ChatGPT Business Usage Guide: Steps for Japanese and Thai Operations at Thailand Offices

ChatGPT is a conversational generative AI (Generative AI) service that uses a large language model (LLM) provided by OpenAI, capable of handling tasks such as document creation, summarization, and question answering in a natural dialogue format. This article is aimed at IT and operations staff at companies based in Thailand, and explains use cases for generative AI in business operations—such as drafting emails and meeting minutes in Japanese and Thai, organizing quotation and delivery-date inquiries, and handling first-line FAQ responses—as well as the introduction process, including how to address Thailand's PDPA and erroneous outputs, and how to verify effectiveness through trial use. After reading this article, you will be able to determine which of your company's business operations to start applying ChatGPT to, and which management rules should be established first.

The decision on introduction depends on whether the handling of personal data falls under the scope of PDPA (Thailand's Personal Data Protection Act) regulations, and on how well the company's internal information management system is established. In addition to these two points, if you decide what to measure through trial use before incorporating it into business operations, you can reduce the effort of later having to revisit access scope and prohibited tasks.

When reviewing information security policy, first decide who can access ChatGPT and which internal company data it can connect to. ChatGPT Enterprise provides single sign-on via SAML and an admin console, allowing administrators to control access rights for each user and whether connectors can be used. For connectors that connect to external services such as SharePoint and Google Drive, administrators should configure settings so that only individually approved scopes are available to employees. When integrating with custom-developed systems, the connection scope can be limited via connectors using MCP (Model Context Protocol).

However, there is no need to prepare all connectors and automated integrations from the start. First, try providing approved input via Excel forms or data exported from business systems as files, and consider integration only for tasks where this manual handoff proves burdensome each time—this keeps the connection scope limited to what is actually needed. Decisions about what information may be entered and with whom outputs may be shared should not be left to the using department's discretion, but approved by the person responsible for information management.

At the same time, the company should internally confirm data retention periods and deletion settings, the regions where data is stored and processed, whether a Zero Data Retention (ZDR) option is needed, and whether a DPA (Data Processing Agreement) can be concluded, and should also decide which department will audit the logs. If these settings are postponed, reviewing the connection scope and access privileges after use has begun will become necessary, increasing the cost of revising operational rules.

When deciding the scope of tasks eligible for use, divide them into three levels based on the impact if an error occurs. For tasks such as drafting emails or summarizing meeting minutes—where errors can be corrected by a person and stopped before sending—ChatGPT's output can be used as-is as a draft. On the other hand, for high-risk tasks such as decisions involving medical or legal matters, or finalizing contract terms, where incorrect output directly affects customers or internal decision-making, OpenAI also advises that these should be treated as areas requiring mandatory human review and additional verification. First-line responses to customers and FAQ handling, which fall in between, should be designated as permitted tasks on the premise that a human will always check the output before sending it. When determining scope, each department should create a list dividing tasks into three categories—"tasks where ChatGPT's output can be used as-is," "tasks requiring mandatory human review," and "tasks where use is prohibited"—and communicate this alongside the management settings of the internal security policy. If this categorization is postponed, there is a risk that staff on the ground will extend use to high-risk tasks based on their own judgment.

Once permitted tasks have been decided, before rolling them out company-wide, have a small group of people responsible for that task try it out and verify its effectiveness compared with the conventional process. Even if ChatGPT can produce drafts faster, if checking and correction take time, the overall workload will not be lightened. The time measured should be the total time including preparation, checking, and correction—not just the time taken for generation.

Before the trial, decide on the target task, the number of cases, the period, the person responsible for evaluation, and the conditions for considering it a pass. If these conditions are decided after seeing the results, there is a risk of judging based only on cases that happened to go well. The indicators to check and the evaluation methods are divided as follows, depending on the type of task.

The test cases should also include requests with missing information and requests involving exceptional conditions. If a task passes with normal cases but produces errors with exceptions, keep it in the category requiring mandatory human review. If the pass conditions are not met, record whether the cause was how instructions were given, insufficient source materials, or poor fit with the task, and decide whether to retry or exclude it from scope.

Document creation is a task where errors can be caught by a person before sending, making it the easiest area to start with when introducing ChatGPT. The speed of draft creation and the accuracy of text organization between Thai and Japanese depend on how granular the prompt instructions are and whether a final check is performed.

