By SOLOMOAT Editorial Team

Source: Tencent Research Institute / Tencent Cloud

Author: Dowson Tong, Senior Executive Vice President of Tencent and CEO of Cloud and Smart Industries Group (CSIG)

August 28, 2026

Executive Editor's Note

On the occasion of Tencent’s annual internal publication Zhidian, Tencent Cloud sat down with Dowson Tong to review two years of frontline AI deployment. His operational observations span foundational models, enterprise applications, autonomous agents, and organizational restructuring.

💡 Core Strategic Takeaway

• Enterprise AI advantage comes from operational integration, not model access alone.
• Autonomous agents shift software from passive tools into active participants in execution.
• First-principles thinking and real user demand matter more than legacy organizational assumptions.

This document was synthesized through a collaborative workflow between human leadership and LLM agents. AI parsed historical speeches and internal records, reconstructing disparate operational concepts into a unified narrative. In modern enterprise workflows, tools no longer merely execute; they participate directly in synthesis and strategic formulation.

Addressing a technological shift of this scale requires grounding in first-principles thinking, stripping away legacy operational dependencies, and listening to user demand.

Capital Endurance Outweighs Early-Mover Advantage

External commentary recently suggested that Tencent was lagging in the AI race. The global foundation model market is fiercely contested, characterized by daily releases, escalating parameter counts, and heavily promoted benchmark rankings. Enterprise clients monitor these metrics closely, and key competitors have maintained a vocal presence in generative AI.

Tencent’s proprietary Hunyuan foundation model underwent multiple internal reconstructions. The streamlined Hunyuan Hy3 architecture prioritized high-quality training datasets, delivering competitive benchmark performance within its parameter class. However, broader internal compute constraints initially slowed training cadences and delayed product rollout schedules.

A cold assessment reaffirms a core market principle: early entry is rarely the decisive variable. The second half of the AI transition has only begun; market structures are fluid, and value distribution across the supply chain remains unfinalized.

The evolution of mobile computing offers a clear historical parallel. In the late 1980s, Nintendo launched the Game Boy; by the early 1990s, Apple introduced the Newton PDA. The late 1990s saw personal digital assistants (PDAs) diverge alongside feature phones, pagers, and specialized electronic dictionaries. The 2000s brought the iPod and dedicated e-readers. Only with the launch of the iPhone and Android did these functions converge into a ubiquitous platform.

Apple required over a decade of iterative product exploration before securing market leadership; meanwhile, early pioneers such as Nokia, Motorola, and BlackBerry lost market position and were marginalized.

Competitive advantage belongs not to the first entity out of the gate, but to the operator that sustains long-term capital deployment, solves concrete bottlenecks within specific operational contexts, filters market noise from genuine demand signals, and maintains operational flexibility to navigate full economic cycles. The foundation model race is a marathon, and the industry has cleared only the first kilometer. Operational endurance is what captures emerging market demand.

The Application Layer: Balancing Model Quality with Systems Engineering

Generative AI’s acceleration traces back to ChatGPT. While the initial breakthrough occurred at the foundation model layer, broad public access was required for capital markets and society to recognize the strategic value of base infrastructure. For platform conglomerates like Tencent, sustained capital allocation toward frontier model research remains mandatory. The upcoming Hy4 architecture is positioned to deliver expanded capabilities.

Yet strategic focus must not remain confined to the foundation model tier. In the mobile era, defining the market solely around handset manufacturers was an overly narrow framework. The primary commercial determinant is whether AI drives measurable utility within target applications.

Tencent’s core moat lies in its established business-to-consumer (B2C) and business-to-business (B2B) application footprint. Every product connects to verified users, concrete workflow demands, and established operating processes. Tencent Meeting uses models for automated synthesis; Tencent Lexiang deploys them as enterprise knowledge engines. Today, Hunyuan supports hundreds of internal production services, while specialized product agents are expanding across business units.

User workflows represent Tencent’s deepest operational asset. Interaction data—whether enterprise-generated or verified through specialized institutions—provides structural context for LLM reasoning. Historical interactions can be distilled into long-term memory or packaged into reusable skills. Many products route queries across multiple foundation models, dynamically categorizing tasks by performance tier and API cost to optimize price-to-performance metrics.

Enterprise deployment requires a synthesis of algorithmic research and systems engineering. Algorithms define the theoretical ceiling; infrastructure engineering determines the velocity of reaching it. From data sanitization and pre-training to distributed inference and workflow integration, every step requires dedicated AI infrastructure.

To match user habits in production, LLMs require a complete operational environment: structured file systems, dynamic tool-calling interfaces, long-context memory stores, and closed evaluation loops. When software harnesses integrate foundation models, enterprise workflows transform. Models now plan multi-step workflows, call specialized APIs, verify execution outputs, and generate high-value economic productivity, securing enterprise procurement budgets.

