiet rooms and dropped to maybe 85 percent when our Shenzhen team called in from what sounded like the floor of a basketball game. Speaker diarization, which is the fancy term for figuring out who is talking when, worked about 80 percent of the time, and when it failed it was usually because two people had similar voices or because someone joined by phone instead of video. The action item extraction was hit or miss. In a clean meeting with crisp decisions, it pulled out action items beautifully, including who they were assigned to. In a chaotic brainstorming session where people were talking over each other and changing their minds, the AI confidently produced a list of action items that were about 60 percent real, 30 percent inferred, and 10 percent completely hallucinated. We learned the hard way that “AI extracted this from your meeting” is not the same as “this actually happened.” We started a practice of having a human reviewer glance over the AI summary before it went out to the client. That sounds like a small thing, but it adds up to a meaningful amount of time across fifty meetings a week.

The bigger surprise with DingTalk was how the AI performed in chat. The assistant inside a group chat, when you @ it, can answer questions using your company knowledge base, which is impressive. I typed “what is our refund policy for enterprise customers” into our sales channel and got back a correct, sourced answer pulled from a Notion page we had imported. I was thrilled. Then I asked a follow-up: “who owns the refund policy and when was it last updated?” The AI confidently told me that Zhang Wei from finance owned it, and that it was last updated in February 2024. Zhang Wei does work here, but he does not own the refund policy, and the policy was last updated in November 2025. The model had hallucinated two facts, packaged them in a confident sentence, and was ready to send them to a sales rep about to talk to a customer. I caught it because I happened to know the answer. A new hire would not have. This is the kind of failure mode that is invisible in a demo and catastrophic in production. I am not blaming DingTalk specifically here, because every LLM does this, but the product surface made it very easy to trust the AI and very hard to catch the errors. There is no “show me the source” by default; you have to dig. Compare that to a simple search, which at least lets you see the document and judge its freshness yourself.

The other DingTalk feature that got a lot of internal hype was the AI-powered low-code platform, sometimes called Dingtalk Yida, which lets you build internal apps and forms using natural language. I tried to build a simple travel request form. I typed “create a form for business travel requests with fields for destination, dates, purpose, estimated cost, and manager approval.” It built the form. It was honestly impressive. Then I tried to add a conditional rule: “if estimated cost is over 5000 yuan, route to the director for approval, otherwise just to the manager.” It struggled. It produced a form that had the conditional logic, but the routing was wrong, and when I asked it to fix it, it broke something else. I spent about forty minutes trying to coax it into the right configuration and eventually gave up and built it manually. My take is that the AI form builder is great for the 70 percent of cases that are simple, and frustrating for the 30 percent that involve any business logic. The promise is real, the maturity is not quite there.

After three weeks we had a clear picture of DingTalk AI. It is best in class for meeting summarization, strong for general chat assistance, and promising but uneven for app building. It is tightly integrated with Alibaba Cloud, which is great if you are already on Alibaba, and irrelevant if you are not. The mobile app is dense and powerful, occasionally to the point of feeling like the cockpit of a 747, and our younger employees found it more cluttered than the alternative we were about to test.

Now let’s talk about Feishu, which I kept wanting to call Lark because that is what the international version is called, and because the people who picked the English name in our company kept saying “no, it’s Feishu, we are in China.” Fair enough. Feishu’s AI, branded as 飞书智能伙伴 or Feishu Smart Companion, is built on ByteDance’s Doubao model family, and the product philosophy feels different from the moment you log in. The interface is calmer, the AI is less aggressive about offering to do things, and the whole experience is built around documents rather than messages. That sounds like a small thing, but it is a huge philosophical difference, and after a few weeks I started to think it was the actual reason to pick one over the other.

Feishu’s meeting AI is called Miaoli, which translates roughly to “wonderful record.” It does the same thing as DingTalk’s meeting AI: transcribes, summarizes, extracts action items, and produces a shareable document. The accuracy was similar, maybe a hair better on technical jargon, possibly because Doubao was trained on a slightly different mix of data. But the killer feature of Miaoli is not the summary, it is the search. Every meeting transcript in Feishu is fully searchable, with timestamps, and the AI can answer questions like “what did the CEO say about the Q3 hiring plan in the all-hands on March 4th” and give you a specific answer with a link to the exact moment in the video. We started using this immediately and it became a small miracle. The number of times someone asked a question in a meeting, someone else said “I think she mentioned that in the March all-hands,” and then two hours later the AI would surface the exact sentence was, over the course of a month, probably fifteen. That is fifteen meetings that did not have to be repeated, fifteen pieces of institutional knowledge that did not get lost. I cannot overstate how much morale this generated. People felt like the company finally had a memory.

The Feishu AI in chat is more conservative than DingTalk’s, and at first I found that annoying. I would @ the assistant in a group chat and it would often say “I’m not sure I have enough information” instead of taking a guess. After the hallucination incident with DingTalk, I started to appreciate the restraint. When Feishu did answer, it was right more often, and crucially, it almost always included a citation. You could click through to the source document and verify. That sounds like a small UX detail. It is not. It is the difference between an AI that is a productivity tool and an AI that is a liability.

