SentraMind

Briefing

BOT AI risk guidelines and third-party AI

How the Bank of Thailand’s 2025 AI risk guidelines treat a GenAI control sold to a bank, and which Thai cybersecurity changes are still draft.

Guidelines of 12 September 2025 class a bank-facing GenAI control as third-party AI

The Bank of Thailand’s AI risk management guidelines are dated 12 September 2025 on the verification sheet. Under those guidelines, a GenAI data-loss control sold to a Thai bank is assessed as a third-party AI system. The materials the sheet attaches to that assessment are governance, testing, and explainability documents. A GenAI control sold to a bank is, under these guidelines, a third-party AI system, and the assessment uses those three kinds of material.

An older technology-risk layer sits under the 2025 text. Notification SorNorSor 21/2562, from 2019, and implementation guidelines refreshed in November 2023, are the instruments on which the sheet verifies that Bank of Thailand IT risk covers data-loss controls. The September 2025 guidelines are the AI layer on that foundation. Parallel regimes for insurance and for securities are noted beside it. The sheet gives those regimes no separate clock in this account, and it does not merge them into the 12 September guidelines.

The classification records an assessment, and the sheet records no pass

The guidelines describe the assessment a bank applies to a third-party AI system. Governance documents, testing, and explainability documents are the pack of materials named for that assessment. The sheet contains no pass result, no approval number, and no sentence in which a product is said to have satisfied the guidelines. A slide from “assessed as a third-party AI system” to “has passed the guidelines” adds a result the sheet refuses.

The table captioned “A bank-facing GenAI control is a third-party AI system here” holds the date, the three kinds of document, and the refused pass in a single row. That row exists so the classification and a claim of passage are not written as one fact. The later rows, on the personal-data clock and on language, do not convert the guidelines into an approval either. This briefing can say how the guidelines classify the control. Passage is a further fact, and the sheet does not state it.

Notice to the PDPC under the PDPA is 72 hours

The PDPA, given on the sheet as B.E. 2562 (2019), has been fully in force since 1 June 2022. No amending act is recorded. Subordinate PDPC notifications carry the working detail, including security measures, breach notification, ROPA, data-protection officer notification, and cross-border rules effective in March 2024. The breach rule in that set is notice to the PDPC within 72 hours, and notice to individuals when the risk is high.

Enforcement of the interval is already in the record the sheet cites. The first fine there, in August 2024, was THB 7 million, and THB 3 million of it attached to missing the 72-hour notice. The figures show the clock being used. They are not a tariff for a later case. Further guideline consultations opened in March 2026. The sheet does not present those consultations as a replacement of the 72-hour rule, and it does not present them as the draft 24-hour critical-infrastructure clock.

The 24-hour critical-infrastructure clock is still a draft

The Cybersecurity Act, B.E. 2562 (2019), is in force, and a revised critical information infrastructure list was gazetted on 16 September 2025. The list contains seven sectors, among them government services and banking. The gazette is a completed publication. The sheet does not give that list a 24-hour incident clock.

The 24-hour clock appears in a draft amendment. The country note dates a hearing to July 2025, and watch-list item 4 still marks the amendment as a draft as of August 2025. The draft would also bring cloud providers and data-centre providers into scope. It remains a draft in this briefing. The personal-data interval that is actually in force stays 72 hours to the PDPC, with notice to individuals when the risk is high. Using 24 hours as today’s infrastructure deadline would brief watch-list item 4 as if it had commenced.

The NCSA cloud and website texts are marked secondary-sourced

The sheet records an NCSA Cloud Security Standard effective 10 September 2026, and Website Security Standard WSS 1.0 effective 17 September 2026. The country row says those texts apply to government, to critical information infrastructure, and to tender vendors, and it marks the row as secondary-sourced, to be confirmed with the NCSA. The market-entry note is narrower. The website standard may reach a management console when the bid is a public tender, and the cloud standard is relevant if a cloud variant is offered. Both notes carry the secondary-source limit.

Watch-list item 4 holds that confirmation in the same item as the draft amendment. The September 2026 dates are what a secondary source supplied. They are not confirmed obligations in this briefing. The sheet’s instruction is to confirm the texts with the NCSA before they are used as bid conditions. Alongside them, the seven-sector list gazetted on 16 September 2025 is stated without that secondary-source mark, and it still carries no 24-hour clock.

A Thai-language interface is a de facto tender practice

The sheet finds no rule that makes a Thai-language interface mandatory for every government deployment. What it records is a de facto tender practice, the sort of requirement that shows up in terms of reference. DGA standard 11-2566 is the formal Thai-language standard, and it covers government websites. A website standard and a statute for every deployment are different things. The sheet has the website standard, and it has the tender practice beside it.

A government website procurement can meet a Thai-language expectation in its terms of reference, and through standard 11-2566 where that standard applies. A deployment that is not a government website does not pick up a statutory duty, from this sheet, to present a Thai interface in every case. The table’s last row draws the same line. One cell is the de facto tender practice for government websites. The cell it refuses is a statute for every deployment. The practice can still shape a tender. It does not, on the verification sheet, govern every deployment.

How to read this

A bank-facing GenAI control is a third-party AI system here
A bank-facing GenAI control is a third-party AI system here
What is in forceWhat is still openDo not collapse into
BOT AI risk guidelines 12 Sep 2025; governance, testing, explainability A statement that a product has passed them
PDPA notice 72 hours to the PDPC The draft 24-hour infrastructure clock
Thai-language UI De facto tender practice for government websites A statute for every deployment

Orientation only. This briefing is not legal advice. A watch-list item is not a statement that the cited instrument is in force.