Enterprise adoption introduces strict governance requirements: access controls for sensitive data, tenant isolation, and granular token budget allocation. Integrating internal enterprise systems through standard skills and the Model Context Protocol (MCP) establishes a long-term enterprise software moat.

Non-Linear Innovation and Capability Accumulation

The emergence of Tencent's workplace agents, CodeBuddy and WorkBuddy, was not the result of a pre-planned corporate blueprint. It grew out of long-term software investments.

Tencent Cloud’s Cloud and Smart Industries Group (CSIG) had long developed software engineering tools. An early acquisition, coding.net, provided source code management and continuous integration/continuous deployment (CI/CD) pipelines, later expanding into cloud IDEs and sandboxed code-execution runtimes. Under legacy SaaS pricing models, developer tools struggled with profitability and dragged on cloud business margins. During subsequent cost-optimization cycles, CSIG sunset several unprofitable SaaS/PaaS lines, but preserved developer tooling investments.

The 2021 launch of GitHub Copilot demonstrated the potential of AI-assisted code generation. Software engineering provides an ideal domain for large language models: code is a structured language with rigid syntax rules and extensive open-source training corpora. Functional boundaries are clear, and synthetic outputs can be validated via compilers, test suites, and runtime logs. Software development became one of the earliest proving grounds for generative productivity.

CSIG’s developer tools team began testing AI code-generation workflows, integrating models into internal IDEs to construct early coding agents and execution harnesses. Concurrently, the Technology and Engineering Group’s (TEG) internal engineering management team was tasked with driving internal software productivity. The two teams collaborated to build and validate an internal AI coding pipeline, with CSIG engineering teams acting as primary internal test users.

Early models were restricted to basic code completion, offering modest efficiency gains. As model architectures matured, code quality and functional scope expanded. In 2024, the team launched CodeBuddy for enterprise clients, generating software revenue through private deployments that funded continued product refinement.

By late 2025, frontier coding benchmarks advanced significantly. Models could sustain long-horizon programming workflows—generating complete modules, optimizing algorithms, validating unit tests, and debugging runtime errors. The proportion of AI-generated code within Tencent climbed rapidly, altering engineering workflows and driving an expansion in internal token consumption.

In early 2026, the team leveraged the CodeBuddy architecture to launch WorkBuddy, featuring an interface designed for non-technical enterprise personnel. Product managers drafted specifications, operations teams compiled business analyses, and managers processed communications and meetings. These workflows share a foundational operating pattern: task decomposition, context retrieval, data synthesis, content drafting, and output verification—forming the baseline for a general-purpose enterprise workspace agent.

When the Tencent Docs business unit was integrated into CSIG, the two groups merged their roadmaps. This added collaborative document editing, bi-directional human-AI co-authoring, and shared knowledge bases to WorkBuddy.

Following its release, WorkBuddy saw rapid enterprise adoption. The engineering team iterated using its own AI-assisted workflows, resolving bugs and shipping more than 40 version updates over three months.

This velocity reflects an altered product development lifecycle:

WorkBuddy and CodeBuddy are now embedded within daily corporate workflows, querying internal databases and administrative systems. Tasks that previously required hours—such as data cleaning, financial variance analysis, or drafting legal briefs—are resolved to a functional initial draft in roughly 20 minutes.

Developing successful AI-native applications requires operational latitude to explore, a focus on workflow friction points, fast prototyping, and continuous user-feedback loops.

The multi-year deployment of Yuanbao, Tencent’s consumer search and chat assistant, provided foundational experience in consumer AI interfaces. While some market observers view the consumer assistant segment as saturated, structured search, information retrieval, and question-answering remain durable consumer demands. As long as response accuracy and synthesis outpace competing alternatives, commercial opportunity remains.

Developing Yuanbao clarified that building AI products extends beyond wrapping an API around a simple chat interface. It requires continuous experimental platforms, multi-model evaluation frameworks, dynamic retrieval strategies, and closed user-feedback loops. These shared infrastructure components are now leveraged across enterprise products like WorkBuddy.

While competitor applications focused heavily on anthropomorphic voice interactions, core utility in search and retrieval centers on output accuracy, verifiable source provenance, and low-latency real-time retrieval.

Over the past year, aggressive user-acquisition campaigns for Yuanbao yielded mixed conversion rates because product-market fit was still early. With Tencent’s broader AI roadmap consolidated, the Yuanbao team has shifted to systematic product refinement. Co-designing Yuanbao alongside the Hunyuan Hy3 model architecture has driven measurable gains in search accuracy, increased user retention, and lowered per-query operating costs.

