Doubao's Conversational AI Team Cut by Nearly Half, Accelerating Commercial Pivot
nashnova research
ByteDance is halving Doubao's core chat team — about 50 people down to roughly 25 — and redeploying staff toward monetization and Feishu; DAU tops 200 million but revenue lags far behind, signaling a strategic shift from conversation to enterprise services and e-commerce.
Which team was cut, and by how much?
The General Session team — responsible for chat product strategy and evaluation — had about 50 people. Roughly half are being let go or reassigned.
The Product Posttrain-Chat team is also shrinking; some staff have already moved to other units.
This means → ByteDance is narrowing both core chat teams at once. The signal is clear: conversation is no longer priority number one.
200 million DAU — so why cut?
Doubao's DAU tops 200 million, the highest of any AI app in China — yet the chat product still has no proven revenue model.
Costs extend well beyond text. Image, voice, and video features are all free; each user gets 5 free Seedance video-generation credits per day, settled separately with Volcengine.
Some comic creators even register accounts in bulk — building "account pools" — to harvest the free quota, pushing costs higher still.
In plain terms = more users means bigger losses. The free-tier strategy built the DAU; it did not build the revenue.
Where will the money come from?
Doubao's revenue operation now runs on two tracks: "Productivity" (Doubao Pro, Enterprise edition, Doubao-inside-Feishu), led by Tong Yao; and lifestyle services plus e-commerce, led by Xia Ning.
A third line — mini-program integrations like ride-hailing inside Doubao, previously run by Yang Kang — has been folded into the two main tracks.
This reflects ByteDance borrowing from Anthropic's playbook: pivot from consumer chat to enterprise services, layered on top of its own strengths in e-commerce and local services.
What was the "sycophantic answers" problem — is it fixed?
In April, users complained that Doubao's replies were "dumb and sycophantic." The root cause: Doubao optimized for retention — every new release had to hold or raise it.
A model shipped before Chinese New Year over-weighted user-preference signals. The model became more agreeable but less accurate; post-holiday user growth made rolling back nearly impossible.
In April, Flow division head Zhu Jun decided to accept a short-term retention dip to fix the experience. The team spent over a month retuning reward-model weights, cleaning data, and retraining. Retention dipped by less than 1%.
In plain terms = to avoid losing users, the model was "spoiled." The fix cost a small, temporary retention drop — but leaving it unfixed would have damaged the brand long-term.
What changed at the model layer?
Around June, the post-training team merged multiple small models into a single model, eliminating routing errors caused by inaccurate query classification.
This means → before the merge, each user query was classified first, then dispatched to a different small model. A misclassification tanked the answer quality. One unified model removes that failure point.
The chat team shrank — what comes next?
A ByteDance insider's view: the company already holds a 200-million-DAU entry point. Once a new direction is validated, Doubao can leverage that user base and its accumulated product and algorithm capabilities to move fast.
But right now, Doubao needs leaner operations and higher revenue.
This reflects a strategic inflection point: the chat-team contraction is not "giving up on chat" — it is shifting from burning cash for DAU to turning DAU into cash.
市场有风险，内容仅供研究参考，不构成投资建议。