The core answer
A company may use AI to help screen applications or evaluate performance, provided that a person with real authority makes the decision and the company can show on what reasons it was made. As at 21 November 2026 we found no Thai law that directly prohibits AI in selection or evaluation. That does not mean there is no risk, because applicant and employee data fall under the PDPA and the treatment of employees falls under the Labour Protection Act.
The risks we see most often are four: the system gives wrong results and nobody notices; the system leans against certain groups; the system uses more data than was disclosed, or uses sensitive data without a lawful basis; and the person affected has no way to ask or contest. This guide takes them one at a time and says what documents to keep.
The main caution is this: do not let the system's score stand in for the decision as a whole. A score should be one input that a person reads, and that person then writes their own reasons.
Who this guide is for
| Your situation | Sections to read first |
|---|---|
| Setting policy before anything happens (choosing a system or just starting to use one) | Sections 1, 2, 3, 4, 5 and the document checklist at the end |
| Something has happened with one employee or applicant (rejected, scored down, or asking how the system decided) | Sections 6, 7, 8, 9 |
| A complaint or dispute already exists (a claim, a threatened claim or a complaint to an authority) | Sections 8, 9, 10 and the documents to keep |
1. What does Thai law say about using AI to screen people and evaluate performance?
We found no Thai law dedicated to AI in personnel matters, as at 21 November 2026. The position may change, so check the latest announcements before you decide. What applies today is the existing law.
| Issue | Law or guidance that bears on it |
|---|---|
| Applicant and employee data | PDPA: lawful bases in section 24, sensitive data in section 26, data subject rights in sections 30, 32, 35 and 36 |
| Discrimination in employment | Section 15 of the Labour Protection Act requires the employer to treat male and female employees equally in employment, unless the nature or condition of the work does not allow it. The section speaks of sex and does not cover every kind of bias |
| Discipline and dismissal | Labour law and the company's work rules |
| AI practice | ETDA guidance, which is not binding |
ETDA, with the Ministry of Digital Economy and Society and its AI Governance Center (AIGC), announced the "Guidelines for the Governance-Based Application of Generative AI in Organisations" on 30 October 2024. It is guidance for executives and relevant staff in organisations and is not binding law. It has five parts: understanding Generative AI; benefits and limitations; risks and how to manage them; application guidance and structures; and considerations for responsible use. Its central idea is to balance benefits against risk management and to involve stakeholders. Selection and evaluation directly affect people's rights, so treat them as a higher-care category than ordinary office use.
For internal AI use generally, such as staff pasting customer data into a chatbot, see company AI policy and customer data. This guide covers only the use of AI to decide things about people.
2. What kind of system is it, and what does it decide for us?
Before the legal points, know what the system does, because the risk differs a lot.
| Use | Example | Level of care |
|---|---|---|
| Sorting or summarising | Summarising CVs, grouping applications by keyword | Low to medium, if a person still opens the applications |
| Scoring or ranking | Scoring fit for the role, ranking applicants | Medium to high, because the score carries weight in the decision |
| Automatic rejection | The system rejects applications below a threshold before anyone sees them | High; the rejected person was never considered by a human |
| Analysing biometric or emotional signals | Analysing face, voice or posture in video interviews | Very high; touches biometric data, which is sensitive data |
| Evaluating employee performance | Pulling sales figures, chat response times or clicks to produce a monthly score | Medium to high, depending on whether the result feeds pay or dismissal |
Before real use, make an inventory: where AI is used in personnel work, what data goes in, who receives the output and who presses the final button. Many companies do not know that the recruiting software or time-keeping system they already use has a hidden AI screening or scoring function.
3. Is the system accurate enough, and how can the company show it?
AI can be wrong. It may fail to read an unusual CV layout, misread a word out of context, or score down someone from a different field whose skills actually match. Section 35 of the PDPA requires the controller to keep personal data accurate, up to date, complete and not misleading. Information a system summarises or generates about a person should be held to the same standard.
What an SME can realistically do:
- Ask the vendor for documents on what the system measures, what data it uses and whether it has been tested on work like yours.
- Trial it on 30 to 50 past applications whose outcome HR already knows, and compare whether the system picks the people HR picked. Record the comparison.
- Set how often to re-check, for example each quarter or whenever the vendor updates the system.
- Let reviewers record clearly wrong outputs when they find them, and keep those records.
The company does not need to audit the model technically itself, but it needs evidence that before and during use a person checked whether the results were reasonable, and what was done when errors were found.
4. How should the company check that the system does not disadvantage certain groups?
Bias in an AI system rarely comes from an instruction to discriminate. It comes from the data used to train it. If the company has historically hired only men for a role, a system that learns from hiring history may learn that "this is what a good candidate looks like". Neutral-looking variables such as a gap in employment, the name of a school or an address can stand in for sex, age or background.
Points under Thai law:
- Section 15 of the Labour Protection Act requires equal treatment of male and female employees in employment, unless the nature or condition of the work does not allow it. If the system screens out one sex at an unusual rate, the company should ask whether the nature of the job really explains it.
