Doubao Work vs Feishu? Verify Before You Commit

Last verified: August 27, 2026

Doubao Work (豆包工作) launched on August 25, but the evidence available two days later is unusually one-sided. ChooseAI explicitly noted that its launch-day capability descriptions all came from official sources because no third-party testing was yet available.[1] The product may be able to create editable office deliverables and operate across workplace tools. What has not been established is how reliably it can do so with a buyer’s own tasks, permissions, and account limits.

There is one limited warning from outside the launch materials. A login-walled Zhihu article carries the headline “豆包专业版实测：4个办公任务，3个没干成,” reporting that three of four office tasks failed in a hands-on run of Doubao Pro.[2] Because only the headline is accessible, it does not reveal the tasks, failure criteria, environment, or recovery attempts. It is a documented headline-level warning about the predecessor—not a detailed test record and not direct testing of the newly launched Doubao Work.

The predecessor connection is narrow but relevant. Doubao Pro was listed on June 24 with an 办公任务 office-task mode built on Doubao 2.1-series models, including browser use, Skills, and scheduled tasks.[3] That establishes lineage around agentic office work. It does not establish that the two products have identical execution behavior.

The pre-commitment pass

A polished demonstration answers whether a prepared workflow can succeed. A buyer needs to know whether representative work completes correctly, whether the resulting files survive review, what the account actually includes, and where the agent can reach. Those questions can be resolved through a contained trial without giving the product a production role.

First, make the work prove itself

Start with work that resembles what the intended user already does, while keeping the data disposable or safely redacted. Do not substitute a vendor template simply because it displays the product at its best. Before running anything, record what completion means: the required deliverable, the facts or source material that must be preserved, the parts that must remain editable, and the errors that would make the output unusable.

Then inspect the deliverable rather than the chat surrounding it. An agent can describe progress convincingly while leaving a spreadsheet structurally wrong, a presentation difficult to edit, or a plan dependent on unsupported assumptions. The evaluator should open the resulting files in the normal work environment, check whether their structure remains usable, and identify every intervention required to finish the task.

• Record whether the requested task actually reached its defined end state, rather than accepting a partial artifact plus an explanation.
• Review factual content, calculations, file structure, formatting, and omitted requirements separately. A visually polished result can still fail the underlying assignment.
• Track manual repair and recovery work, including access corrections and reruns. This is labor transferred to the reviewer, not automation.
• Repeat important work enough to see whether success depends on a lucky run. A single successful demonstration establishes possibility, not dependable execution.
• Preserve the prompt, inputs, output, intervention points, and final disposition so that another reviewer can understand the result.

Keep the evidence categories separate while judging these runs. Official pages can document claimed features. The Zhihu headline supplies a warning signal about the predecessor, but not enough detail to diagnose why those tasks failed. The separate capabilities analysis discusses an APPSO hands-on result involving editable plan, spreadsheet, and presentation output. That is evidence that particular outputs were produced in that reported run; it is not a reliability rate for every user or workflow.

Efficiency claims deserve the same treatment. Percentages generated by Lark’s marketing calculator describe a vendor’s modeled benefit, not an independently observed outcome for the evaluating team. The trial record should measure what happens locally: what finished, what required correction, and how much supervision remained.

Watch what the agent can reach—and when it gives control back

Execution quality is only half of an office agent evaluation. The agent operates through accounts, documents, and organizational boundaries, so its permission behavior must be visible during the trial.

Published product information says Doubao Work inherits the user’s Feishu permission scope, separates personal and enterprise data, and requires the user to take over for logins, captchas, and sensitive operations. The same source advises reviewing outputs because the AI can misrecognize information or perform an incorrect operation.[4] These are vendor-described safeguards and cautions. They do not by themselves prove that access will behave correctly in a particular tenant.

Inherited scope must be inspected, not assumed

If Doubao Work inherits Feishu access, the relevant question is what the test account can already see. Review that account’s group memberships, shared spaces, connected tools, and inherited document access before starting the agent. A trial account with broad legacy permissions can expose far more material than the evaluator intended, even if the product follows the account’s formal scope exactly.