For drafting emails and reports, providing the recipient, purpose, and any required numbers or proper nouns in advance as bullet points, along with specifying the writing style (polite/concise, internal/external), reduces the back-and-forth of revisions. The official prompt pack demonstrates use cases such as extracting line items from invoices and classifying them—extracting structured information from business documents—which can also be applied to compiling the numerical sections of reports.

Subject: Please draft an apology email to a business partner regarding a payment delay.
Conditions:
- Recipient: Contact person at the business partner
- Reason for delay: Delay in internal approval procedures
- New scheduled payment date: Next Friday
- Writing style: Polite but concise, apology limited to one sentence at the beginning

The output should not be sent as-is; a person must cross-check specific information such as amounts, dates, and company names against the original text before use. Saving the summary alongside the original text makes it easier to verify the content later. For reports, having drafts created in bullet points by item, and explicitly specifying the unit of figures (baht or Japanese yen) and the aggregation period, helps prevent mix-ups. It is advisable to differentiate the strictness of checks between emails sent externally and reports for internal use, and to establish a rule that a supervisor or responsible person must visually confirm external emails before sending, which helps minimize the risk of sending errors.

In text organization work involving Thai and Japanese, ChatGPT can handle text in either language, so translation and summarizing of key points can be delegated to it. However, nuances in the level of politeness and company-specific phrasing are prone to shifting between languages—the formality of "desu/masu" style may be converted into an overly stiff expression in Thai, or conversely, into an overly casual one. In practice, it is easier to prevent loss of meaning by using a two-step instruction: first have the key points summarized in bullet points in either Japanese or Thai, and then have them refined into the target language.

For proper nouns such as personal names, company names, and department names, instructions must explicitly specify that the original-language notation should be preserved as-is, since the model may paraphrase proper nouns or alter their notation. For internal technical terms such as product code naming conventions or defect names, it is advisable to compile a Japanese-Thai glossary and attach the approved version to the instructions, which prevents the same term from being translated differently across different documents. A person should be assigned to maintain the glossary, and a revision date should be recorded whenever a translated term is changed. The organized text must always be visually checked by either a native Thai-speaking employee or an employee who can accurately read Japanese, and particularly for documents involving amounts, dates, and contract terms, the practice of cross-checking against the original text before sending must be strictly enforced.

For inquiry handling, the information sources to connect and the strictness of verification required differ depending on whether the inquiry is internal or external. The key distinction is that internal FAQs are based on internal documents authorized by an administrator, whereas for customer support, the information sources referenced and the design of the verification process differ depending on each individual inquiry.

An internal FAQ chatbot is built using the Company Knowledge feature by designating internal documents as the reference source. An administrator connects documents such as HR regulations, employment rules, and operations manuals to the connector via SharePoint or Google Drive, and only the synced documents are permitted as targets for search and citation. When an employee asks a question such as "What is the procedure for applying for paid leave?", an answer is generated based on the relevant section of the connected documents.

The scope of connection should be divided by department and configured in conjunction with viewing permissions—for example, ensuring that accounting department regulations are not accessible to all employees. When a document is revised, the sync timing should be checked, and a person should be assigned to periodically verify that answers are not being based on an outdated version. Configuring the system to display the document name and relevant section in the answer makes it easier for employees to verify the original text, and also facilitates smooth reporting when an incorrect citation is noticed.

In drafting first-response replies for customer inquiries, make it a strict rule to have ChatGPT produce a draft of the reply text, which a person then verifies and edits before sending.

For inquiries with many conditions, such as quotation requests or delivery date inquiries, have ChatGPT list out the conditions before drafting the reply text. Provide the approved request email, quotation, and delivery date response, and have it organize the part number, quantity, delivery date, and amount, mapping each to the document in which it was stated. Then, having it extract into a separate list the items with no record and the conditions that conflict between documents lets the person in charge know what points need to be checked before replying. Draft the Japanese and Thai reply proposals only after reflecting these items to be confirmed. The same procedure can be used for purchasing tasks that involve inquiring with suppliers about delivery dates.

For generating first responses and FAQ chatbot answers, aligning the scope of the documents used as the basis helps prevent discrepancies in content. For responses involving pricing, contract terms, or returns and warranties, never remove the operational step of having the person in charge cross-check against the original regulations before sending. Skipping this check risks guiding customers based on outdated regulations or incorrect conditions.

Information organization and analysis is a task where, the faster summarization and aggregation become, the easier it is to skip cross-checking against the original source materials. Whether handling meeting minutes or numerical data, whether the output is structured so that results can be traced back to and verified against the original materials or data is the dividing point that determines the burden of the verification work.