- Thai law has no clear general provision on discrimination by age, religion or ethnicity in private-sector recruitment, so do not conclude that something is acceptable merely because no law says no. Consider the facts, the dispute risk and the reputational risk.
- If the system uses or infers sensitive data, such as disability, health or biometric data, section 26 applies and in principle requires explicit consent unless a statutory exception applies.
Checks the company can do:
- Compare pass rates by group, such as sex and age band, if the company holds such data lawfully.
- Remove job-irrelevant variables from the input, such as photo, age and marital status.
- Ask the vendor whether it tests for bias and get a written summary.
- Have a person sample the rejected applications from time to time.
5. Personal data: what can be used and what must be disclosed?
This builds on PDPA and employee data in HR, which explains lawful bases and general privacy notices. For AI, ask five further questions.
- What is the lawful basis? Section 24 prohibits collecting personal data without consent unless a statutory basis applies, for example necessity to take steps at the data subject's request before entering a contract, or the controller's legitimate interests, which cannot be relied on where they are overridden by the data subject's fundamental rights. Consent from an applicant may be questioned as not freely given because bargaining power is unequal, so do not rely on consent alone without thought.
- Is any sensitive data involved? Section 26 protects health, disability, biometric data, criminal records and other listed data, and in principle requires explicit consent unless an exception applies. A system that analyses face or voice in video interviews can involve biometric data, so check with the vendor before switching it on.
- Have applicants and employees been told? The privacy notice should say that an automated system helps to screen or evaluate, what data it uses, who receives the result and how long it is kept.
- Where does the vendor keep the data? If the vendor is a processor, there must be an agreement with the company. If the servers are abroad, check the conditions for cross-border transfer. See company AI policy and customer data for details.
- Does the vendor train its model on our data? Read the contract and switch it off where possible.
Where a system measures staff, for example face scanning or location tracking, see face scan, fingerprint and GPS monitoring of employees.
6. Who actually decides, and what makes human review real?
Writing "a person reviews" in a policy is not enough. If in practice the reviewer clicks approve on every system result without reading, a court or authority is likely to see a decision made by the system.
What a meaningful review needs:
- Real authority. The reviewer can change the system's result and is not pressured by targets to agree with it.
- Enough information. The reviewer sees the underlying data, such as the full application, not only the score.
- Enough time. If hundreds of applications must be cleared per hour, the review means nothing.
- Enough understanding. The reviewer knows what the system measures and its limits, through short training.
- Own written reasons. When rejecting an applicant or giving a low evaluation, the decision-maker writes their own reasons and does not copy the system's explanation.
- A named person. The record shows who approved, when and on what data.
A sound policy line: the AI system provides information, a named person decides, and that person must be ready to explain the decision in their own reasons.
7. How can applicants and employees contest a result?
Thai law has no dedicated provision on a right to contest automated decisions. But the PDPA rights apply to the data the system uses and produces:
- Access and copy (section 30). The company must act without delay and within thirty days of the request. It may refuse only where a law or court order requires, or where it would harm the rights of others, and must record the refusal with reasons.
- Correction (sections 35 and 36). If the company does not act on a request, it must record the request and the reasons.
- Objection (section 32). This applies to collection or use of data based on section 24(4) or (5), and the other cases that section lists. The company need not stop if it shows a compelling legitimate ground or that the data is needed to establish, exercise or defend a legal claim.
Beyond the statutory rights, the company should have a clear internal channel:
- A contact point (email or form) stated in the privacy notice and in the outcome email.
- A recipient who is not the same person as the original decision-maker.
- A clear response period, for example 15 working days (an internal policy period set by the company, not a statutory one).
- A record of the request, the outcome and the reasons.
What remains uncertain: Thai law does not state expressly that the company must disclose the system's logic or formula. Agree with the vendor in advance what information can be supplied to explain a result when asked.
8. What evidence should be kept so the decision can be explained later?
In a dispute, the first question is "what was the decision based on at the time?" If the company cannot answer, it is at a disadvantage even if the decision was reasonable.
| Evidence | What to keep |
|---|---|
| System register | Name of system, version, start date, the work it is used for and the data fed in |
| Vendor documents | Manual, accuracy and bias test results, contract and data-use terms |
| Criteria | The criteria for the role written before applications opened, with the weight of each |
| Per-person record | The system's output, the reviewer, date and time, the reviewer's own reasons and any change to the system result |
| Test and re-check results | Date, method, outcome and what was fixed |
| Requests and replies | The applicant's or employee's request, date received, recipient, outcome and date of reply |
| Training | Names of reviewers who completed training on the system |
Retention: set by data type and purpose. Unsuccessful applicants' data should not be kept longer than needed. See retaining contracts, emails, chats and employee records for the approach to employee records.
9. What if an AI evaluation is going to lead to reduced pay or dismissal?
A system result is one piece of information. Discipline or dismissal must still follow labour law and the company's work rules, and a court looks at the facts of each case. What to do:
- Separate clearly whether this is about performance below standard or about a disciplinary offence, because the legal consequences differ.
- Check that the metrics the system uses match the work the employee was actually assigned, and that the employee knew them in advance.