Observe the resources the agent searches, reads, creates, and modifies. The record should distinguish authorized access from appropriate access: permission to open a document does not necessarily mean that document belongs in the current task’s context. If the evaluator cannot reconstruct where the agent went or why a source appeared in an output, the trial has not established enough visibility for workflow adoption.

Verify the personal–enterprise boundary in the actual account setup

A stated separation between personal and enterprise data is important, but the evaluator still needs to identify which identity is active, where generated artifacts are saved, what context is available to the agent, and whether switching accounts or workspaces changes that context clearly. The objective is not to defeat the boundary. It is to make sure an ordinary user can see and understand it before submitting real material.

Takeover prompts are control points

When a login, captcha, or sensitive operation requires takeover, note what information the user receives before acting. Can the reviewer see the destination, intended action, and likely consequence? Can the action be declined without losing unrelated work? Does control return to the agent predictably afterward? A prompt that merely interrupts execution is weaker than one that enables an informed decision.

The vendor’s warning about misrecognition and mis-operation makes output review part of the operating requirement, not an optional precaution.[4] Any adoption plan therefore needs a named reviewer, a clear approval boundary, and a recovery route for incorrect actions. Teams unable to provide that supervision should not place the agent in a workflow where errors can alter important records or trigger external consequences.

Confirm what the payment actually buys

Pricing is not stable enough to copy from a secondary listing and proceed. AIHub associates ¥68, ¥200, and ¥500 tiers with Doubao Work,[4] while this site’s earlier verification associated those amounts with Doubao Pro and had not independently confirmed standalone Doubao Work pricing. Until the seller or checkout flow resolves that attribution, those figures should not be presented as verified Doubao Work plans.

Before paying, capture the product name shown at checkout, the included quota, what consumes it, the reset period, overage behavior, renewal terms, and whether agent actions use a separate allowance. A plan name or monthly amount alone cannot establish practical value when the unit of use is unclear.

The detailed purchasing protocol is covered in Is Doubao Work Worth the Price? Individual users evaluating note-taking value should also use Is Doubao Work Membership Worth It for Note-Takers? Neither decision should rely on a tier table whose product attribution remains disputed.

The product boundary is still moving

Baidu Baike dates the merger of the Feishu and Doubao teams to July 30, 2026.[5] ChooseAI also reports the subsequent renaming of aily to “豆包工作伙伴” on August 13 and the folding in of TRAE and Coze office capabilities on August 24.[1] Its page prints 2024 for those latter organizational dates, however, which conflicts with the surrounding 2026 launch sequence. Without another source confirming their year, they should be treated as a reported sequence with a date-label problem—not silently rewritten as fully verified 2026 dates.

The independently anchored July 30 merger, the reported rename, and the reported integration still signal a rapidly changing product surface. That does not prove instability or explain the organizations’ motives. It does mean evaluators should record the product name, version or entry point, account type, and visible limits when testing. A result tied to one interface or organizational setup may need rechecking after a material product change.

A conditional verdict

As of August 27, the available evidence does not justify a paid or workflow-level commitment to Doubao Work without a verification pass. The absence of independent launch testing leaves reliability unknown. The predecessor headline supplies a reason to test carefully, but not a basis for declaring that the new product will fail. The permission and data protections remain published product claims until they are observed in the buyer’s own environment.

A low-stakes trial is enough to move forward if representative work survives review, the purchased quota is independently confirmed, and access behavior stays inspectable and controlled. Failure at any of those gates is a reason to postpone commitment rather than make the eventual operator repair the workflow after adoption.

Doubao Work is not for you if no one can supervise its outputs, audit the access inherited by its account, or take responsibility for recovery when an operation is wrong. That judgment can change if stronger testing and clearer commercial terms emerge; it should not be relaxed merely because the agent produces an impressive first artifact.

This verification answers whether Doubao Work is established enough to trust. It does not decide whether an agent layer or a collaboration-platform layer better fits the organization. For that separate choice, use Doubao Work vs Feishu? Ask Which AI Layer You Need.

References

1. 豆包工作正式发布：字节把 AI 办公收拢进一个 Agent — ChooseAI
2. 豆包专业版实测：4个办公任务，3个没干成 — Zhihu
3. Doubao Pro — Microsoft Store
4. 豆包工作 - 字节跳动推出的 AI 办公智能体平台 — AIHub
5. 飞书 — Baidu Baike