Architectural Divergence: Personal vs. Service Agents

The enterprise AI ecosystem is bifurcating into two distinct architectural patterns:

Building production-grade service agents using open-source frameworks involves significant engineering overhead and maintenance costs.

Tencent Cloud’s Agent Development Platform (ADP) provides a managed environment for configuring and monitoring enterprise service agents. ADP packages workflow orchestration, retrieval-augmented generation (RAG), tool calling, offline evaluation, security guardrails, and enterprise systems integration into an out-of-the-box management plane accessible via open APIs. Forward Deployed Engineers (FDEs) partner directly with enterprise clients to map internal operating logic, establish data boundaries, and deploy reusable skills and MCP interfaces.

Personal and service agents form a complementary operating loop. When an employee drafts an enterprise client analysis, the personal agent scopes the project and outlines the deliverable. When the task requires audited customer records, current price books, or approved contract templates, the personal agent queries the corresponding enterprise service agent. The service agent verifies identity, enforces data permissions, isolates sensitive records, and returns validated outputs. Integrating these two models provides the structural foundation for enterprise-scale AI deployments.

Restructuring Enterprise Operations Around AI-Native Workflows

AI is fundamentally altering corporate organization and software engineering cadences.

Legacy digital product development followed a sequential assembly line: product managers wrote requirement specifications; UI/UX designers drafted mockups; software engineers implemented business logic; QA ran automated testing matrices; and systems engineers managed production deployment. This structure created high communication overhead, long feedback loops, and friction during requirement revisions.

LLM workflows flatten this structure. Product managers can generate functional prototypes and user interfaces directly; engineers can generate end-to-end codebases alongside comprehensive test coverage; non-technical team members can deploy lightweight tools via natural-language programming. By collapsing operational silos, smaller cross-functional teams can deliver project scopes that previously required large engineering departments.

Tencent is integrating WorkBuddy across its core enterprise suite, including WeCom, Tencent Docs, ima, Tencent Meeting, and iWiki. This enables consolidated document processing, enterprise knowledge retrieval, meeting synthesis, and human-agent collaboration within a single interface. By unifying these high-frequency touchpoints, the underlying model receives real-world task context, user interaction telemetry, and specialized application feedback, driving iterative capability improvements.

Transitioning to an AI-native organization requires structural management adjustments:

Flattening team structures and reducing communication overhead.

Establishing clear evaluation benchmarks upfront.

Integrating automated experimentation and failure analysis directly into core development pipelines.

Moving validation checkpoints upstream, shifting cross-functional assessment across product, engineering, and algorithm design to project inception.

Enduring Operating Principles: Capital Discipline and Human-Centric Utility

Amid rapid technological shifts, two strategic principles remain foundational:

Strategic Discipline and Capital Endurance: The commercialization of AI is a multi-decade transition. Market leadership will not be decided by early sprint velocity, but by balance-sheet strength, long-term capital allocation, and the ability to compound value across market cycles.

Human-Centric Value Delivery: Technology remains an enablement mechanism; user utility is the commercial objective. Enterprise AI deployment must focus on solving structural friction points, improving operational interfaces, and driving measurable productivity gains rather than chasing novel technical features.

An operational deployment in Beijing’s Miyun District demonstrates this dynamic: a frontline municipal administrator utilized WorkBuddy to configure an automated emergency response application integrating geological hazard alerts, precipitation telemetry, and municipal evacuation registries. A workflow that previously required dozens of personnel coordinating manual phone verification overnight was consolidated into a single operator distributing automated warnings and aggregating verified evacuations across 16 townships within an hour.

AI does not replace human responsibility; it strips out repetitive, high-friction operational overhead, allowing human capital to focus on strategic execution and critical decision-making. Commercial returns and market positioning follow products built on measurable real-world utility.

The AI transition is accelerating. Navigating rapid market shifts requires anchoring operations in foundational business principles: compounding structural capabilities, verifying product-market fit, and developing deep integration within core vertical workflows.

Market performance is rarely captured by participants reacting to short-term industry trends. It is secured by long-term operators who understand their structural competitive advantages, allocate capital toward high-conviction vertical problems, and maintain disciplined execution across economic cycles. Filtering out transitory market noise and focusing resources on measurable operational transformation forms the foundation for durable enterprise value.

❓ Frequently Asked Questions

What did Tencent learn from two years of AI deployment?

AI value compounds when models, agents, workflows, and organizational structures are redesigned together rather than deployed as isolated tools.

How should enterprises approach autonomous agents?

Start from measurable user demand, define clear execution boundaries, and rebuild workflows around agent capabilities and human accountability.

🎓 Deepen Your Strategic Mastery

Turn these strategic insights into a focused, defensible one-person-company operating system with the Garbo Decodes China mini-MBA.

Explore Garbo Decodes China