When summarizing meeting minutes, if you pass audio transcriptions or handwritten notes to ChatGPT and have it extract the three items of decisions made, person in charge, and deadline separately, the burden on the person reviewing it later is reduced. Having it summarize a long transcript of remarks as-is can omit the speaker's intent, so it is necessary to add an instruction that explicitly specifies "who decided what."

OpenAI's safety best practices recommend, even for summarization tasks, retaining references to the original source material and having a person verify it. For meeting minutes, have it append the relevant portion of the remarks and the speaker's name to the end of the summary text, and save it in a form that can later be cross-checked against the original text. In particular, for decisions involving numerical targets or budgets, rather than sharing the summary alone, establishing a procedure where the person in charge checks the relevant part of the original text before internal distribution prevents mishearing or transcription errors from propagating to later processes.

When a meeting proceeds in multiple languages, having the language recorded for each remark prevents variations in expression between Thai and Japanese from becoming mixed during the summarization stage, making the verification work easier. The granularity of summaries should be standardized by meeting type, with the level of detail required for regular meetings and decision-making meetings kept separate, so that the person in charge does not have to decide on the verification standard each time.

When handling numerical data such as sales, inventory, or inspection records, rather than leaving the aggregation entirely to ChatGPT, keep a procedure in place where the same results can be recalculated in Excel or similar tools. By assigning ChatGPT the role of comparing aggregated results with each other and drafting explanations for points with large changes, numerical errors become easier to notice.

As preparation, align item names and units, such as part numbers, processes, and defect names, into a form without inconsistent notation. If the same defect is recorded under different names depending on the person in charge, the count gets scattered and trends become invisible. For quality comparisons, look not just at the number of defects but at the occurrence rate, with the number inspected as the denominator. Since part numbers with higher production volumes tend to have higher counts, ranking by count alone leads to mistaken prioritization of countermeasures. When looking at the gap between plan and actual results, have it list the processes and periods where discrepancies occurred, and organize the reasons written in daily reports alongside them.

Use the output aggregated values only after cross-checking them against the original data and existing aggregation tables. Judgments on causes, the validity of countermeasures, and approval of formal reports should be made by the quality or production department in charge, checked against on-site facts, with ChatGPT's explanations treated as material for narrowing down the points that need to be checked.

Failures during implementation occur when either of the two operational rules—input data management and output verification—is missing. The cause is not insufficient tool performance, but rather that use began without defining the procedures that should have been determined internally.

In OpenAI's FAQ for the consumer-facing ChatGPT, it is stated that input data may be used to improve the model. If employees paste customer names, contract amounts, personal medical histories, and other data directly into the consumer version to request summaries or translations, confidential information ends up being transferred to an external system in a way that administrators cannot track. In contrast, ChatGPT Enterprise explicitly states a default policy of not using enterprise data for model training, and ownership of both inputs and outputs remains with the customer. The steps to avoid this issue are as follows.

1. Limit business use to workspaces under an enterprise contract, and prohibit use of personal accounts for work purposes through internal company rules
2. When handling customers' personal data, first determine whether it falls under the PDPA (Thailand's Personal Data Protection Act), and if so, complete anonymization or pseudonymization before entering it
3. As needed, check with the administrative department whether a DPA (Data Processing Agreement) or Zero Data Retention (ZDR) option is required

If these steps are neglected, confidential information cannot be retrieved once it has been passed to the external consumer version.

Text generated by ChatGPT may include hallucinations—content not based on fact, output in a natural writing style. OpenAI itself officially acknowledges that the model may generate misinformation, and advises against automatic application—or recommends requiring human verification—in high-impact areas such as medicine and law, where errors directly affect decision-making. In practice, common issues include distributing numbers or proper nouns from a meeting minutes summary internally without checking them, or sending FAQ answers to customers that still contain outdated conditions. The steps to avoid this are as follows.

1. In advance, distinguish between tasks where the output can be used as-is and tasks that require mandatory human verification
2. For numbers, dates, and proper nouns, have the responsible staff member cross-check the generated results against the original text or internal documents connected via a connector
3. Have the approver record whether verification has been completed, and establish a system so that unverified output does not flow into the next step

If this verification is skipped, incorrect numbers or conditions may be used as-is in customer interactions or internal decisions.

By combining the three tasks of document creation, inquiry response, and information organization, ChatGPT can also be applied to work that consolidates documents from multiple departments—such as manuals and training materials—into a single piece of material. The broader the scope of application, the more likely it is that the question of which department's approval is needed before internal publication becomes unclear, which is a key point to watch.

