ByteDance released the consumer version of its Doubao phone assistant on September 14, with the first device carrying it being the Nubia NaviX Ultra, co-developed by ZTE and ByteDance. The phone launches and goes on sale at 2 p.m. on September 16, with preorders already open.
The same day, ByteDance also released a protocol called SAEP — the Screen Automation Execution Protocol — with a 30-day public comment period. According to the company, third-party apps can independently declare the boundaries of automation within their own app, explicitly allowing or refusing the AI assistant's screen-automation actions. For apps that explicitly opt out, the assistant will not carry out automated actions.
What This Version Adds
Near the top of the feature list is how the assistant is woken: voice activation, plus a dedicated AI key with fingerprint authentication. By the company's description, the key doubles as both an entry point and an authorization checkpoint — only after the fingerprint passes does the assistant proceed.
Next is on-screen Q&A — users can ask about whatever is currently on the screen. Local data search now covers photos, text messages, and notes. Voice recording has been upgraded and now connects to Feishu Miaoji (Lark's transcription tool). The part that executes tasks across apps is called "Operate Phone," currently tagged Beta.
In a demo scenario given by the company, a user points a camera at furniture and says: "Based on this decor style, pick me a cabinet under 1.2 meters and under 1,000 yuan." The assistant then searches an e-commerce platform for matching items. That single sentence packs in three constraints — style, size, and price — spanning vision, search, and price comparison, and completing it requires the assistant to actually navigate inside the shopping app.
Data processing uses a hybrid on-device/cloud approach. Listed compliance certifications include ISO 27001, ISO 27701, ISO/IEC 42001, and China's MLPS Level 3 (等保三级). Published hardware specs put the phone at 7.62mm thick with a 7,100mAh battery.
How far the assistant actually works, the company itself qualifies: it depends on the device model, system version, app adaptation, and user authorization.
The App Being Operated Gets a Switch, for the First Time
Screen automation has followed roughly the same playbook for more than two years: the assistant obtains accessibility or screenshot permissions, recognizes on-screen elements, and simulates taps and swipes. Anthropic's computer use, OpenAI's browser operation features, and the agent solutions from several domestic phone makers all sit within this same basic framework.
One thing this approach has never resolved: the app being tapped has no say in the matter. It cannot tell whether the input is coming from a human or an assistant, and it has no formal channel to say "don't touch this page." In the past, the workaround was either the app building its own anti-automation detection, or the phone maker striking private deals with major apps.
SAEP is aimed at exactly that layer, giving the operated app a way to declare its boundaries. For apps built around transactions, payments, or content distribution, this addresses a real concern: automated traffic mixed in with genuine user traffic distorts risk control, billing, and recommendation models.
But how much the protocol actually accomplishes depends on implementation. During the 30-day comment period, ByteDance has not specified who is authorized to make a declaration, whether declarations apply at the page level or only the app level, or how the assistant is supposed to back off once an app refuses. Whether the declaration list itself will be made public is also unclear. The protocol is drafted and maintained by ByteDance alone, not by an industry standards body, so whether third parties adopt it ultimately comes down to each company's own judgment.
A Few Practical Questions for Users in China
First, device exclusivity. This version of the assistant ships only with the Nubia NaviX Ultra for now, and ByteDance hasn't said when — or how fully — it will reach other phone models. The recurring pattern with domestic AI assistants over the past couple of years has been to debut a feature on in-house hardware first and expand later; this rollout looks set to follow the same path.
Second, permissions. Screen automation requires a higher tier of access than an ordinary app, and combined with local search across photos, messages, and notes, the scope of authorization is substantial. The fingerprint-gated AI key adds a checkpoint at the entry point, but once authorization is granted, it's hard for users to verify item by item what the assistant can reach in the background.
Third, reliability. "Operate Phone" is still in Beta, and the places where this kind of feature tends to break down in China's app ecosystem are CAPTCHAs, payment confirmations, and login states — precisely the three steps least suited to letting an assistant act on a user's behalf.
Sources: Doubao official announcement, IT Home (IT之家), Nubia's official site, CocoLoop; verified against the SAEP comment-period timeline, device launch schedule, and stated feature limitations.