--- title: "Shipping an AI Product Safely Under PDPA: The Practical Compliance Checklist for Thai Startups Using LLMs" slug: "shipping-an-ai-product-safely-under-pdpa-the-practical-compliance" locale: "en" canonical: "https://ireadcustomer.com/fr/blog/shipping-an-ai-product-safely-under-pdpa-the-practical-compliance" markdown_url: "https://ireadcustomer.com/fr/blog/shipping-an-ai-product-safely-under-pdpa-the-practical-compliance.md" published: "2026-07-12" updated: "2026-07-12" author: "iReadCustomer Team" description: "Many Thai startups are unknowingly violating the PDPA by shipping AI features that send customer data overseas. Here is your battle-tested, practical compliance checklist." quick_answer: "Shipping an LLM product under PDPA requires startups to run local PII scrubbing on prompts before API dispatch, upgrade to enterprise API tiers that contractually forbid model training, and explicitly disclose automated data processing in updated privacy notices." categories: [] tags: - "pdpa compliance for startups" - "thai privacy law ai" - "llm data protection" - "cross-border api transfer" - "saas compliance bangkok" source_urls: [] faq: - question: "Why do startups using LLMs need to worry about Thailand's PDPA?" answer: "Every prompt sent to an external LLM API counts as a cross-border personal data transfer under PDPA. If these prompts contain customer emails, names, or identifiable actions, they violate Section 28 of the law unless backed by secure DPAs and proper structural privacy protections." - question: "Can startups use LLMs without requesting explicit consent from users?" answer: "Yes, if you scrub all identifying data before the API call or establish a legitimate interest. If the data is fully sanitized beforehand, it falls out of the scope of personal data processing under the PDPA." - question: "Consent vs Legitimate Interest: which is better for shipping AI features?" answer: "Legitimate interest is usually better for core optimization and standard product features because it avoids user friction. Consent should be reserved for processing sensitive data, high-impact automated profiling, or marketing integration." - question: "What is the cost of achieving PDPA compliance for an early-stage startup?" answer: "Basic compliance for a single AI feature can be done in days for under 15,000 Baht using open-source tools and template agreements. This is highly cost-effective compared to administrative fines that can peak at 5,000,000 Baht." - question: "How do you handle a user's deletion request within an AI model structure?" answer: "Never use raw user data for fine-tuning. Instead, use Retrieval-Augmented Generation (RAG) so that you can delete the user's data from your active database or vector storage instantly when requested without having to re-train the entire model." robots: "noindex, follow" --- # Shipping an AI Product Safely Under PDPA: The Practical Compliance Checklist for Thai Startups Using LLMs Many Thai startups are unknowingly violating the PDPA by shipping AI features that send customer data overseas. Here is your battle-tested, practical compliance checklist. Shipping an AI feature that sends raw customer data to an overseas LLM API without a legal framework violates Thailand's Personal Data Protection Act (PDPA). For many fast-moving founders, the priority is always speed-to-market. The temptation to simply hook up OpenAI's API to your product and worry about the legal details later is incredibly high. However, under Thailand’s data privacy regime, there is no such thing as a "we are just a startup" exemption. As regulatory bodies step up audit procedures and users become increasingly sensitive to where their prompts travel, you cannot afford to launch blind. Implementing a robust pdpa compliance checklist for thai startups is not just a defensive play; it is a critical requirement to build a defensible, investment-ready tech asset from day one. Failing to build privacy safeguards directly into your product’s architecture exposes your company to heavy administrative fines and permanent reputational damage. Many startups unknowingly leak proprietary data and customer secrets when feeding free-text inputs directly into external models. Every piece of telemetry, feedback, or text your user types is protected personal data if it contains identifying elements. * **Unencrypted API pipelines:** Transferring user-inputted prompt streams across the public internet without proper transit security. * **Zero vendor vetting:** Adopting consumer-tier AI API terms that legally permit providers to use your customer data to train their future systems. * **Indefinite prompt logging:** Storing prompt histories containing personally identifiable information (PII) on your internal databases without a clear expiry schedule. * **Absence of automated deletion:** Having no physical pipeline or mechanism to remove data from your system when a user invokes their right to be forgotten. * **Hidden processing practices:** Integrating generative models to parse sensitive profiles without disclosing the processing activity to the end user. --- ## Understanding Cross-Border Data Transfers in AI Systems The Thai PDPA treats API calls to overseas LLM servers as international data transfers that require explicit legal justification. Under Section 28 of the PDPA, sending personal data to a destination country that does not have adequate data protection standards is highly restricted. Since the leading foundation models are hosted on hyperscale data centers in the United States or Singapore, your application is transferring user data across borders on every single generation loop. **Failing to register or justify cross-border data transfer api solutions can freeze your startup’s development during its first major enterprise procurement audit.