Graphic | Sister Tang

The AI Agent craze sparked by OpenClaw at the beginning of the year has now been ongoing for half a year. Back then, the Chinese internet was almost abuzz with talk of Agents. Major companies launched OpenClaw-like products, and startups flocked to desktop assistants, as if everyone could grab a ticket to this new entry point.

Looking back six months later, Tencent's WorkBuddy has truly captured this demand and made a name for itself in personal office entry points.

What makes WorkBuddy special is that it didn't rely on a globally leading proprietary model to break through. While Codex and Claude Code had top-tier foundational models paving the way, WorkBuddy chose to integrate multiple models, Skills, local tools, and the Tencent ecosystem at the product level. For those without programming backgrounds, it's a product worth trying out.

Even though I'm not a WorkBuddy user, I've felt its impact firsthand. When I first used Claude's desktop version earlier this year, I never touched Claude Code because I didn't think I needed programming. But after actually using it, I realized it handles files, tools, and multi-step tasks from my existing work, proving useful for both investment research and daily office tasks.

WorkBuddy's rise on the personal side at least confirms one thing: besides competing on models, delivering a great product is also a viable path. Almost simultaneously, Alibaba and ByteDance began consolidating their office Agent resources. On August 3, Alibaba integrated three Agents into QianWen Office and launched a public beta; on August 6, ByteDance emphasized AI productivity and enterprise business at its mid-year all-hands meeting, having already reorganized teams related to Doubao, Feishu, and Volcano Engine.

But the buzz on the personal side is still far from a sustainable, cash-flow-generating business. What the three companies showcase most visibly now are features, free quotas, and social media presence. What's hardest to see are payment scales, usage depth, renewal rates, and unit costs.

Moreover, office Agents face a natural disconnect: employees use them, but the work results belong to the company. For a single report, individuals might occasionally buy some Tokens; but when Agents start running long-term, accessing internal data, and acting on behalf of employees, costs, permissions, and responsibilities should no longer be left to employees. To truly scale, they must eventually enter enterprise budgets.

01  Usability Is Just the Beginning

The reason office Agents now seem like a personal competition isn't complicated. Employees are the first to discover new tools and are easily swayed by Skills, multi-model support, and free quotas. Clicking a Skill to see it write a report or create a spreadsheet delivers instant value.

Such demonstrations are crucial—they transform an unfamiliar technology from incomprehensible to hands-on, making individuals proficient quickly. But personal proficiency only proves the product has crossed the usability threshold; it doesn't mean the entire workflow is ready for enterprise production, let alone that enterprises will pay for it.

What individuals feel most acutely are immediate, visible results. Having an Agent write reports, organize spreadsheets, or handle code is something today's foundational models can already accomplish to a significant extent. But what enterprises pay for isn't these isolated results—it's whether the Agent can handle a specific job end-to-end.

For the same customer, order, or profit, even within the same company, different departments may use different metrics, with knowledge scattered across customer systems, business systems, data lakes, and legacy systems. While models can write coherent reports using public data, inside enterprises, they must adhere to company definitions and deliver usable results.

Even if data and metrics align, the job is only half done. Reports need approval, data must be written back to systems, tasks stuck in limbo require human intervention, and work must resume seamlessly afterward. Agents must handle not just a single task but an interrupted, error-prone process that must keep moving forward.

As workflows lengthen, costs rise. Amazon Web Services (AWS) example calculations show that when a request requires three tool calls, inference costs can reach 5–10 times that of a single tool call, as each round sends accumulated conversation and tool results back to the model. Added orchestration, storage, network, and security overheads also influence how long enterprises let Agents run and how many tasks they handle.

If Agents can spend money, modify systems, or take external actions, enterprises face more than just costs. Who grants permissions? Who checks results? Who takes over when things go wrong? These must be clarified before action. Individuals can redo a task if the result is wrong, but companies must ensure a single mistake doesn't propagate through subsequent work.

Costs escalate with task complexity, and responsibilities expand with actions—naturally, enterprises won't initially entrust entire workflows to Agents. A more realistic and common approach is to first define an observable, controllable scope, then gradually expand AI tool usage and authorization based on results.

Banco Bilbao Vizcaya Argentaria (BBVA) began collaborating with OpenAI in May 2024, initially granting ChatGPT Enterprise access to about 3,300 people, expanding to around 11,000 by May 2025, and announcing a rollout to about 120,000 seven months later. By June this year, actual users exceeded 100,000—two years after the first accounts were issued.

JPMorgan Chase moved faster. It launched LLM Suite in summer 2024, placing large models in a secure internal environment, covering about 200,000 people from zero in eight months. But this first established a unified, secure employee access point for models, not that the enterprise had handed over data, systems, and action permissions to Agents all at once. Deeper data integration and workflow authorizations remain ongoing.

While entry points are managed by enterprises, Agent operational resources may span internal and external environments. Enterprises must decide which capabilities can leverage external resources and which must remain internal. IDC and the China Academy of Information and Communications Technology surveyed 250 enterprises, finding that among IDC's intelligent computing cloud service users, 85.8% adopted a hybrid architecture combining external AI computing power with in-house data centers.

These choices may seem varied, but they follow the same logic: organizations must regain control over what Agents can access and how far they can go. Models provide capabilities, but a product mechanism outside the model is needed to organize tasks, verify results, and restrict or take over actions when necessary.

