Public Procurement and AI: The Compliance Questions
A checklist of questions Thai government buyers should ask when procuring AI: data residency, PDPA, security, explainability, oversight and exit. General information only.
AI public procurement in Thailand raises questions that ordinary software purchases do not. Buyers need to know where data is stored, how the system reaches its answers, who is accountable when it is wrong, and how to switch suppliers later. This article lists those questions so a procurement team can write them into requirements and evaluate bids with confidence.
It is written for procurement officers, IT leads and legal teams in Thai government agencies and state enterprises. It is general information, not legal advice. Rules on public procurement, data protection and cybersecurity change, so confirm the current position with your own counsel and the relevant regulators before you act.
Why AI public procurement in Thailand needs different questions
A traditional system does what its code says. An AI system produces outputs that depend on a model, on training, on the data you give it and on settings that can change. Two identical questions can get slightly different answers, and a model update can shift behavior without any change on your side.
That affects how you write requirements and how you test deliveries. A specification that lists screens and fields is not enough. You also need to describe acceptable accuracy, what happens when the system is unsure, and how you will check quality after go-live.
The compliance questions, grouped
The table below is a working checklist. Use it to shape requirements, not as a substitute for your agency's own rules.
| Area | Question to ask | What a good answer looks like |
|---|---|---|
| Data residency | Where are data, models and backups hosted? | In Thailand, on sovereign cloud or on-premises, stated in writing |
| Personal data | How is personal data collected, limited and protected? | A documented approach aligned with Thailand's PDPA, with named responsibilities |
| Security | What standards does the supplier follow? | Recognized practices such as ISO/IEC 27001, plus evidence of testing |
| Explainability | Can the system show why it answered as it did? | Citations to approved sources, logs, and a readable decision record |
| Human oversight | Who decides when the AI is unsure? | A named officer, with a defined handoff path |
| Accuracy | How is quality measured on our documents? | A test on your own samples, with agreed thresholds |
| Audit | What is logged, and for how long? | Searchable logs held in your environment |
| Exit | How do we leave or change supplier? | Data export in open formats, documented interfaces |
| Subcontractors | Who else touches the system or data? | A disclosed list, with the same obligations passed down |
Data residency and sovereignty
Public-sector data often has to stay in Thailand. Ask for the exact location of every component: the application, the database, the model, backups and any support access from abroad. Check current guidance from the Digital Government Development Agency (DGA) and the standards that apply to your agency.
Watch for services that send prompts to a model hosted overseas. The application may be in Thailand while the model call is not. The bid should say so plainly.
Personal data and the PDPA
Many public AI uses involve names, ID numbers or contact details. Thailand's Personal Data Protection Act applies in many public-sector cases, and the PDPC publishes guidance. Ask the supplier how they limit the data the AI sees, how they mask identifiers, who can access it and how deletion works. Your data protection officer should review the answers.
Security and standards
Ask which security frameworks the supplier follows and what evidence exists, such as audit reports or test results. ISO/IEC 27001 is a widely used reference for information security management, and the National Cyber Security Agency (NCSA) provides guidance for Thai organizations. A certificate alone does not prove that your deployment is secure, so ask how your environment will be tested.
Explainability and records
For decisions that affect citizens, you need to explain afterward what the system did. The practical test: can an officer open a record and see the question, the sources used, the answer and who approved it? If a supplier says the model is a black box, treat that as a finding.
Human oversight
State in the requirements which decisions the AI may only suggest. Anything that affects a person's rights, money or access to services should have an officer making the final call. Define what happens when confidence is low, such as routing the case to a person with a summary.
Writing requirements you can test
Vague requirements produce vague bids. Prefer statements that can be checked.
- Name the task. "Answer citizens' questions about permit steps in Thai from approved regulations" is testable. "Provide an intelligent assistant" is not.
- Supply test material. Give bidders a set of real, anonymized documents or questions, and ask them to show results on it.
- Define unsure behavior. Require the system to say it does not know and route to an officer.
- Require citations. Each answer should point to a source passage.
- Specify language. State Thai handling needs, including formal and informal forms, and mixed Thai-English text.
- Set acceptance criteria. Agree on how quality will be measured at delivery and again after go-live.
Evaluating bids
Ask suppliers to demonstrate rather than describe. A working demo on your sample documents tells you more than a slide deck. Be careful with demos that use clean, curated inputs, since real files are messier.
Check what is included and what is not. Questions to put to every bidder:
- Which parts are built by you, which are third-party, and which models are used?
- What happens if the underlying model is withdrawn or changed?
- Who maintains the system after handover, and how is knowledge transferred?
- What does your team need from our staff, and for how long?
- How do you handle a mistake the system makes in production?
Prefer clear scope and phased delivery to a single large commitment. A short proof of concept on real data, with a decision point before wider rollout, limits risk for both sides. For a view of how a pilot-first approach looks in practice, see our AI for government and public sector page.
What can go wrong
Requirements copied from a vendor. If the specification mirrors one supplier's brochure, competition suffers and you inherit their blind spots. Write requirements from your own process.
Lock-in. Proprietary formats and closed interfaces make switching hard. Require open export and documented interfaces from the start.
Accuracy promised, not measured. A bid may claim high accuracy without saying on what data. Insist on your own samples.
Forgotten operations. AI systems need monitoring, regular updates to rules and review of failures. If the contract ends at delivery, quality will decay.
Unclear accountability. Decide in advance who is responsible when the system errs: the agency, the supplier, or both under defined conditions.
When to pause before buying
Pause if you cannot say which decision the system supports, if the data is not ready, or if no officer has been assigned to own it. Buying first and finding a use afterward is a common reason public AI projects stall. A short assessment of your processes and data is often the better first step.
How to start
- Choose one service. Pick a citizen-facing or internal process with clear volume and rules.
- Draft the checklist. Adapt the table above with your legal, IT and data protection teams.
- Collect sample material. Prepare anonymized documents for testing bidders.
- Plan a pilot. Agree on scope, success measures and an exit decision.
To see what a government deployment can look like, open our working demo with sample data or read what a provincial leader sees on an AI governance dashboard. You can also review our solutions overview.
Next step
If you are preparing requirements for an AI system, our consultants can review your draft checklist and point out gaps before it goes to market. Get in touch.
Frequently asked questions
What should a government AI tender specify?
The task in plain terms, test material from your own documents, behavior when the system is unsure, source citations, Thai language needs, acceptance criteria and data location. Testable statements produce comparable bids.
Does government AI have to be hosted in Thailand?
Public-sector data often has to stay in the country, but the rules depend on your agency and the data. Check current guidance from the DGA and confirm with your own counsel.
How do we avoid supplier lock-in with an AI system?
Require data export in open formats, documented interfaces, a clear plan if a model is changed or withdrawn, and knowledge transfer at handover. Raise these in the requirements, not after award.
Is this article legal advice?
No. It is general information. Procurement, data protection and cybersecurity rules change, so confirm the current position with your counsel and the relevant regulators.
Ready to put this to work?
Turn this playbook into a working system
From a first agent to a national platform, the same team plans it, builds it and runs it with you.