** Large clients and foreign investors will inspect your transfer mechanisms to ensure you are not creating a massive compliance liability for their operations. ### The AWS Singapore vs OpenAI US Dilemma * **Jurisdiction limitations:** US cloud instances are subject to different foreign intelligence surveillance laws compared to Southeast Asian servers. * **Data transit pathing:** Routing data through local regional endpoints like Singapore reduces latency and simplifies international compliance paperwork. * **Adequacy ratings:** Choosing model endpoints that comply with internationally recognized adequacy standards established by the PDPC. * **Contractual protection:** Deploying private VPC instances on local clouds guarantees your pipelines stay isolated from public internet transit routing. ### Why "We Are Just a Startup" is Not a Legal Shield * **Statutory flat penalties:** Regulatory authorities do not scale down maximum administrative fine ceilings for small businesses or pre-revenue projects. * **Investor due diligence failure:** Early-stage VCs routinely abandon deals during tech diligence when they discover illegal data transfer practices. * **Customer churning risks:** Modern tech-savvy B2B buyers require detailed data mapping before they agree to integrate any SaaS tool into their tech stack. * **Class action liabilities:** The [cost](/en/pricing) of settling collective lawsuits initiated by affected user groups can completely bankrupt an early-stage company. Before launching any public-facing bots, it is highly recommended to study [LINE Chatbot PDPA Compliance 2026: The Ultimate Consent & Opt-Out Checklist](/en/blog/line-chatbot-pdpa-compliance-2026-the-ultimate-consent-opt-out-checklist) to avoid basic structural errors in messaging-based user pipelines. --- ![| Compliance Attribute | Standard Consumer API | Enterprise API Instance | | :--- | :--- |…](https://land-admin.ireadcustomer.com/api/images/6a531b8440f2afa7c374546f) ## Step 1: Prompt Minimization and Sanitization Protocols Minimizing the personal data that enters your prompt payload is the easiest way to reduce your compliance surface area. By running an automated, open-source scrubbing tool like Microsoft Presidio on your application backend, you can wipe out social security numbers, emails, and names before they are dispatched to your external model. If your external LLM provider never receives any identifying data in the first place, your system is no longer performing a cross-border personal data transfer for those variables. **Implementing automated sanitization ensures that you are shipping ai products safely under pdpa without degrading the output quality of the application.** Stripping identifiers from user inputs acts as a failsafe against accidental prompt leaks and unexpected system injections. ### Stripping Direct Identifiers * **Regular expression matching:** Setting up strict search rules to flag and clean national ID formats, passport numbers, and phone digits. * **Named Entity Recognition (NER):** Utilizing light, local machine learning models to detect real-life names, addresses, and physical locations. * **Standardized masking tags:** Replacing sensitive entities with neutral bracketed placeholders such as "[REDACTED_NAME]" or "[USER_ID_1]" on the fly. * **Payload routing control:** Isolating non-personal transactional markers and forwarding only necessary contextual prompts to foreign servers. ### Contextual Data Masking * **Time resolution downgrading:** Generalizing specific timestamps into broader intervals like "Q3" or "morning sessions" before API transmission. * **Generalizing demographic markers:** Converting specific birth dates into simple age brackets to prevent individual re-identification downstream. * **Vocabulary simplification:** Substituting industry-specific terms or unique company handles with standard generic expressions. * **Anonymity-preserving routing:** Creating temporary non-traceable keys on your local secure database to link outputs back to real users safely. For a deeper look into maintaining technical performance alongside structural security, review the [Building an LLM Evaluation Suite for Business: Stop AI Features from Ruining Your Reputation](/en/blog/building-an-llm-evaluation-suite-for-business-stop-ai-features-from) guide to safeguard your product’s reliability. --- ## Step 2: Running a Legitimate Interest Assessment for AI Most AI-driven features do not actually require user consent if you establish a valid legitimate interest under the PDPA. The Thai Personal Data Protection Committee recognizes that business optimization and technical service improvements are valid processing grounds, provided that you conduct a formal Legitimate Interest Assessment (LIA) and document the results. Forcing users to click through five different popups just to use an autocomplete feature creates horrible friction and encourages dark-pattern designs that regulators actively target. **Documenting a rigorous legitimate interest assessment for ai provides you with a robust paper trail that serves as your primary defense during regulatory audits.