- Tell the employee the result and the data used, and let them explain before deciding.
- Have a supervisor who knows the real work add an opinion; do not rely on figures alone.
- Consult counsel before dismissal, because severance and the fairness of the dismissal depend on the cause and the evidence. For the overview see employee rights on dismissal.
If the system scores from behavioural monitoring, such as screen time or mouse movement, check against the PDPA whether it is necessary and was disclosed.
10. What should the contract with the AI vendor say?
Responsibility to applicants and employees stays with the company. Pointing to an outside vendor does not release it from responsibility to the people affected. So write into the contract:
- Scope: what the system does and does not do, and the data it uses.
- Accuracy and testing: the vendor provides test results and known limits, and tells you about updates that change results.
- Personal data: processor status, a ban on training the model on your data without permission, where data is stored, how long, and deletion at the end of the contract.
- Support with requests: the vendor helps retrieve data an applicant or employee asks for, within a time the company can meet.
- System logs: kept so the company can check results afterwards.
- Liability and indemnity: who is responsible if the system breaches the agreement or data leaks.
11. Pre-use checklist and documents to have
- Inventory where AI is used in personnel work, the data used and who receives the output
- Set the lawful basis under section 24 and check for sensitive data under section 26
- Update the privacy notice for applicants and employees to mention automated systems
- Review the vendor contract: processor terms, model training, storage, deletion and help with requests
- Test the system on past data and record the result, including a bias check by group
- Write a policy that names who decides and requires their own written reasons
- Turn off automatic rejection, or at least have a person review the rejected group
- Train reviewers and keep the list
- Set up a channel for requests with a response period, separate from the original decision-maker
- Set retention for applicant data and decision records
- Set a re-check cycle (for example quarterly, or when the system updates)
- Before using results for discipline or dismissal, have counsel review the documents
Hypothetical example
Hypothetical example (not a real case; all figures are illustrative). A distribution company with 120 staff installs a recruiting system with an AI function that scores applications and is set to reject automatically anything below 60. For an accounting-clerk vacancy there are 200 applicants and the system rejects 140. One applicant, who has just returned to work after two years of childcare, scores 55 and is rejected. She emails to ask why and to see her data.
A company that did no preparation cannot say what the score was based on. Nobody has seen her application, and the vendor says it is "an internal formula" and will not provide data.
A better-prepared company would have set the system to "rank", not "reject automatically". HR would open applications below the threshold, at least by a sample set in the policy, and record who checked what and with what result. Before use it tested the system on past applications and found that an employment gap lowered scores, so it set that variable to have no effect. When the applicant asks, the company supplies the data it holds about her within the PDPA time frame, has a different person from the original reviewer re-examine her application, and replies with the decision-maker's reasons.
The difference lies in whether the company can explain the decision person by person, not in whether it uses AI. The legal outcome still depends on the actual facts and documents of each case.
Common misunderstandings
- "AI is neutral because it has no feelings." → The system learns from past data and can inherit past bias. Test it and check it regularly.
- "We bought the system from a large vendor, so the risk sits with them." → The company that decides to hire or discipline must still answer to the person affected. Liability between company and vendor is a matter of contract.
- "Someone clicked approve, so there was human review." → If the person had no information or time to read and never changes a result, it is no different from the system deciding.
- "The applicant ticked our terms on the website, so we can use any data." → Consent must be specific and freely given. Sensitive data needs explicit consent or an exception. A blanket tick-box may not be enough.
- "Thailand has no AI law, so anything goes." → We found no AI-specific law as at 21 November 2026, but the PDPA, labour law and civil law apply fully.
- "The system's score is enough evidence to dismiss." → A system result is supporting information. Dismissal needs a cause under labour law and evidence that can be explained.
- "The ETDA guideline is law, so following it makes us safe." → It is non-binding guidance. It helps you set up a sound system but does not remove liability under the law.
Further reading
- PDPA and employee data in HR — lawful bases, applicants and the privacy notice
- Fake qualifications or false work history — establishing facts, the right to explain and choosing a response
- Company AI policy and customer data — AI use inside the company and vendor contracts
- Face scan, fingerprint and GPS monitoring of employees — biometric data and employee monitoring
- Employee rights on dismissal — when an evaluation leads to dismissal
AI can help screen people or evaluate performance when a person with real authority decides and writes their own reasons, the system is tested for accuracy and bias, data use is disclosed under the PDPA, a channel to contest exists, and evidence is kept so the decision can be explained later.
We help review where the AI systems a company uses in personnel work carry risk, and draft AI-in-HR policies, privacy notices for applicants and employees, terms for vendor contracts, and documents for answering requests and complaints. We are not a technical auditor of AI models and do not certify that any vendor's system is accurate or unbiased, and we cannot guarantee the outcome of a dispute. Fees depend on the stage of the matter and the work involved. Send us the documents first, and we will give you the scope of work and fees in writing before we start. For an initial consultation, call +66 92 254 2045 or send us the details. See also HR documents and PDPA compliance.
This guide is published by Suwanvara Law Firm — a Khon Kaen law firm established in 1986. General information only, not legal advice on a specific matter.

