What is the difference between AI agents and AI workflows? Explaining the key points for using them

AI Agent and AI Workflow Characteristics

First, let's organize the characteristics of each mechanism.

To achieve a certain goal, an AI agent autonomously generates and executes steps, such as "Let's do this next." While it has the flexibility to change its behavior depending on the situation, it can be difficult to predict the results and costs.

In contrast, AI Workflow is a system in which AI is incorporated into procedures determined in advance by humans, and processing is carried out along a set route. While the advantages of this system are stable quality and the ease of tracking the process later, it also has the disadvantage of being halted if an unexpected event occurs.

In actual work, rather than using only one or the other, it is more realistic to use them in a workflow for routine tasks and leave the parts that require more difficult decisions to agents.

AI Agent: AI thinks autonomously and tries and errors

When given a goal, the AI agent takes a dynamic approach, repeating the cycle of "gather information, plan, act, and reflect" as it moves forward, sometimes changing course as new facts emerge and sometimes conducting additional research if necessary.

This flexibility is a great advantage when the desired outcome is a little unclear or when the situation is constantly changing. On the other hand, even if you ask someone to do the same thing, they may take a different route each time, which can make it difficult to explain why the result was achieved.

The main features of AI agents are as follows:

• Autonomous decision-making: AI agents dynamically choose the next action to take depending on the situation at hand. There is no need to determine all branching conditions in advance; for example, if they determine that there is insufficient information, they will conduct their own search, or if a hypothesis appears incorrect, they will change the approach to analysis.
• Tool-based extensibility: AI agents can extend their capabilities by using tools such as web searches and database searches. The more tools available, the more they can do, but the greater the risk of making mistakes or handling information inappropriately.
• Resilience in the event of an error: If the execution result is not satisfactory, consider what the cause was and try a different prompt or use a different tool, thus creating a self-correcting loop.

AI Workflow: Execute defined procedures in collaboration with AI

AI workflow is a method of running AI and systems along rails laid out by humans. Even if there are branching paths, the conditions and destinations are predetermined.

On the other hand, if an exceptional case that is not in the manual occurs, the process will come to a halt and will not be able to be handled. How to handle such exceptions is usually handled by a human outside the workflow, or by partially combining agent-like elements.

The main features of the AI workflow can be summarized as follows:

• Reproducibility of procedures: Because AI workflows follow set steps, they tend to produce similar results for similar inputs. This makes them ideal for tasks where a certain level of quality must be maintained or where deadlines are somewhat fixed.
• Process visibility: Because the workflow is clearly divided into steps, it is easy to see at a glance where time is spent and where errors have occurred. This is an operational benefit as it allows for quicker identification of the cause when a problem occurs.
• Processing controllability: With AI Workflow, you can precisely control the content of instructions, the degree of freedom of the AI, the data to be referenced, etc. It is important to use it appropriately, leaving some freedom for tasks such as "compiling ambiguous information," which is AI's forte, and using logic processing to control numerical calculations and registration tasks, which must never be wrong.

Back to Top

Case study of different usage depending on business characteristics

Rather than discussing which system is better, it's easier to imagine choosing from the perspective of whether the procedures for the work you want to handle can be determined in advance. The basic idea is that if the procedures can be fixed, it's Workflow, and if on-the-spot decisions are required, it's Agent.

When is an AI agent suitable?

AI agents are well suited to exploratory tasks where there is no single answer and the next step is not clear until you try it.

• Sales data analysis and insights: For example, when working on a task such as "finding the cause of a drop in sales," initial predictions often turn out to be wrong. This requires trial and error, such as first looking at the overall picture, then re-examining by region, and if that doesn't solve the problem, narrowing down to a specific product category and digging deeper. An AI agent can support this process of "looking at the results and then deciding on the next analysis method."
• Customer support support: Customer inquiries are often fragmented, and the true problem is often hidden behind the words. The process of narrowing down the situation by asking additional questions and searching for an appropriate solution is an area where agents excel. By changing the tone of their responses depending on the situation or finding the most appropriate part of the manual, they can help ensure a smooth resolution.
• Searching for internal knowledge: To answer vague questions like, "Is there any documentation on that somewhere?", it's necessary to search across internal chats, files, wikis, etc. An AI agent can understand the intent of the question, devise search keywords, and compile an answer by collating information from various sources.

Cases where AI workflow is suitable

AI workflows are suited to routine tasks with clear quality checkpoints.

• Customer analysis and reporting: Workflows are suitable for processes such as aggregating customer data, creating reports, and registering the results in a sales force automation (SFA) system. This is because the steps of data acquisition, aggregation, document creation, and review are clearly separated.
• Data preparation for utilization: Workflow is also good at organizing a database by extracting necessary items from documents in various formats and tagging them. Accuracy can be measured at each process, making it easier to identify where errors are likely to occur.
• Operational anomaly detection: The task of detecting and reporting system anomalies is standardized, from detection to reporting and ticket issuance. Numerical values are checked using an accurate system, and AI is used to convert complex logs detected into human-readable summaries.

Back to Top

Key points for improving accuracy in mounting

Even if you have a great system, if it is not well-designed, it will not be used in the field. We will organize the points to take the accuracy of both agents and workflows to the next level.

Tips for implementing AI agents

The more autonomous an AI agent is, the more likely it is that it will become unclear what constitutes success and when to stop. The essence of improving accuracy is to create a framework that prevents deviation while taking advantage of its freedom.