** It demonstrates that you actively evaluated the balance between user privacy and business utility before writing a single line of production code. ### The Three-Part Legitimate Interest Test * **The Purpose Test:** Clarifying that your startup has a genuine, non-trivial business need to process the data with machine learning. * **The Necessity Test:** Confirming that the AI-powered task cannot be achieved through alternative, non-personal data processing techniques. * **The Balancing Test:** Evaluating whether the individual’s reasonable privacy expectations override your startup’s commercial interests. * **Document archiving:** Keeping the signed LIA evaluation on file so it can be instantly provided to the PDPC if an inquiry arises. ### When Consent is Non-Negotiable * **Biometric classification:** Utilizing user facial scans, voices, or fingerprints to build personalized profiling databases. * **Medical data scanning:** Processing health records, diagnosis histories, or wellness inputs via external neural networks. * **Third-party ad-network sharing:** Forwarding user behavioral prompts to external marketing platforms for behavioral targeted advertising campaigns. * **Automated life-altering decisions:** Employing predictive systems to dynamically reject credit scores or employment applications without human oversight. --- ## Step 3: LLM Vendor Due Diligence Checklist You cannot safely deploy AI without auditing your model provider's data processing agreement (DPA) to ensure they do not train on your prompts. Leading providers like OpenAI and Anthropic separate their personal consumer tiers from their commercial APIs. If you route sensitive startup data through consumer channels, your inputs might be logged to train next-generation public models, leading to catastrophic corporate intellectual property exposure. **Conducting a rigorous llm vendor due diligence audit is an essential operational requirement to satisfy both local PDPA guidelines and international client requirements.** Choose partners that legally commit to isolation, short retention windows, and transparent processing locations. | Compliance Attribute | Standard Consumer API | Enterprise API Instance | | :--- | :--- | :--- | | **Model Training Use** | Yes (often active by default) | No (contractually excluded from training sets) | | **Data Retention Limits** | Indefinite or up to 30 days | Deleted immediately after execution or up to 14 days | | **Sovereignty Options** | Fixed central servers (primarily US) | Regional server options (EU, Singapore, local) | | **SLA & Audit Support** | Generic web terms of service | Custom DPA and security audit certificates (SOC2) | * **Confirm retention timelines:** Verify exactly how long your chosen provider caches your inputs on their internal debugging storage. * **Review human evaluation terms:** Ensure that no human contractors are actively reading your prompts to test system alignment. * **Validate encryption standards:** Confirm the partner enforces TLS 1.3 for data in transit and AES-256 for data at rest on their platforms. * **Verify data sub-processors:** Identify if your chosen vendor utilizes external contractors or secondary cloud providers to run their models. --- ![Unencrypted API pipelines:](https://land-admin.ireadcustomer.com/api/images/6a531b8940f2afa7c3745475) ## Step 4: Redrafting Your Privacy Notice for AI Processing Your current privacy policy is likely illegal if it fails to explicitly state that personal data is analyzed by automated AI models. Section 23 of the PDPA demands that you inform users of the purposes of processing, the categories of data collected, and who that data is shared with. Simply having a vague clause that says "we share data with third parties" will not pass regulatory scrutiny during an audit or a user dispute. **Deploying an updated, clean disclosure via an ai privacy notice generator thailand guarantees transparency while keeping your user onboarding seamless.** The notice must detail exactly what happens to a user's prompt the moment they hit the generate button. ### Mandatory Disclosures for LLM Integration * **The nature of AI tooling:** Clearly informing users that their inputs are processed using advanced automated natural language models. * **Cross-border transfer destinations:** Explicitly naming the foreign countries and target servers where your provider’s models are run. * **The logic behind decisions:** Sharing high-level overviews of how the algorithms compute suggestions, recommend items, or filter entries. * **Opt-out mechanics:** Providing direct instructions on how users can choose to use alternative non-AI components within your software. ### Explaining AI Decisions to Customers * **Interactive privacy paths:** Splitting long legal policies into