Goldman Sachs' judgment: Claude Cowork becomes the first "trusted proxy workflow case," and OpenClaw showcases the future direction of human-computer interaction.

AI applications are entering a critical next phase—full integration from developer tools to knowledge workers. Goldman Sachs’ latest research report outlines a clear path for the large-scale deployment of general intelligence tools through two flagship products—Claude Cowork and OpenClaw.

On March 7, according to ZF Trading Desk, Goldman Sachs stated in its latest report that applications of AI as the next stage of the general intelligence layer have already emerged, with Anthropic’s Claude Cowork and the open-source agent OpenClaw being two key data points defining this trend.

The report notes that Claude Cowork, as the first trusted “agent workflow” for knowledge workers, is reshaping value distribution in the SaaS industry—it won’t directly replace existing record systems (like CRM or ERP), but will shift market share toward SaaS providers that effectively integrate AI capabilities.

Meanwhile, the viral spread of OpenClaw not only demonstrates a new form of human-machine interaction but also its unique hosting economics are breaking the monopoly of hyperscalers, channeling massive new inference computing demands directly to alternative edge computing and VPS providers like DigitalOcean and Cloudflare.

Claude Cowork: The First Trusted Case of Knowledge Worker Agent Workflow

On January 12, Anthropic officially launched Claude Cowork—a proxy tool embedded in the Claude desktop app, designed for non-developers, capable of automating file management and task workflows, including multi-step operations such as file organization, document processing, and workflow management.

On January 30, Anthropic further released a series of Cowork plugins targeting “business functions,” covering sales and marketing, finance, legal, customer support, data analysis, productivity and enterprise search, as well as product management scenarios.

Goldman Sachs believes that the core value of Cowork lies in filling the critical gap between developer-only proxy tools and the daily use of chatbots by knowledge workers. Software engineers have seen productivity gains from terminal proxy loops for over a year, but AI interactions for knowledge workers have long been limited to chat interfaces. Cowork extends this capability to non-technical users through integrations with Excel and standard business tools via application interfaces.

Specific deployment scenarios include: accountants classifying credit card transactions, HR teams compiling tax reports and tracking PTO, CFOs accelerating project-specific DCF analyses. Among digital proxy users, marketing is the most prominent—clients are integrating Klaviyo and automating entire reporting workflows.

Cowork’s core competitive advantage is its role as an intelligent layer across SaaS stacks: it is not dependent on any single SaaS product, can switch freely between tools, and performs reasoning over data spanning multiple systems and “glue layers.” Industry research by Goldman Sachs shows that Anthropic is clearly leading in both tool capabilities and underlying model intelligence.

Cowork will not replace system records but will reshape SaaS value distribution

Goldman Sachs explicitly states that tools like Cowork will not replace core system records such as CRM or ERP, and companies have no intention of dismantling their existing SaaS stacks. However, as AI becomes a key differentiator, wallet share will concentrate among SaaS vendors that provide the most effective AI workflows.

In practice, clients are replacing lightweight “glue” software—tools mainly used for moving data between systems—with Cowork, rather than core platforms. Conversations with Goldman Sachs partners indicate that companies still prefer to pay third-party software vendors for security, governance, and operational control.

The partnership between Intuit and Anthropic exemplifies this logic. On February 24, Intuit announced a collaboration with Anthropic: Intuit will access Anthropic models and custom AI agents via the Claude Agent SDK, while Anthropic will promote Intuit products through MCP integration on Cowork, Claude for Enterprise, and Claude.ai platforms.

Goldman Sachs believes this approach is similar to Intuit’s previous $100 million annual deal with OpenAI: the foundational models are monetized through capabilities and reach, without building custom workflows; meanwhile, Intuit maintains control over data and preserves its unit economics.

The potential risk is that adding a layer between clients and Intuit could, over the long term, abstract value; but the potential benefit is better customer acquisition, akin to the collaboration models of Web 2.0 and search engines.

Notably, on March 5, OpenAI launched GPT-5.4—its first general model with native computer operation capabilities. Similar to Claude Cowork, it can operate computers and execute proxy workflows across applications. OpenAI emphasizes that this model reduces token consumption by about 47% on certain tasks while maintaining the same accuracy.

OpenClaw: Breaking Context Limits and Showcasing Future Human-Machine Interaction and Computing Demand

Goldman Sachs believes OpenClaw, by granting AI agents full control over computers and unlimited memory, has achieved viral growth on the consumer side; its unique local and low-cost VPS hosting model is a significant boon for edge computing providers.

OpenClaw (formerly Moltbot and Clawdbot) was first released in November 2025 and achieved “viral” spread in late January.

This is an open-source, self-hosted, 24/7 AI agent that runs on the user’s local machine, capable of receiving commands via personal messaging platforms like iMessage, WhatsApp, Google Chat, and executing any task a human can—opening applications, browsing web pages, completing transactions.

OpenClaw addresses a key technical bottleneck limiting most AI agents—the context window problem. Its solution is a file-based memory system: all agent activities are stored locally in Markdown files, with the agent loading that day’s memory and the previous day’s context at each session start. Users can also establish long-term memories through commands (e.g., “Always prioritize OpenTable when booking restaurants”), creating persistent preference databases.

Matthew Prince, co-founder and CEO of Cloudflare, notes that OpenClaw’s development trajectory over the next three years could mirror ChatGPT’s past three years: currently catalyzing numerous experiments, but similar to early ChatGPT, it may take years to achieve sustained daily scale.

OpenClaw’s computing economics: Opportunities for edge and VPS providers

OpenClaw’s hosting economics offer important insights for investors. Due to the high cost fluctuations of hyperscalers, most users run OpenClaw on home hardware (Mac Mini, old laptops) or low-cost VPS providers (like Cloudflare and DigitalOcean). Some users have spent thousands of dollars in token costs within just a few days of initial use.

Many are adopting a layered model architecture: front-end models orchestrate, smaller models perform sub-tasks, reducing per-query costs.

DigitalOcean’s data is particularly illustrative: within days of OpenClaw’s release, nearly 30,000 native one-click OpenClaw Droplets were deployed on DigitalOcean, with thousands more activated. AI revenue from DigitalOcean’s customers continues to grow—by December 2025, AI customer ARR reached $120 million, with 70% from non-metallic services, and inference services grew at 254% year-over-year.

Cloudflare has not yet seen significant impact on logo growth or web traffic from OpenClaw, but Goldman Sachs believes this further confirms Cloudflare’s edge network and Workers software’s high cost-performance advantage.

This insightful content is from ZF Trading Desk.

For more detailed analysis, real-time insights, and frontline research, please join【**ZF Trading Desk▪Annual Membership**】

![](https://img-cdn.gateio.im/social/moments-7ad4881ca3-36f0f348ca-8b7abd-ceda62)

Risk Disclaimer and Terms of Use

Market risks are present; investments should be cautious. This article does not constitute personal investment advice and does not consider individual users’ specific investment goals, financial situations, or needs. Users should consider whether any opinions, viewpoints, or conclusions herein are suitable for their particular circumstances. Investment is at your own risk.