- Your staff already use AI. A ban only moves it to personal phones the company cannot see.
- A policy people follow fits on one page, has ten sections and arrives together with tools the company has approved.
- Personal data gets a class of its own, because under the PDPA the company stays responsible after the data leaves.
- Name one owner, give people a way to report mistakes, and book a review every six months.
A ban leaves the company blind
Sales asks AI to draft replies to customers. Accounting asks it to summarize an Excel file. Most of this happens on free accounts opened with personal email addresses. AI use the company has not approved is called shadow AI, and it goes on even in companies that have announced a ban.
A ban does not stop anyone. People keep using AI on their own phones, and when customer data leaks, the company is the last to find out. A policy that works says plainly which tools to use, with what data, and whom to ask. If you have not yet chosen between banning AI, buying business accounts and running AI in-house, read When employees paste customer data into ChatGPT first.
One page, ten sections
The policy people remember is one page long. The tool list and the data-class table go in appendices you can update without reissuing the policy. The sample sentences below can be used as written.
1. Purpose and scope: "This policy applies to employees, contractors and interns who use AI for company work, including on personal phones." 2. Approved tools and accounts: "Use only the tools listed in the appendix, signed in with a company account. Request new tools through IT." 3. Data classes: "Data falls into four classes: public, internal, personal and confidential. What may go where follows the appendix table. When unsure, treat data as the higher class." 4. Customer and employee personal data: "Remove names, phone numbers and ID numbers the task does not need. Sensitive data, such as health information, needs the DPO's approval first." 5. Checking AI output: "AI output is a draft. Whoever uses it checks it against the source before it goes out, and is responsible for it as if they had written it." 6. Telling customers: "Customers talking to an automated system are told so in the first message and can ask for a person." 7. Ownership and records: "Work done with AI falls under your employment contract like any other work. Prompts and output for client deliverables are kept for as long as that document is kept." 8. Incidents and who to call: "If data goes into the wrong tool, report it to [name and channel] the same day. Nobody is disciplined for reporting." 9. Training and acknowledgement: "Everyone completes one hour of training and acknowledges this policy before receiving a company AI account." 10. Six-month review: "[Policy owner] reviews this policy every six months, or sooner if the law or a vendor's terms change."
What the PDPA asks when staff paste customer data into AI
For the customer data it collects, the company is the data controller under Thailand's Personal Data Protection Act of 2019 (B.E. 2562), the PDPA. When an employee pastes a customer list into an AI tool, the data goes to an outside provider. The responsibility stays with the company.
In October 2026, Thailand's Office of the Personal Data Protection Committee (PDPC) published guidelines on protecting personal data in AI. They are not legally binding. Read together with the PDPA, they point to four things your policy has to cover.
- Prompts, usage logs and AI answers can be personal data if a person can still be identified from them. - A provider that trains models on customer data for its own purposes may become a second data controller. - If the provider stores data outside Thailand, the destination country needs adequate protection standards, unless a legal exception applies. - After a breach, the company notifies the PDPC without delay and, where feasible, within 72 hours of becoming aware of it, unless the breach is unlikely to put people's rights at risk. If the risk is high, the people affected are told as well.
Thailand's draft AI law is still being revised, and the generative AI governance guideline for organizations from the Ministry of Digital Economy and Society and ETDA dates from late 2024. So build your policy on the PDPA, which is in force today. Rules abroad are covered in The AI rules on the way. This chapter is practical guidance and is not legal advice. Have your legal team or DPO read the policy before you issue it.
Hypothetical case
An appliance retailer in Khon Kaen that sells on installment (names and details invented). The managing director once told the staff LINE group that AI was banned for company work. One day a collections officer pasted two hundred overdue customers, with phone numbers and balances, into a free AI chatbot on her own phone and had it draft a reminder for each one. Then she mentioned it, proudly, in a meeting.
The HR manager, who also handles PDPA matters, gathered the facts, deleted the chats and worked out with the company's legal adviser whether the PDPC had to be notified. Everyone agreed on the cause: the officer had no proper tool and did not know whom to ask. The next week the company wrote a one-page policy, bought business accounts for the collections team and had AI draft one template message that never sees a customer's name. The company's own system fills in names and amounts.
Four classes of data, and where each may go
Staff do not need to memorize the law. They need to know which class the data in front of them belongs to. Many companies already have a data classification scheme. The AI policy adds one change: personal data becomes its own class, because the PDPA has rules specific to it and it turns up in daily work, in customers' LINE chats and job applications for example.
The four-class table: where data may go
- Question 1 Does it contain a name, phone number, ID number or anything else that points to a person? If yes, it is personal data.
- Question 2 If it leaked, would the company breach a contract or lose an advantage? If yes, it is confidential.
- Question 3 Has it already been published? If yes, it is public. Everything else is internal.
When unsure: go up one class. If both question 1 and question 2 apply, follow the confidential rules.
