[Government AI GENAI] Intermediate agent chat
Table of Contents
- Introduction: Intermediate Agent Chat: Practical Examples and Important Notes
- Case 1: Cross-section consistency check of a collection of questions and answers in preparation for answering questions in the Diet
- Case 2: Progress management and organization of huge reports
- Precautions for Use
- Summary and Inquiry Desk
Introduction: Intermediate Agent Chat: Practical Examples and Important Notes
In "Elementary Agent Chat," we introduced how to use Agent Chat for daily administrative work. In the intermediate edition, we will look at more applied usage, such as handling documents that span multiple departments and massive documents that exceed 100 pages.
The basic flow is the same as Getting Started - you can simply drop off items on My Drive and ask for them - and you can save the steps for repetitive tasks as "skills."
In this article, we will focus on checking the consistency of the anticipated collection of questions and answers in the Diet and confirming the progress and structure of the report, which has a huge number of pages. Please refer to this and grasp the utilization image for large-scale operations. In addition, we will explain the points to note when using it and the points for security.
Case 1: Cross-section consistency check of a collection of questions and answers in preparation for answering questions in the Diet
The first case is an example of verifying the consistency of the anticipated questions and answers for the Diet answers across divisions and offices.
The anticipated Questions and Answers are generally prepared by several sections and offices sharing the work to be done. However, when you read the entire document, there may be subtle differences between sections and offices in terms of numerical values, terminology, and the degree to which answers are discussed.
This last-minute cross-checking is one of the most onerous aspects of dealing with Congress. Your AI agent can help.
This is an explanatory illustration in which the AI agent supports the consistency check of the anticipated questions and answers compiled by each section.
[Operation Flow]
This time, the anticipated questions and answers prepared by each of the 5 sections will be uploaded to My Drive.
In Agent Chat:
Please summarize the questions and answers that each section was in charge of into a single document, and then thoroughly confirm that there are no discrepancies in the numerical values, terminology, and the degree to which answers are taken between departments. If there are any discrepancies, please list them clearly with the question number and the name of the relevant section.
You do not need to manually reassemble the documents beforehand.
After you submit your request, you will see on your screen how the AI agent proceeds through the process. Once completed, a list of all discrepancies will be saved to My Drive.
In the list, discrepancies in numerical values and differences in the degree to which answers are taken are sorted out by item.
Because the question number and the name of the department in charge are also specified, it is easy to understand where to confirm with which department.
[Notes]
Agent Chat cannot determine which statement is correct. Please be sure to check with each section and office on the points you have pointed out, and coordinate the overall response policies to unify them. It is important not to leave the decision on which one to unify to an AI.
[Save and reuse as skills]
Save these steps as a Skill for future use.
Save your current workflow as a skill
The AI Agent will then create a markdown file of these steps and place it in the "skills" folder on My Drive.
The next time you perform the same type of task, simply request it using these skills, and the process will be the same as before.
Case 2: Progress management and organization of huge reports
The second case study is about a large report written by several people. In this case study, the authors check the writing progress for each chapter and point out duplications and omissions in the report structure.
In reports of more than 100 pages, it is difficult to understand which chapters have been completed and whether there are any duplications or omissions in the chapter structure throughout.
The longer a project goes on, the more likely it is that even the authors lose sight of the big picture. It's up to the AI agent to make sure.
[Operation Flow]
This time, I will upload the draft report of all 12 chapters and over 100 pages to My Drive. The density of writing differs greatly depending on the chapter, and it is impossible to read how far it has progressed only from the table of contents.
In Agent Chat:
This draft is written by a group of authors. For each of the 12 chapters, list the status of the work in three stages: Completed, Not Started, and Needs Revision. If there are any overlaps or omissions in the structure of the chapter structure, please indicate these in a document with the chapter number clearly written.
Ask.
After submitting your request, you will see on screen what the AI agent is doing. When the process is complete, a document will be saved to your My Drive, summarizing the writing and organization of each chapter.
The document lists the status of each chapter and the reasons for the decision. For example, chapter 4 was marked as "Revision Needed" because "many of the sections are bullet points only with missing paragraphs."
In addition, Chapter 7 has been judged as "not started yet" because "all the sections have only' to be written' and' TBD', and there are three placeholders (temporary empty frames) for figures."
In addition, it has been pointed out that there are structural problems, such as the content of Chapters 3 and 8 being duplicated.
Stay on top of writing and organization issues without having to re-read through more than 100 pages of material.
[Notes]
The Complete rating does not guarantee the quality or veracity of a chapter's content. The ratings are based on the presence or absence of writing and on the amount of writing, so please be sure to check the content yourself before submitting it.
[Save and reuse as skills]
This procedure can also be saved as a "skill" if you want to reuse it in future business operations. Once saved, there is no need to rewrite the request every time from scratch, and business know-how can be accumulated as an asset.
Precautions for Use
The following sections introduce three points to keep in mind when using Agent Chat.
(1) Output error of AI
The first is an output error due to AI. Not only in agent chats, generative AI may output content that differs from the facts and views that are not supported by evidence. Please use the output results as a basis for discussion, and be sure to make the final confirmation and judgment yourself.
② Files with unknown reliability
Second, you can load untrusted files. When you allow AI to load an externally received file, malicious instructions embedded in the file can cause AI to behave differently than originally intended. These instructions are difficult to notice by looking at them. Before loading such instructions, make sure that you know who sent the file and how it was obtained.
(iii) Confidentiality of information handled
The third is the sensitivity of the information handled. GENAI can handle information up to Confidentiality 2 level. However, please judge what to enter in accordance with the information handling rules of your affiliated organization. For information that is difficult to handle, please check with the relevant department of each organization before entering it.
Summary and Inquiry Desk
This time, we have introduced two examples and points to note when using it. From checking the consistency of a hypothetical collection of questions and answers that spans multiple sections and offices, to confirming the progress and structure of a report that exceeds 100 pages, the method of placing a request on My Drive remains the same, even for large documents.
In addition, by saving work procedures as "skills", it is easy to proceed with the same quality when doing similar work next time.
On the other hand, when using generative AI, please be sure to be aware of the following points.
- Be sure to check the AI output yourself.
- Note the Origin of Imported Files
- The confidentiality of input information shall be determined in accordance with the information handling rules of the affiliated organization.
If you have any questions, please contact the Help Desk in GENAI.
Please refer to the following links for the introduction and beginner versions of Agent Chat.