AI Chatbots Lack Agency — Why the Shift to 'Digital Colleagues' Redefines Infrastructure Spending

A research paper published by Tencent and several Chinese universities identifies a fundamental ceiling in current Large Language Model (LLM) utility: the lack of agency. The study argues that AI will remain a mere tool rather than a 'digital colleague' until it can operate within persistent work environments and execute multi-step tasks without human prompting.

The Agency Gap Prevents AI from Replacing High-Value Labor

Current generative AI models function primarily as sophisticated autocomplete engines that respond to discrete user inputs. This reactive nature limits their economic value to information retrieval and content drafting rather than true labor replacement. The Tencent research paper (published in 2024) suggests that true productivity gains require a transition from chatbots to agents capable of autonomous task completion.

The researchers argue that the current paradigm relies too heavily on the 'prompt-response' loop, which requires constant human supervision. To move beyond this, AI must possess what the researchers call 'agency'—the ability to set goals, use tools, and correct its own errors over long durations. Without this capability, the ROI (Return on Investment) on enterprise AI-integration remains capped by the necessity of human oversight.

This gap creates a massive distinction between 'generative' AI and 'agentic' AI. While generative models create content, agentic models execute workflows. Until the industry bridges this gap, the massive capital expenditures (CapEx) seen in data center construction may face a period of diminishing returns as companies realize chatbots alone cannot automate complex business processes.

Persistent Workspaces Will Drive the Next Wave of Cloud Demand

Most current AI interactions are ephemeral, meaning the model 'forgets' the context once a session ends. The Tencent study highlights that for an AI to become a colleague, it requires a 'persistent workspace'—a digital environment where it can store long-term memory and track progress on ongoing projects. This shift represents a fundamental change in how data must be stored and processed.

If AI agents are to work alongside humans, they cannot exist solely in a stateless API (Application Programming Interface) call. They require stateful environments where they can access files, interact with software, and maintain a continuous presence. This requirement will likely drive a secondary wave of demand for specialized cloud infrastructure designed for long-running, autonomous processes rather than short, bursty chat sessions.

This evolution moves the bottleneck from raw compute power for training to sophisticated memory architectures for inference (the process of a trained model providing an answer). Companies that provide the 'operating systems' for these agents—rather than just the models themselves—may capture the most significant value in the next phase of the AI cycle. The transition from'model-as-a-service' to 'agent-as-a-service' will redefine the competitive moats of the hyperscalers.

Reusable Skills and the End of Single-Use Prompting

The research posits that the path to digital colleagues lies in the development of'reusable skills' rather than just larger parameter counts. Currently, an LLM's ability to perform a task is highly dependent on the quality of the prompt provided by a human user. This creates a friction point that prevents scaling across an entire enterprise.

A true digital colleague must be able to learn a specific company's workflow and store that knowledge as a modular skill. Instead of a human explaining how to file an expense report every time, the agent should develop a'skill' for that specific software interface and internal policy. This modularity allows the AI to accumulate value over time, much like a human employee does through experience.

For investors, this suggests a shift in focus from the 'foundation model' layer to the 'application and orchestration' layer. While the giants like OpenAI and Google dominate the foundation models, the companies building the 'kill libraries' and the orchestration layers that manage these agents will likely see higher stickiness. The ability to turn a generic model into a specialized, skill-bearing agent is the key to enterprise-grade-reliability.

The Shift from Chatting to Executing Redefines the Labor Moat

The current fear of AI-driven job displacement is often predicated on the ability of AI to write text or code. However, the Tencent study suggests that the real disruption occurs when AI can navigate a complex software ecosystem to complete a goal. This requires more than just linguistic intelligence; it requires 'world models'—an understanding of how different software tools and business processes interact.

This capability changes the-nature of the 'human moat.' If an AI can execute tasks, the value of human labor shifts from 'doing' to'verifying' and 'orchestrating.' The workers who thrive in this new era will not be those who can write the best prompts, but those who can manage fleets of digital agents and audit their outputs for accuracy.

This transition will likely be uneven across sectors. Industries characterized by highly structured digital workflows—such as accounting, legal discovery, and supply chain management—are much closer to the 'agentic'-ready stage than creative or highly empathetic-driven fields. The economic impact will be felt most heavily in sectors where the primary task is the movement and transformation of data within software environments.

Key Developments to Watch

• MSFT (Microsoft) — Watch for updates to Copilot-driven 'agentic'-capabilities in M365-the integration of autonomous task-execution within the Office suite (by Q4 2025)
• OpenAI — Any release of 'Operator' or similar agentic frameworks that move beyond the ChatGPT chat interface (through 2025)
• NVIDIA Earnings — Monitor guidance regarding inference-specific chips (L40S and similar), as agentic workloads require high-speed memory for long-context-window processing (quarterly)

If the value of AI shifts from 'answering questions' to 'completing tasks,' will the current leaders in LLM development be able to pivot fast enough to own the workflow layer?