The other place Feishu shines is documents. Feishu Docs, or 飞书文档, is the best document collaboration tool I have ever used, full stop, and I have been using tools like this since Google Docs was an invite-only beta. The AI integration is tight. You can highlight a paragraph and ask the AI to rewrite it, summarize it, expand it, or translate it. You can ask it to generate a first draft of a document from a prompt, and the draft will be saved as a real document you can edit with your team. The most underrated feature is the AI-powered table, called Bitable, which is essentially a smarter Airtable. You can have the AI generate a table from a description, populate it with data from your other docs, or summarize the contents in natural language. I built a customer feedback tracker in about twenty minutes that would have taken me two days in a spreadsheet. I am not exaggerating. I described what I wanted, the AI built the structure, I imported a CSV, and the AI summarized the trends. This is the kind of AI feature that does not show up in a vendor demo because it is not flashy. It is just quietly useful every single day.

The area where Feishu is weaker is its app-building platform. Aily, which is Feishu’s answer to Dingtalk Yida, is more polished for pre-built templates but less capable when you start asking for custom business logic. We did not test it as hard, because by the time we got to Aily we were already leaning toward Feishu for the other reasons, but the feeling I got from a few hours of poking around is that Aily is optimized for sales and HR workflows, and gets flakier when you try to build something custom for engineering or operations. If your company lives in standard HR and sales processes, you will be fine. If you have weird internal tools, you will probably end up calling a developer.

So let me get to the comparison proper, because that is what you came for, and I owe you a real verdict after all this setup. The two platforms are converging in their headline features, but they have very different personalities. DingTalk AI is more aggressive, more chatty, more willing to take initiative, and that is both its strength and its weakness. It is the coworker who is always volunteering to do things, and sometimes does them brilliantly, and sometimes does them wrong and confidently. Feishu AI is quieter, more document-oriented, more citation-friendly, and the experience is built around the idea that you should be able to verify what the AI said. The contrast is almost like the difference between a fast-talking consultant and a careful research assistant. Both are valuable, but you want different ones in different situations.

For meeting notes and general workplace chat, DingTalk is slightly better in raw features, and Feishu is slightly better in trustworthiness. For document collaboration, Feishu is in a different league, and the AI integration feels like it was designed from the ground up rather than bolted on. For internal app building, DingTalk is more flexible but flakier; Feishu is more polished but more constrained. For mobile experience, DingTalk is more feature-dense, Feishu is cleaner. For integrations with the broader Alibaba versus ByteDance ecosystems, the choice is obvious based on where you already are.

The single most important factor for us turned out to be search and memory. We picked Feishu, and the reason is that Feishu made our company feel like it had a memory. The combination of fully searchable meeting transcripts, AI-powered document search, and citation-friendly chat answers meant that for the first time in the history of our company, someone could ask “what did we decide about X” and get a real answer instead of getting forwarded to three people who would each have a different version of the story. That is not an AI feature in the sense of “look, the robot can write a poem.” It is a knowledge management feature, and it is the thing that actually moves the needle for a company that is drowning in its own information. The second most important factor was document collaboration, where Feishu is so far ahead that there is not really a comparison. The third most important factor was trust. The Feishu AI was less likely to hallucinate, more likely to cite its sources, and less likely to confidently tell a sales rep something wrong. For a company where bad information in a customer conversation is a five-figure mistake, that matters.

Now, the recommendation, with all the caveats that come with picking software for someone else’s company. If you are a heavy meeting culture, with lots of cross-team communication and a lot of “wait, what did we decide” moments, both platforms will serve you well on the meeting front, and you should pick based on the other features. If you are a document-heavy organization, Feishu is the obvious choice, and the AI features are a bonus on top of an already superior collaboration product. If you are a small company that mostly needs chat and basic meetings, the differences will not matter much, and you might as well pick the one your team finds more pleasant to use, which in my informal survey was Feishu by a wide margin, especially among employees under thirty. If you are deeply integrated with Alibaba Cloud, or you are using DingTalk’s specialized hardware or logistics features, the choice is already made for you. If you are a global company with significant English-language operations, Feishu’s international version Lark is significantly more mature, and that should be a factor.

For us, after sixty days, the answer was Feishu, and we are now three months into the rollout. The number of repeated meetings is down, the number of “where do I find the policy on X” questions is down, and the new hire onboarding time has dropped by about a third because the AI can now answer most of the questions a new hire would have asked a human. The platform is not perfect. The AI still hallucinates occasionally, the mobile app still feels slightly less powerful than DingTalk’s, and there are a few edge cases in Aily that have made our ops team grumble. But the net effect has been a company that is faster, calmer, and more confident in its own information. That is the kind of AI win that actually shows up on a P&L, even if it never makes a flashy press release.

So that is the story. DingTalk and Feishu are both genuinely good, and the fact that you are choosing between them is already a win compared to where most companies are. Pick based on your culture, your existing ecosystem, and most importantly, your tolerance for the AI confidently making things up. If you want the AI to be a research assistant that shows its work, pick Feishu. If you want the AI to be a fast-talking intern who will do anything you ask and sometimes gets it wrong, pick DingTalk. Either way, your employees will be more productive than they were on whatever Frankenstein stack you have today. And that, my friends, is the only metric that actually matters.