02  Who Can Turn Implementation Experience into a Product

Even with the same model, different products can achieve vastly different outcomes. Some products require constant human reminders at every step, while others remember progress, call tools, and halt when errors occur. These differences aren't written into model weights but determine whether Agents can deliver results.

The product runtime framework (Harness) organizes model capabilities into deliverable products. It feeds task-required materials to the model, arranges steps, calls tools, remembers progress, and checks and retries when results are incorrect. When AI Agents enter enterprises, identity, permissions, and auditing must also fit into the same mechanism.

The stronger the model, the less Harness needs to replicate what the model already does, shifting product design focus toward verification, recovery, and governance.

In July 2026, Anthropic revealed that for Claude 5, it adjusted Claude Code by removing over 80% of system prompts without performance loss in programming evaluations, then added necessary mechanisms where the model repeatedly failed. Claude Code lead Boris Cherny also argued that enabling models to verify their own work during execution is one of the most underrated capabilities today.

Unlike individual users, when Agents enter enterprises, the product layer beyond the model extends from a single product to a cloud platform. Cloud vendors combine model services, runtime environments, and organizational governance into one layer. In April 2026, Google Cloud announced Vertex AI would deliver via the Gemini Enterprise Agent Platform.

Breaking it down: a single Agent's task completion is organized by Harness; managing multiple Agents across an enterprise requires a unified registration, authorization, and management system—the enterprise Control Plane. Domestically, Tencent Cloud's WorkBuddy Managed Agents, Alibaba Cloud's BaiLian and AI Gateway, and ByteDance's Volcano Engine AgentKit and HiAgent are filling similar organizational governance capabilities with different product combinations.

Harness and the enterprise Control Plane can enforce rules but cannot create them for enterprises. When Agents are used within a single team, metrics, permissions, and acceptance criteria can be locally agreed upon; once they cross departmental boundaries, these rules must be realigned within the organization.

Anthropic's survey of over 500 U.S. enterprise technology leaders quantified this divide: 57% of respondents used Agents for multi-stage processes, but only 16% for cross-team workflows. In the same survey, 46% cited existing system integration as a major obstacle, and 42% mentioned data access and quality.

Crossing this organizational boundary requires enterprises to both integrate systems and align data. A 2026 KPMG report proposed that for enterprise data to be usable by Agents, it must be searchable, carry sufficient business context, and establish trust through data lineage, machine-readable permissions, and decision rules. Yet data meanings, quality standards, and permission ownership lie with different departments.

For office Agents to enter enterprise production, two transformations must occur. Harness turns models into practically usable enterprise products, while enterprises must embed these products into their production systems and processes. The former determines whether Agents can complete tasks; the latter determines whether they can enter formal production and generate recurring purchases. Agent vendors can provide connectors, permission tools, and evaluation methods, but data interpretation, business definitions, and result acceptance must be consensus-driven by client enterprises.

At this stage, each enterprise's unique data, permissions, workflows, and acceptance criteria first increase the implementation burden per client. For each new client, vendors must reunderstand the business, configure systems, and verify results, with only partial reuse of prior experience.

Compared to traditional SaaS, office Agents cover broader work scopes and theoretically larger market boundaries. Currently, different vendors arrive here with different strengths. Cloud vendors and enterprise software companies are closer to data, systems, and identity frameworks, while independent Agent products can rely on cross-model, cross-cloud neutrality or deep industry methodologies. Multi-model support alone doesn't create a moat; only by managing quality, cost, permissions, and evaluation together can model routing deliver enterprise value.

Tencent, Alibaba, and ByteDance's personal traffic, internal promotion, and cloud resources are certainly advantages; but a replicable business must ultimately be validated by third-party clients. Only when enterprises outside their ecosystems are willing to allocate budgets, data, and action permissions does this competition truly enter the enterprise market.

03  Conclusion

Personal-side popularity can generate buzz, but enterprise-side adoption leaves actionable pricing data.

For investors, the Agent boom isn't entirely good news—it's turning the already opaque cloud and software businesses into an even deeper black box. Previously, markets struggled to discern whether revenue came from computing, storage, databases, or software; now, application seats, model inference, enterprise Control Planes, and implementation delivery are lumped into a single "AI business" category by many companies.

Microsoft discloses an AI annualized revenue run rate exceeding $37 billion and over 30 million paid seats for Microsoft 365 Copilot; Salesforce discloses Agentforce's annualized recurring revenue (ARR) and transaction volumes; Chinese vendors publicly share more active rankings, seat prices, and the proportion of Feishu new customers purchasing AI products.

All seem to prove AI is generating revenue, but these metrics haven't become more granular alongside the numbers. Markets still can't distinguish whether what's being sold are replicable products or projects requiring repeated implementation.

Investors must answer two key questions: Have Agents moved beyond seats and pilots into clients' formal production? Has enterprise deployment begun achieving economies of scale? The former can be measured by production workflow Proportion (market share), manual takeover rates, and renewal expansion; the latter by implementation and customization resources consumed per client, as well as onboarding speed and template reuse rates for subsequent clients.

More detailed disclosures won't automatically boost stock prices but will serve as operational signals. Companies willing to public (publicly disclose) production usage, delivery reuse, and client expansion at least make this business comparable; those only sharing seats, Tokens, and contract numbers must accept market discounts for the opaque portions.

Disclaimer: This article is for learning and communication purposes only and does not constitute investment advice.

Like, share, and repost—your support fuels our updates!