AI-based enterprise: Practical thinking on AI agents and workflows

Written by

Clara Bennett

Updated on:June-28th-2025

If you are interested in the following questions, this article is for you:

1. What is the difference between workflow and agent?

2. Can AI be used in all scenarios? What kind of scenarios are more suitable for combining with AI?

3. What are the key points to pay attention to when building an agent?

The major direction I want to focus on at the moment is enterprise AI, that is, how to integrate AI well into corporate business to help companies achieve the goal of reducing costs and increasing efficiency.

The idea of integrating AI into enterprise business may be what we AI people are more concerned about. For enterprises, the focus is on how to reduce costs and increase efficiency. As for whether to integrate AI, it may not be the most important concern of enterprises. Of course, many enterprises have also seen the capabilities of AI and believe that AI will affect and change the future, so they are more concerned about: with AI, how to reduce costs that could not be reduced before and increase efficiency that could not be increased before.

When it comes to reducing costs and increasing efficiency, the first thing that comes to mind is automation (of course there are other methods). The question then becomes: what kind of business processes can AI automate, and what is the difference between automation that AI can achieve and automation that traditional programs can achieve.

To illustrate this issue, three concepts are divided here: logic program , LLM workflow, and AI agent.

(Abbreviation: LLM: Large Language Model, workflow: workflow agent: agent, AI agent: intelligent agent)

Logic program is a traditional automation program that can handle deterministic tasks with clear rules , certain logic , and structured data flow . Its characteristic is that it completely relies on logical judgment, so it is relatively stable and reliable.

LLM workflow: refers to a manually-choreographed LLM-based workflow. It is centered on manually-choreographed processes, and the execution path, steps, and nodes are all determined. LLM is just a node, just like stringing together LLMs. It can handle tasks with clear processes and require understanding of unstructured data. Compared with logic programs, it has a certain degree of intelligence, can understand information, classify, organize, and make weak decisions on information.

AI agent: It is an intelligent agent with LLM as the core that can understand, plan and execute tasks. The biggest difference between it and LLM workflow is that it has a high degree of autonomy. Autonomy here does not mean that it can do whatever it wants, but that it can plan and execute tasks autonomously after receiving them. The number of iterations is also determined by the agent autonomously. However, the problem brought by autonomy is that the accuracy and consistency of the results of each execution cannot be guaranteed. To make a vivid metaphor: AI agent is like an intern with very rich experience and extensive knowledge. It is very capable, but not necessarily reliable.

After understanding the characteristics of the above three, you will know which method should be adopted in different scenarios.

AI is not omnipotent, it needs to be put in the right place.

Finally, what key points should we pay attention to when building AI agent products?

1. The most important thing is to write the prompt well (it must be accurate).

2. It is necessary to build a self-feedback mechanism for the agent so that the agent output can converge to the desired result (process feedback).

3. When building an Agent product, we should not just focus on improving its reliability and accuracy. This is determined by the underlying technical characteristics of LLM. It is better to wait for the underlying model to evolve than to improve the product. We should pay more attention to facilitating users to check, confirm and modify the product design.

Two viewpoints:

1. Agent is more suitable for ToB enterprise scenarios rather than ToC scenarios.

2. For valuable and complex task scenarios in the enterprise, it is worth doing even if the agent cannot save time, because the agent can be used to expand on a large scale at low cost.