Summary of Key Points
The global AI industry is shifting from a focus on “which model is larger or can reason faster” to one where the goal is to create real value with intelligent agents (AI assistants that can complete tasks independently). Companies are no longer just experimenting with the use of these agents; instead, they are seeking large-scale applications. However, this requires not only technical expertise but also cognitive and organizational capabilities. As large-model technology becomes more widespread, the true differentiator lies in a company's ability to transform general AI into a tool tailored for its own needs. The article discusses the development trends of intelligent agents, the specific paths companies are taking to deploy them, the practical challenges they face, and corresponding solutions.
1. The Shift in Focus of the AI Industry: From “Building Large Models” to “Using Agents to Get Work Done”
Previously, there was a competitive race to build the largest models with the most parameters, the fastest reasoning speeds, and the ability to process text, images, and videos (multimodal capabilities). Now the focus has changed:
- Technologically: AI has evolved from being able to communicate effectively to being capable of taking action. For example, while it could previously only answer questions, it can now call on tools (such as data retrieval or operating systems) and perform tasks like planning, executing, checking, and correcting errors.
- Architecturally: Instead of relying on a single large model, composite systems are used. An intelligent agent typically consists of multiple components, including a task planner, a tool invocation layer, a memory system, and a security sandbox, all connected through a unified interface (similar to how all smartphones use the Type-C charging standard).
- Value-wise: The source of profit has shifted from the model itself to its applications. Intelligent agents in specialized industries (such as law, healthcare, and customer service) are growing rapidly. One company achieved $100 million in revenue within 12 months by breaking down industry workflows into small tasks that AI can perform reliably, such as analyzing contracts into steps like extracting clauses, checking for risks, and generating reports.
- Security-wise: The focus has shifted from ensuring content compliance (e.g., avoiding illegal content) to ensuring the behavior of AI is trustworthy (e.g., preventing it from misusing systems or granting excessive permissions). Companies now use security measures like sandboxes to limit AI’s actions and include human review processes to prevent issues.
2. The Four Critical Steps for Companies to Deploy Intelligent Agents
To successfully implement intelligent agents, follow these steps:
1. Choose the Right Use Cases: Prioritize tasks that are frequent, have clear rules, and can tolerate errors. Examples include customer service responses (high frequency, fixed rules, and possible corrections) and expense reimbursement reviews (repetitive, with clear standards). For more complex tasks like credit approvals or medical diagnoses, use AI as a supplement rather than making decisions on its own.
2. Leverage Data Engineering: Train the agent using your company’s proprietary data. Organize this data (from emails, documents, and systems) into a format that AI can understand, such as structured graphs. This is time-consuming but creates a competitive advantage, as large models are available to everyone, but your company’s specific knowledge is unique.
3. Reengineer Processes: Modify existing processes to be executable by AI. For example, break down tasks like report generation into steps like data retrieval and chart creation, and grant AI access to ERP systems. Critical steps should still be reviewed manually to prevent errors.
4. Organizational Change: Implementing AI requires organizational transformation. Reassign roles: some tasks can be automated (e.g., data entry), while others require human-AI collaboration (e.g., AI drafts contracts for review). Train employees to give clear instructions to AI and overcome concerns about AI taking over jobs.
3. Three Common Challenges in Implementing Intelligent Agents
Many companies find that the reality of using intelligent agents differs significantly from expectations:
1. High Performance but Instability: AI may perform well in tests but fail in real-world applications. This is because AI uses probabilistic reasoning, and the nature of errors can vary with model updates.
2. Blurred Lines between Autonomous Decision-Making and Human-AI Collaboration: Companies either rely too heavily on AI (leading to panic when issues arise) or use it merely as a tool without fully leveraging its potential. For example, if AI is used to manage systems without clear guidelines for human intervention, trust may be lost.
3. Cost Overruns and Difficulty in Measuring Value: Unlike traditional software, intelligent agents require ongoing costs (for computing power and tokens). The operational expenses can exceed development costs, and it’s hard to quantify the benefits in terms of increased revenue or cost savings. For example, while AI may save time in customer service, there is no clear way to measure whether customer satisfaction has improved.
4. Practical Recommendations for Companies
1. Map Out Tasks: Identify high-frequency, high-value tasks that can be verified automatically. Use AI for these tasks first.
2. Consolidate Knowledge: Transform implicit company knowledge (e.g., from experienced employees or cross-departmental practices) into structured data that AI can process. Start with mature business processes and refine them through practical use.
3. Gradual Process Transformation: Large companies should start with stable areas (e.g., expense reimbursement) before making sweeping changes. Small companies can use simpler, natural-language-driven intelligent agents.
4. Dual Measurement: Evaluate both the financial benefits (e.g., cost savings) and the effectiveness of AI’s operations (e.g., accountability). This ensures that investments in AI are worthwhile.
In Conclusion
Implementing intelligent agents is not just a technical issue; it requires a combination of technology and organizational changes. Companies need to choose the right use cases, leverage their data effectively, reengineer processes, and adapt their organizations to turn AI into a valuable tool for generating revenue.