Summary of Key Points
This article highlights a critical issue through a dispute involving an AI customer service on a used car platform: the AI claimed to include the transfer fee in the refund, but the sales staff overruled this claim. It also discusses two experiments conducted by Taobao with AI customer services. The main takeaway is that when AI evolves from a “tool to assist customer service” to an “agent that can handle transactions independently,” companies cannot simply integrate AI into existing processes. Instead, they must redesign the underlying data, rules, permissions, and procedures. Otherwise, there will be discrepancies between what AI says and what it actually does, reducing efficiency. The true value of AI lies in whether a company can adapt itself to work with AI, rather than just using more powerful models.
I. Two Roles of AI Customer Service: Assistant or Autonomous Agent?
Taobao conducted two experiments to clearly illustrate the differences between these two roles:
- First Experiment: AI as an Assistant
An AI assistant was provided to new customer service agents. The AI helped identify users’ issues, find relevant information, and draft responses. However, the final decisions on whether to refund the money and whether to send the notifications were still made by humans. The result was improved efficiency—agents could handle issues faster, and novices performed as professionally as experienced staff, with a lower user dissatisfaction rate. In this case, AI was merely a tool that did not disrupt the company’s existing processes (such as approval processes or responsibility assignments), resulting in minimal friction.
- Second Experiment: AI Takes the Lead
AI directly handled user interactions, with human staff supervising from the background. The processing speed increased by 16.8%, but user satisfaction decreased. Why? Although AI completed the conversations, many issues were not resolved properly. For example, the AI suggested a refund, but the backend system was unable to process it, or the process had to be restarted when human staff took over. This shows that while AI can handle some tasks, the company’s outdated processes (such as data inconsistencies or slow approvals) hinder its effectiveness.
In short, using AI as an assistant only optimizes individual tasks, while using it as an autonomous agent requires a complete overhaul of the entire process, which significantly increases the complexity.
II. Why Does Integrating AI into Old Processes Lead to Efficiency Gaps?
The article uses a historical analogy: in the steam age, factories relied on central shafts to power all machines, with the layout of the buildings and equipment designed around these shafts. When electric motors were introduced, simply replacing the steam engines did not improve efficiency until the buildings and processes were redesigned to accommodate the new technology. Similarly, integrating AI into old systems without making corresponding changes results in suboptimal performance.
III. The Gap Between AI’s Promises and Actual Actions
It’s easy for AI to say it can perform a task (e.g., refunding 500 yuan), but implementing it requires two critical processes:
- Task Process: AI must complete a series of steps sequentially: identify the user’s need, locate the order, check the payment status, apply refund rules, determine eligibility, process the refund, and record the transaction. If any step is interrupted (e.g., the AI finds the relevant rule but needs manager approval, and the manager is in a meeting), all previous efforts are wasted. In Taobao’s experiments, only 5.8% of interactions were successfully handled by AI due to disconnected task processes.
- Authorization Process: Companies must grant AI the necessary permissions to access real order data, apply refund rules, and manage company funds. For example, AI may be able to automatically process small refunds, but larger amounts may require additional verification. Both processes are essential; the first determines the scope of AI’s actions, and the second ensures the company’s systems can handle the tasks securely.
IV. AI in Procurement: Similar Challenges
These issues are not unique to customer service. For example, in procurement:
- AI as an Assistant: AI can help sort supplier information and draft inquiry emails, improving efficiency without changing processes.
- AI as an Autonomous Agent: AI can automatically identify inventory shortages, contact suppliers, and submit purchase orders. However, challenges arise when it comes to accessing real inventory data and making purchasing decisions. If each step requires manual approval, the benefits of AI are lost.
The solution is to start with simpler tasks where rules are well-defined and the outcomes are controllable (e.g., small-scale replenishments), and gradually expand the use of AI as the processes become more streamlined.
V. The Advantages of Mature Companies and AI-Native Companies
Some may wonder if companies without existing processes are better suited for AI. However, both types face challenges:
- AI-Native Companies: They can design processes from scratch but lack historical data and experience, which can lead to errors in handling unusual situations.
- Mature Companies: They have extensive data and rules, but these are often embedded in outdated systems. Transforming this experience into usable formats for AI is crucial.
The winner will be the company that can effectively integrate AI into its operations, whether by developing new processes or adapting existing ones.
In Conclusion
AI is not just a “plugin” but a tool that can fundamentally transform a company. The real value of AI depends on whether a company can restructure its processes, data, and rules to make them compatible with AI. Otherwise, even the most advanced AI will be limited by outdated systems.