As LobsterAI Starts 'Acting,' Office Assistant Valuation Logic Shifts

The most valuable insight from this article is that LobsterAI has truly bridged the gap from "generation" to "execution," but the underlying "operational controllability" and "data closed-loop" are what actually determine its valuation ceiling.

At the WAIC 2026 venue, NetEase Youdao's LobsterAI had people lining up to try it out. That itself isn't unusual. AI booths have long lines every year, but most people just walk away after trying it, resulting in extremely low conversion rates. What really made me stop was a user feedback quote: "It's not about having AI write a paragraph for me, but having it actually get the job done." This hits the core pain point of current office assistants—many products stay at "can answer," but fail at "can execute."

From an investment perspective, LobsterAI's differentiation lies in building an abstraction capability at the "operation layer." Traditional AI office assistants (like the Copilot model) are essentially "generators"; they output text, code, or tables, but users still need to manually copy-paste, open software, and set parameters. LobsterAI attempts to directly take over the operation interface of desktop software, evolving from "helping you think" to "helping you do."

User Command: Organize last week's sales data into a PPT, including a bar chart of growth rates
LobsterAI Execution Flow:
1. Identify and open the Excel file
2. Locate the "Last Week Sales Data" worksheet
3. Calculate the growth rate column (using formulas)
4. Generate a bar chart and adjust formatting
5. Open PowerPoint and create a new slide
6. Embed the chart and adapt to the template
7. Save and send email notification

This flow looks simple, but the technical barrier is extremely high. It requires three core capabilities: first, deep understanding of the DOM structure of desktop software; second, cross-application action orchestration; third, robustness in exception handling (e.g., software crashes, file format errors). Currently, similar products on the market (like AutoGPT's desktop version) mostly stay at coarse-grained operations like "open browser," whereas LobsterAI can achieve fine-grained control like "select cell - set formula - adjust column width."

[!note] From a valuation logic perspective, LobsterAI's moat lies in the startup speed of its "data flywheel." Every time a user executes a task, operational trajectory data is generated, which can feed back into the model to improve execution accuracy. However, the issue is that this data loop needs enough paying users to work, and data privacy compliance costs are very high. NetEase Youdao has a natural user base in educational scenarios, but in enterprise office scenarios, it faces competition from platform players like DingTalk and Feishu.

I noticed that LobsterAI's pricing strategy might adopt a "pay-per-task" or "subscription + task pack" model. This is more flexible than traditional SaaS per-account billing, but it also tests user retention more. If users only execute 5 tasks per month, ARPU will struggle to cover model inference costs. A true business model must convert users from "occasional trial" to "daily dependency."

• Core Moat: Operational controllability (deep embedding into desktop software)
• Competitive Risk: Platform players (like Microsoft, Google) with OS-level control
• Valuation Key: Data flywheel startup speed, not user count

From a technical roadmap perspective, LobsterAI chose a hybrid architecture of "edge-side models + cloud API." The edge side handles real-time operations and private data processing, while the cloud handles complex task orchestration and model updates. This architecture keeps inference costs relatively controllable, but the parameter scale of the edge-side model (likely 7B-13B) determines whether it can handle complex scenarios.

[!abstract] Investment Judgment: LobsterAI has a first-mover advantage in the "Office Agent" track, but its ceiling depends on breaking through the "narrow scenario" dilemma. If it can only do "Excel organization" and "PPT generation," its valuation won't exceed $1 billion. But if it can expand to "ERP system operations," "CRM data entry," and other core enterprise processes, it could become the next "RPA 3.0," with valuation logic comparable to UiPath's early stage.

So the question arises: How much are users willing to pay for "execution"? Over 500k? Or over 5 million?

Original Link: https://www.leiphone.com/category/industrynews/59N2nNCnS8JoskbA.html