In creating business manuals, the main use case involves handing ChatGPT existing work procedure notes or transcribed verbal explanations and having it organize them into numbered steps. Official prompt pack examples demonstrate a method of extracting and classifying items from business documents, and this can also be applied to organizing scattered procedure notes into stages such as "Preparation," "Execution," and "Verification" to form a manual draft. When multiple departments perform the same task, it is advisable to first lay out each department's procedures, responsibilities, judgment criteria, and forms used before having a draft created, and to decide which parts should be standardized and which exceptions should remain department-specific. This prevents the draft from being biased toward one particular department's way of doing things.

Unlike meeting minutes summaries, manuals are actually used on-site once completed, which changes the weight of the verification process. The generated draft must be checked by employees who actually perform the task to confirm they can work according to the procedures, and the department responsible for the process must approve the content before it is published internally—this flow must be strictly followed. The approved version should record the person in charge of management, the version number, and the revision date, and when procedures are changed, a revision history must be kept before replacing the document. When creating both Japanese and Thai versions, the approved version should be expanded into the other language using terms from the glossary, with the version numbers of both aligned. Since gaps in steps or errors in sequence often stem from insufficient information at the transcription stage, it is also essential to confirm that the original procedure notes contain sufficient content before creating the draft.

In creating internal training content, ChatGPT can be used to draft key-point explanations, practice questions, and comprehension tests for new employees, based on business manuals and past training materials connected via Company Knowledge. While creating business manuals was mainly about organizing procedures, training content differs in that it requires clarity of explanation that learners can understand, as well as natural expression between Thai and Japanese.

When the training target is Thai staff, generating content directly from Japanese materials can result in unnatural translations of honorific language or company-specific phrasing. Therefore, a two-step instruction approach is effective: first have the key points organized into bullet points, then have them translated into Thai in line with the glossary's terminology. The generated practice questions and explanations should be checked for accuracy and appropriate difficulty level by the employee actually responsible for conducting the training or by the HR department, and distributed only after approval. If this check is skipped, incorrect business knowledge may spread through the training.

Whether to adopt ChatGPT depends less on the tool's mechanics and more on how clearly you have defined your procedures for handling personal data in Thailand and for checking its outputs. Here, we answer two questions: what kind of tool ChatGPT is, and whether it can be used for Thai-language work.

ChatGPT is a conversational generative AI (Generative AI) service built on a large language model (LLM) developed by OpenAI. It is characterized by its ability to handle writing, summarization, translation assistance, and question answering through natural-language conversation. For businesses, ChatGPT Enterprise is available, offering management features such as single sign-on via SAML/OIDC, an admin console, and integration with internal documents through Company Knowledge. Because the free/individual version and the enterprise version differ in how input data is handled and in access management mechanisms, it is necessary to first understand the differences in contract types before using it for business purposes.

Yes, it can be used. ChatGPT can also write and summarize text in Thai. However, since there can be variation in the strength of honorific expressions and in how proper nouns are translated, visual review by a native Thai-speaking employee or an employee who can accurately read Japanese should always be carried out—especially for documents involving monetary amounts or contract terms.

Introducing ChatGPT at a Thailand office becomes easier the more clearly the risk classification of tasks and the verification framework have already been decided. Regardless of whether you start with document creation, inquiry response, or information organization, deciding on the order of implementation through the following four stages will help avoid confusion.

1. Check whether personal data subject to PDPA is involved, and for applicable tasks, decide on anonymization or pseudonymization procedures in advance
2. Have administrators set internal access permissions and the scope of connector connections, avoiding business use via personal accounts
3. Classify tasks by department into those where output can be used as-is and those requiring a human to cross-check against the original text, and designate approvers in advance
4. Trial the approved tasks with a small group, then expand the scope only after comparing the total time—including checking and correction—and the number of errors against the previous process

Satisfying just one of these is not enough; some tasks, such as drafting initial responses to inquiries containing customers' personal information, require both a PDPA judgment and output verification. By following these four stages, you can start with low-risk tasks such as drafting emails or summarizing meeting minutes, while safely expanding into tasks involving customer interactions or internal regulations—keeping human verification in place. If you are moving on to the stage of establishing a management framework in-house, also check AI Consulting Thailand/Bangkok | 2026 Implementation Guide, which will help clarify the operational items you should decide on next.