clear, readable sections using collapsible accordions and visual iconography. * **Use plain language descriptions:** Avoiding overly technical concepts and explaining data paths in terms a normal consumer understands. * **Dynamic UI tooltips:** Showing helpful info markers directly beside prompt input boxes explaining how that specific field is handled. * **Dedicated support channels:** Establishing an easily accessible email address for users to submit technical data queries. --- ## Step 5: Handling Data Subject Rights in LLM Ecosystems Honoring a user's request to delete their data becomes an engineering nightmare if their personal information is baked into your fine-tuned model weights. Under Section 33 of the PDPA, users have a legal right to request the deletion or restriction of their personal data. If you have fine-tuned a custom model using raw, un-anonymized database logs, you cannot easily "un-train" that model. You would have to re-train the entire model from scratch, costing your startup thousands of dollars and valuable engineering weeks. **Isolating your fine-tuning pipeline from raw, direct user identifiers is the only way to scale your AI systems without inviting operational debt.** Your technical architecture must support atomic deletion across all storage, vector DBs, and caching nodes. * **Decouple training from raw logs:** Never include raw customer identifiers or personal messages in fine-tuning datasets. * **Set up vector DB deletion routes:** Ensure your engineering team can locate and wipe specific embeddings when a deletion request is initiated. * **Automate API prompt purges:** Program your backend to clear chat histories and telemetry fields after predefined operational periods. * **Implement a 72-hour breach response:** Design a standard operational procedure to notify the PDPC within 72 hours of detecting any data leak. * **Build an internal audit trail:** Track every user deletion request and record the precise timestamp the action was completed across your stack. * **Utilize real-time consent ledgers:** Keep a secure ledger of consent states to verify you always have the right to process live data streams. --- ## Cost vs Compliance: Right-Sizing PDPA for Early-Stage Startups Achieving basic PDPA compliance for a single AI feature is a weekend engineering sprint, not a six-month corporate legal project. Early-stage founders often freeze because they assume compliance requires a massive data protection team and expensive enterprise legal advisors. In reality, a lean startup can implement safe prompt handling, select enterprise-grade APIs, and update their privacy notices for less than 15,000 Baht in initial tooling and template costs. **Investing a few hours of prep work before you write your launch code saves your business from catastrophic retrospective system redesigns.** To ensure you are prioritizing resources correctly, check out [The Ultimate SMB AI Governance Checklist Without a Data Team](/en/blog/the-ultimate-smb-ai-governance-checklist-without-a-data-team) to balance engineering speed with legal security. To safely launch your AI product under PDPA, execute these steps in chronological order: 1. **Draft a complete data map:** Document exactly how data flows from your front-end interface, through your backend, to the external LLM, and back. 2. **Activate automated prompt scrubbing:** Integrate a localized PII filter on your application server to clean incoming data payloads before API dispatch. 3. **Upgrade your vendor accounts:** Switch your API connections to dedicated enterprise accounts that explicitly disallow model training use. 4. **Publish your updated privacy notice:** Update your web and app portals to clearly detail your AI processing practices and host countries. 5. **Formulate a simple data-deletion runbook:** Establish a documented process so any team member can wipe a user's records within the legal window. --- ## Wrapping Up Your PDPA Compliance Checklist for Thai Startups Securing your AI architecture before you launch is the only way to build long-term venture value and scale without regulatory interference. Since Thailand began fully enforcing the PDPA in May 2022, compliance is no longer a superficial checklist item to display in your website footer. It is a fundamental indicator of your company's technical maturity and operational integrity. **Taking action to secure your prompt payloads today prevents devastating user trust damage and legal system bottlenecks tomorrow.** By keeping your systems lean, clean, and transparent, you prove to both users and global venture funds that your technology is built on a resilient, scalable foundation. * **Enforce continuous automated audits:** Regularly verify that no un-scrubbed PII is leaking past your backend filter into external logs. * **Monitor vendor policy updates:** Check your model provider's legal documentation quarterly to detect silent changes in retention or training policies. * **Maintain strict internal logs:** Record every legal assessment, DPA signature, and processing decision within your internal wiki. * **Engage with your user community:** Be open about your privacy-first architecture; transparency is a massive competitive differentiator in the AI era.