For a chatbot implementation across Thailand and ASEAN, choosing a model is rarely the hardest decision. The difficult work is defining what the bot may answer, which source is authoritative, how Thai and mixed-language queries are handled, when a person must take over, and what evidence will prove the system is ready. This guide helps management, IT, procurement, operations, legal and DPO teams turn those decisions into an actionable RFP and acceptance plan.
Executive decision: seven items to approve before chatbot implementation
A “chatbot” can mean a public FAQ, an employee search interface or an agent that updates business systems. Quotations are not comparable until bidders receive the same scope and constraints. Put these seven items on the investment-approval page:
- Users: employees, distributors, existing customers or the public.
- Work: inquiry handling, document search, procedure guidance or incident triage.
- Channels: web, LINE, Microsoft Teams or an internal portal.
- Languages and the accountable owner of terminology in each language.
- Approved knowledge, personal data, confidentiality classes and access boundaries.
- Conditions under which AI must not answer, plus the human queue, operating hours and SLA.
- Business, quality, safety and cost metrics used for acceptance and ongoing operation.
If these decisions are not yet possible, a constrained proof of concept is more responsible than negotiating an enterprise rollout. Our related guides explain a broader generative AI implementation approach for overseas subsidiaries and how to frame AI PoC costs and success criteria in Thailand.
Go, conditional go and no-go
| Decision | Required condition | Management action |
|---|---|---|
| Go | Knowledge owners, sources, access, escalation and acceptance criteria are agreed | Proceed to a limited implementation |
| Conditional go | Value is plausible, but language quality or personal-data flow remains unresolved | Restrict users, knowledge and channels in a PoC |
| No-go | No authoritative source, no accountable owner, indiscriminate ingestion of confidential files, or unsupervised public answers | Redesign requirements before ordering |
A 90-day plan later in this article is an illustrative project structure, not a delivery guarantee. Adjust it for integrations, legal review, data readiness and local stakeholder availability.
FAQ bot, enterprise search, RAG and AI agent are not the same
Use behavioural requirements in the RFP rather than relying on product labels.
| Approach | How it responds | Good fit | Main limitation |
|---|---|---|---|
| Rule-based FAQ bot | Matches configured questions, answers and flows | Opening hours, fixed guidance, application links | Maintenance and poor coverage of unexpected wording |
| Keyword/full-text search | Returns matching documents or passages | Finding policies and manuals | Users must know useful search terms |
| RAG | Retrieves relevant passages and asks a generative model to answer from them | Cross-document questions, summaries and procedures | Retrieval failure, stale knowledge and unsupported synthesis |
| AI agent | Answers and invokes approved tools or APIs | Ticket creation, stock lookup, request-status checks | Incorrect action, excessive privilege, duplicate execution and audit complexity |
RAG does not simply mean “training AI on company files.” In a typical retrieval-augmented generation flow, the system finds passages at question time and supplies them as evidence. Document parsing, metadata, access filtering, ranking, citations and update propagation therefore matter as much as model quality.
An AI agent adds business-system actions. A read-only inventory lookup is a different risk from creating a purchase order or changing customer data. For an initial chatbot implementation, separate answers and searches from write operations. Release write access only after explicit confirmation, renewed authorization, allow-listed parameters, idempotency and an auditable activity trail exist.
Select a chatbot implementation use case
Score value and feasibility separately
Do not prioritise solely by inquiry volume. Score frequency, handling effort, answer standardisation, knowledge readiness, impact of a wrong answer, access complexity and language coverage.
| Candidate use case | Initial suitability | Conditions |
|---|---|---|
| IT help-desk procedure guidance | High | Authoritative procedures and escalation are manageable; account actions need separate controls |
| HR and general-affairs policy guidance | Medium to high | Country, legal entity, employment class and policy version must be identified |
| Product and maintenance manual search | Medium to high | Model, revision and machine identifiers need reliable metadata |
| Quotation or delivery response | Medium | ERP freshness, price authority and external-response approval are required |
| Final quality or safety decision | Low | Limit the bot to evidence and routing; keep accountable human judgement |
| Contract or legal conclusion | Low | Document retrieval may help, but route conclusions to legal counsel |
State exclusions before features
List what the first release will not do: final medical or safety decisions, employee assessment, unapproved price promises, beneficiary changes, irreversible approval or deletion, and free-form collection of sensitive personal data. These can be reconsidered only after explicit controls and approval.
