Office Assistants Are Now Fighting Over Your History

Office assistants are starting to fight over your history

The most noteworthy move on the Chinese internet these past two days is not which model topped another leaderboard, but a feature that is easy to dismiss as a product detail: Doubao Work has launched one-click import, which can move the skills, memories, and past conversations in ChatGPT and Claude Code into your own account.

In the same week, Alibaba put an e-commerce edition of QwenWork into an invitation-only beta, aimed at sellers who work across multiple platforms, with the pitch that you can get a store up and running at low cost. A few days earlier, xAI's Grok Voice 3.0 took first place on Artificial Analysis's voice agent ranking, with response time squeezed under one second.

These three developments look unrelated, but they point to the same conclusion: for office agents, the deciding factor is shifting from how strong the model is to whether users are willing to switch over.

Why moving house is a signal

Anyone who has used AI office tools for a while knows that what you really cannot bear to delete is not a particular model but the things you have built up: a whole set of tuned prompts, dozens of notes about your projects and preferences, and the workflows that already run smoothly.

In other words, sunk cost is what creates stickiness. Switch to a smarter assistant, and the price is teaching it from scratch who you are, what you do, and where your boundaries lie.

One-click import neatly removes that barrier. It turns sunk cost, which used to keep you in place, into luggage you can carry out. This is not a neutral efficiency feature but a direct strike at a competitor's switching costs. The documentation says import; the subtext is that you no longer have to start over.

From being able to chat to being able to be trusted with the job

At the chatbot level there is not much mystery left. Ask the same question of ten models and the answers are roughly the same, while prices keep falling. What really separates them is whether a model can take a task from start to finish.

The requirements for an office agent are concrete: it has to understand the documents you throw at it, remember the preferences you mentioned last time, and connect the work across several apps. The three items in Doubao Work's import list, skills, memory, and past conversations, map exactly onto those three needs.

Skills determine what it can do, memory determines whether it knows who you are, and history determines whether it can pick up from last time. Miss any one of the three, and the user has to fill the gap themselves, and that filling-in is where churn begins.

Memory is both a moat and a liability

No one disputes that memory makes an assistant smoother the more you use it. But memory is a double-edged sword, and viewed from the other side it is a different story.

If memory can be exported with one click, then it protects no one. The work profile you painstakingly built can become a gift to a competitor at any moment. Conversely, if memory cannot be exported, users are locked tightly into one platform, unable even to switch to a cheaper option.

This contradiction has no elegant answer. For users, the reasonable expectation is that the data is mine and export is my right. For platforms, the burden is to prove that keeping you depends on being more useful, not on making it impossible for you to leave.

The more practical issue is compliance. Once an office assistant starts storing document snippets, calendars, contacts, and even meeting content for the long term, it becomes a database access point attached to your workflow. Who can see it, how long it is kept, and what happens after an employee leaves are questions that will eventually be asked.

Voice is becoming the new entry point

If moving house solves the first step, the entry point determines how many times a day you use it.

The point of Grok Voice 3.0's first place is not the score but the latency. The barrier in voice interaction has always been that awkward pause: you finish a sentence, wait while it thinks for a second, and start to wonder whether it even heard you. Only when response time is pushed under a second does the rhythm of conversation hold together.

When voice is fast enough, the entry point in office settings shifts from opening a box and typing to just saying a sentence. While driving, while your hands are busy, while sitting in a meeting room, being able to assign a task out loud changes the frequency of use by an entirely different order of magnitude.

China plays work, overseas plays frontier

Pull the camera back, and the two paths are already clear.

Several overseas players take the closed frontier model plus memory route, selling a more human-like assistant to individuals and enterprises at no small price, but what they sell is the ceiling of capability. This batch of domestic products takes a different path: focusing on specific job roles and scenarios, pushing costs down, and assembling a full toolchain.

For developers, there are indeed more choices, but choosing has also become harder. For the same task, a closed frontier model might nail it in one pass, but the per-unit price is high; a cheaper small model might take three rounds plus a verification layer, yet the total bill is lower. Which to pick depends on how much of your task must be right the first time and how much can be retried.

Final thoughts

It is easy for the narrative of this round of competition to drift into spec comparisons, but what really determines retention is a pile of unglamorous things: how smooth the migration is, how reliable the memory is, how natural the entry point is, and how predictable the bill is.

Whether Doubao Work's one-click move deserves applause depends on how it handles the memories it will itself accumulate. If it only moves users in from elsewhere and then builds a higher wall in place, the fight is still following the old script.

What might be genuinely new is this: a user does not have to start from scratch the first time they open it. It sounds small, but it means assistants have finally begun competing on the experience of use rather than on whose leaderboard screenshot looks better.