How to create dummy data for generative AI training? Practice tasks similar to practical work and confirmation before inputting
While we would like to practice in generative AI training that is similar to actual practice, we cannot simply input customer names and internal documents into an external service. What is useful here is dummy data that reproduces only the business structure. If you simply replace the text with a fictitious company name, you may miss personal names, transaction terms, and undisclosed information that remain in the text. In this article, we will explain everything from how to create practice assignments to checking them before inputting them.
First, decide on one task that you can do after the training.
Before choosing the topic, decide what deliverables the students will take home. If you're in charge of sales, you might create the next email to send from notes from a fictitious business meeting, if you're a manager, you might organize decisions and responsibilities from notes of a fictitious meeting, if you're in charge of general affairs, you might categorize inquiries. If your goal is to try out multiple tools, it will be difficult to decide how to use them in your work after the training.
The exercise consists of three points: input materials'', expected outputs'', and ``points for people to check''. For example, if you are practicing writing a memo for a business meeting, make sure that even if the e-mail is well-written, it does not arbitrarily promise delivery dates or prices.
Procedure for creating dummy data
- Decide on the type of source material: Match it to the format used in practice, such as emails, minutes, tables, etc.
- Create fictitious content from scratch: Rather than slightly rewriting actual customer materials, creating a new fictitious project will make it easier to avoid forgetting to delete the original information.
- Reproduce variations that occur in practice: Include a small number of abbreviations, ambiguous requests, omissions, etc., and test whether the AI can confirm the unclear points.
- Prepare examples of correct answers and confirmation criteria: Rather than making the only correct sentence expression the correct answer, score the retention of facts, existence of evidence, and compliance with prohibitions.
- Check the usage conditions of the input destination: Check the contract, storage, learning usage, and internal usage rules according to the training environment.
Even with "anonymized" materials, it may be possible to guess the identity of the recipient based on a combination of dates and amounts. When using public information or actual projects, please check with the person in charge of information management within your company before using it.
Example of training assignment: hypothetical business meeting memo
In a hypothetical negotiation memo, write The customer wants a trial implementation within the next month,'' The cost has not been determined,'' and ``The person making the decision is not participating.'' Students will have the AI create a draft for their next confirmation email and revise it from the following perspectives.
| Check items | Points to see |
|---|---|
| Fact | Are the costs and implementation date written as fixed matters? |
| Missing information | Have you been able to ask questions about the person making the decision and the scope of the test? |
| Expression | Is your request to the other party clear and not overly assertive? |
| Reuse | Is there a confirmation procedure that can be used in other projects? |
These types of assignments go beyond just memorizing "good prompts" to practicing judgment before using the output for work.
You can check the content and prices of AI training for managers based on your company's business on the training site below.
https://ceo-claude-code-kennsyuu.onrender.com/
When measuring effectiveness, compare the same tasks before and after training
We look not only at course satisfaction, but also at the time required from creating the first draft to human confirmation, the number of revisions, and the number of factual errors. Since it is difficult to compare if the input conditions are significantly different, we will use hypothetical tasks of the same level before and after the training. If the number of missed checks increases even if the time is shortened, it is necessary to review the procedure.
FAQ
Will training be less effective without actual customer data? The first basic exercise can also be done with fictitious information. Validate it under your company's approved environment and information management rules and against actual formats and exceptions before deployment.
Can dummy data be generated by AI? It is possible, but a person will check to make sure that the generated fictitious data does not accidentally include real names and that it meets the correct answer conditions for the exercise.
Related articles and resources
- How to choose generative AI training for corporations
- Preparation for AI training for president
- IPA's AI human resource development information
https://www.ipa.go.jp/digital/ai/hr.html
If you would like to design training assignments tailored to your company's business, please organize the target audience, the work you want to handle, and the range of available data and Contact us.
Information on training and introduction support
If you are the president and would like to try automating your work, please see Contents and prices of AI training for presidents. This is a training program in which we listen to your business and requests, create something, try it out, and make improvements.
Contents and prices of AI training for presidentshttps://ceo-claude-code-kennsyuu.onrender.com/
When considering what to automate with Grokbot, you can specify the work you want to delegate from Grokguide usage examples/prompts/how to make.
Grokguide usage examples/prompts/how to makehttps://grokguide.jp/
If you would like to consult with us about building a system tailored to your business, please check Details of AI agent construction support.
Details of AI agent construction supporthttps://zettai-ai-agent-site-render.onrender.com/