Strip before you paste: replace names with codes such as "Customer A". Once the data can no longer be linked to a person, it drops to internal.
| Data class | Tools allowed | Conditions | Examples |
|---|---|---|---|
| Public | Any AI, free accounts included | Check the output before use | Product details on the website, news already announced |
| Internal | Approved business accounts | Company sign-in only | Internal procedures, meeting notes without customer names |
| Personal | Business accounts approved for this class, with a contract that rules out training on the data | Use only for the purpose people were told about; remove what the task does not need; sensitive data needs DPO approval | Customer chats on LINE OA, job applications, debtor lists |
| Confidential | Private LLM or in-house systems | The data owner approves each use; usage is logged | Production formulas, contracts under NDA, financials before release, source code |
When someone slips: who does what, and by when
The incident to plan for is data of the wrong class going into the wrong tool. The reporting route has to be short enough that people report the same day, because the PDPA's 72-hour window starts when the company becomes aware.
1. The person who slipped, or whoever noticed, tells the policy owner the same day. 2. Record the incident: what data, how many people, which tool, whose account. 3. Delete the chat if you can, revoke shared links, and change the password if an account was exposed. 4. The DPO decides whether the PDPC or the people affected must be notified, inside the 72-hour window. 5. Fix the cause and add the case to training.
The six-month review checks whether the tool list matches what people actually use, whether vendors have changed their terms, and whether a new law applies.
One week from draft to launch
A policy that takes three months to write is often out of date the day it is issued. Aim the first week at a version people can use, and improve it at each review.
- Monday: run an anonymous survey of which AI tools each department uses, and for what. - Tuesday: choose tools and open business accounts. ChatGPT Enterprise, Microsoft 365 Copilot and Gemini in Google Workspace state that they do not train models on business customers' data without permission, but read the terms of the plan you actually buy. - Wednesday: draft the one page and the four-class table, and have legal or the DPO read them. - Thursday: train each department for one hour, using its own real work. - Friday: issue the policy, have everyone acknowledge it, and put the first review date in the calendar.
Three ways a policy fails
The first is a total ban: people keep using AI where the company cannot see. The second is a policy without approved tools, so staff drift back to free accounts. The third is a policy with no owner, so nobody answers questions or runs the review. Before you issue yours, check the draft against this list.
- The scope covers contractors and personal phones
- There is a tool list and a way to request new tools
- The four-class table uses your own company's examples
- Someone is named to approve sensitive data
- It says whoever uses AI output is responsible for it
- It says when customers must be told
- It says which records are kept, and for how long
- It names who takes incident reports and promises not to punish whoever reports
- Everyone acknowledges it before getting an account
- It has a policy owner and the next review date
For the development team · SSO, DLP, logging and private endpoints
SSO: approved tools sign in through the company identity provider, such as Microsoft Entra ID or Google Workspace, so an account closes the day someone leaves; tenant restrictions at the proxy keep personal accounts out. DLP: add Thai data patterns (the 13-digit national ID with its checksum, mobile numbers starting 06, 08 or 09, bank account numbers) and warn or block when they are pasted into unapproved AI domains. Logging: send the business accounts' audit log to the SIEM, set a retention period and limit who can read it, because logs that contain prompts can be personal data themselves. Private endpoints: route API calls to models through a private endpoint in a region you choose, or through an internal gateway that strips PII before anything is sent. Confidential data goes to a private LLM on company hardware.
A policy written in the training room, and the systems that make it stick
The policies people follow are usually written by the people who have to live with them. For DNA Maker's in-house AI training, see In-house AI training for companies.
When the rules need systems behind them, our FDE team works inside yours to set up role-based access, audit logs and PDPA data handling on the tools you already pay for. For confidential data, we deploy a private LLM in your own environment. Everyone on the project is DNA Maker's own team; nothing is subcontracted. Our 500+ projects include work for banks, hospitals and large Thai groups. Talk to us on the AI Transformation page or ask Dream in the chat.
SOFTWARE ENGINEERING GLOSSARY
Software engineering glossary
| Term | What it is | A simple example | What executives should ask the development team |
|---|---|---|---|
| Shadow AI | AI used for work without the company's approval | A free chatbot on a personal phone | Do we know which AI tools each department uses? |
| Data classification | Sorting data into classes by the harm a leak would cause | Keeping announced news apart from contracts under NDA | Who decides when nobody is sure which class the data is in? |
| PII | Data that identifies a person, directly or indirectly | A debtor's name, phone number and balance | Can our systems strip it out before anything goes to AI? |
| DLP | A system that stops important data from leaving the company | A warning when someone pastes an ID number into an AI website | Does our DLP recognize Thai data formats yet? |
| Audit log | A record of who did what, when, with which data | Seeing who sent a customer file to AI last month | How long are logs kept, and who can read them? |
Tilleke & Gibbins, Thailand Issues Personal Data Protection Guidelines for AI Systems
IAPP, Key developments in Thailand's PDPA regulations
Multilaw, Data Protection Laws Guide: Thailand
Digital Policy Alert, MDES and ETDA Generative AI Governance Guideline for Organisations
Bangkok Biz News (Thai), progress of Thailand's draft AI law
Tilleke & Gibbins, Thailand Resumes Development of AI Regulatory Framework
Microsoft Learn, Data, privacy and security for Microsoft 365 Copilot
OpenAI, Enterprise privacy
Google Workspace, Generative AI in Google Workspace Privacy Hub