Furthermore, it is more important for an agent to "fail gracefully" than to "behave intelligently." We design behaviors such as returning the agent to the person when it fails, or not responding when there is no basis for doing so. For example, it is important to consider the following points when designing:

• Error handling: Validate the output using checklists and constraints, identify any shortcomings, and then retry. Retry also requires design, classifying the type of failure and deciding whether to compensate with additional searches, change the prompt, or switch to a different tool. Blind retries will increase costs.
• Tool usage definition: For each tool, clearly define its purpose, input format, output format, restrictions, permissions, and prohibitions as a description. Designing a priority order is particularly effective. For example, you could decide on an order such as searching the internal knowledge base first, and only allowing web searches if that is not enough.
• Maintaining context: The longer the task, the greater the problem of the limit on the amount of information that LLM can retain. If important information is buried, you may mistakenly assume certain things or repeat the same tasks. The solution is to separate the roles of short-term and long-term memory. Keep the assumptions necessary for decision-making, unresolved issues, and evidence that should be cited, while summarizing and compressing chatter and unnecessary intermediate products.

Tips for implementing AI workflow

The accuracy of a workflow depends on how finely and simply you divide tasks. Specifically, it's important to design it with the following points in mind:

• Breaking down tasks: If you give instructions to do everything at once, it's hard to know where the mistake was made and it's difficult to correct it. Breaking down tasks into stages, such as input understanding, extraction, conversion, generation, and verification, and having intermediate deliverables allows for early detection of errors. Stopping small errors in the early stages prevents them from spreading to later stages.
• Evaluation and return by LLM (LLM-as-a-Judge): This method involves having the generated content evaluated by another LLM, clarifying the evaluation criteria and scoring based on accuracy, format, rationale, and rules. If the content does not meet the threshold, it is returned to the previous process.
• Distinguishing between deterministic logic and AI: If LLM is used to perform processes where the answer is unique, such as numerical calculations, aggregation, normalization, string replacement, and ID matching, even the slightest fluctuation can lead to accidents. It is essential to implement these processes reliably in a program. On the other hand, areas where ambiguity remains, such as classification, summarization, text generation, and intent understanding, are areas where LLM excels.

Back to Top

How data integration can bring out the value of AI

The success of AI applications depends not only on model performance, but also on the quality of the data available and the integration design that allows it to be retrieved when and in the form required. data integration is not simply a connection; it is an important function that includes search quality, freshness, permissions, and even preprocessing.

▼I want to know more about data integration
⇒ data integration / data integration platform | Glossary

The relationship between AI agents and data integration

The agent will select a reference source depending on the situation, and retrieve additional information if necessary. Therefore, aggregating data, making it easy to search, and providing it as a tool directly impacts performance. To summarize, data integration should be considered from the following perspectives:

• Collect data into a knowledge base: Internal data is often scattered across file servers, wikis, tickets, SFA/CRM, logs, etc., and is only accessible to those who know where it is. First, consolidate it into a knowledge base that can be referenced by AI.
• Data preprocessing for AI: Preprocessing can have a major impact on the accuracy of searches and answers. Dividing documents into appropriate units and adding metadata through tagging and entity extraction makes it easier to find the information you need and improves the interpretability of AI.
• Use of tools for direct integration: For agents to access internal data sources, it is practical to provide tools such as APIs, search platforms, and DB queries. Tooling allows you to define input schemas and constraints, reducing the chance of operational errors.

The relationship between AI workflow and data integration

Workflows function as the plumbing that connects the upstream and downstream business systems. By designing the processing order, dependencies, and even the reflection of results, it becomes automation that can be used in the field. For example, data integration is effective in realizing AI workflows in the following ways:

• ​Workflow definition: A workflow is defined as a series of processes from trigger to result reflection. Decide whether it will be executed periodically or event-driven, and connect data acquisition, LLM processing, validation, notification, and registration.
• Synchronization between workflows: On-site automation is not a one-off process; multiple workflows are chained together, including pre-processing, analysis, and registration. Dependencies must be managed, and synchronization is required to ensure that subsequent steps are started only after the previous steps are completed.
• Writing back output results: To turn AI output into value, it is important to design it so that it is written back to the places where it is used in the field, such as SFA/CRM, tickets, and BI (so-called reverse ETL). Being able to always feel the effects of AI in the business systems you use leads to usability in the field.

Back to Top

Finally

AI agents and AI workflows are not competing concepts, but rather it is effective to use them in a complementary manner depending on the uncertainty and quality requirements of the business.

The strength of AI agents is their flexibility, allowing them to change their behavior depending on the situation, but if they are not controlled, results and costs are likely to fluctuate.The strengths of AI workflows are reproducibility and auditability, but they are vulnerable to exceptions.

The first step to successfully utilizing AI is to determine whether your company's operations require "flexible thinking" or "reliable execution." In either case, it is essential to consider not only AI but also data integration.

Please feel free to contact Saison Technology regarding data integration with AI.

Back to Top

The person who wrote the article

Affiliation: Data Integration Consulting Department, Data & AI Evangelist

Shinnosuke Yamamoto

After joining the company, he worked as a data engineer, designing and developing data infrastructure, primarily for major manufacturing clients. He then became involved in business planning for the standardization of data integration and the introduction of generative AI environments. From April 2023, he will be working as a pre-sales representative, proposing and planning services related to data infrastructure, while also giving lectures at seminars and acting as an evangelist in the "data x generative AI" field. His hobbies are traveling to remote islands and visiting open-air baths.
(Affiliations are as of the time of publication)

Recommended Content

Related Content