For every use case, record the owner, users, sample inputs, sources, output format, prohibited answers, exceptions, human destination, peak profile, languages and retention. This register becomes the common source for scope, quotation, tests and operations.
Multilingual and Thai-language chatbot design
Multilingual service is more than translating the interface. Language detection, retrieval, terminology, answer language, citations, routing and analytics must remain aligned.
Separate three locale fields
- User locale: the preference stored in a profile or channel.
- Conversation locale: the language currently used in the session.
- Content locale: the language, country and site of the source material.
Test code-switching, such as a Thai sentence containing an English product name, acronym or model number. Do not translate identifiers mechanically. A useful policy may answer in the requested language while showing the authoritative source in its original language and, where appropriate, a clearly marked summary.
Microsoft Copilot Studio language support, updated on 29 June 2026, lists Japanese, Thai and Vietnamese for relevant generative-answer and user-language capabilities. It also shows that support varies by feature and that a primary language cannot be changed after creation. Verify the exact feature, tenant, region, channel and language combination before contracting.
Microsoft’s multilingual agents documentation describes dynamic language switching under generative orchestration, while a classic chatbot handles one language and manually authored topics remain the author’s translation responsibility. This is one product example, not a universal recommendation; it demonstrates why an RFP must distinguish automatically covered content from content maintained by people.
Thai variants and terminology governance
Thai retrieval can be affected by segmentation, English loanwords written in Thai or English, internal abbreviations, model codes and differences in politeness. Treat the glossary as governed data:
| Field | Purpose | Rule |
|---|---|---|
| Concept ID | Stable identifier for a product or policy concept | Connect equivalents across languages |
| Approved term | Company-approved Thai wording | Prefer in answers and UI |
| Alias/acronym | English, transliteration and departmental shorthand | Use for retrieval expansion after ambiguity review |
| Prohibited term | Wording disallowed by legal or brand owners | Use in output checks |
| Scope | Country, entity, plant, product and validity period | Prevent cross-site policy errors |
| Owner and review date | Business and translation accountability | Drive review and audit |
Translation acceptance must test whether meaning, conditions, negation, deadlines and accountable actors remain unchanged. For safety procedures, employment rules and contractual terms, use an approved local-language document as the source wherever possible rather than making machine translation the authority.
Reference architecture for chatbot development
- Channel: web, LINE, Teams and portal adapters.
- Identity and session: user, entity, plant, role and conversation state.
- Orchestration: language, intent, retrieval, tool calls, refusal and escalation.
- Knowledge: approved documents, metadata, indexes and version control.
- Models: generation, classification, translation and embeddings.
- Integration: CRM, ERP, MES, ITSM, ticketing and notification.
- Governance: logs, evaluation, monitoring, access, secrets, retention and deletion.
Require the bidder to identify the provider of each layer, countries or regions through which data travels, storage locations, subprocessors, failure responsibilities, change notifications, and data return and deletion at exit. A security certification for one SaaS component does not describe the risk of the assembled system.
RAG knowledge pipeline
- Collect only approved content from SharePoint, file stores, websites or databases.
- Normalise PDF, tables, OCR, headers, footers and duplicates.
- Chunk without breaking procedures, model applicability or revision context.
- Attach language, entity, site, department, product, classification, effective and expiry dates.
- Configure lexical/vector retrieval and reranking for the use case.
- Enforce access before or during retrieval.
- Answer from evidence and abstain when evidence is insufficient.
- Show source, revision, passage and link.
- Monitor update and deletion propagation from source to index and cache.
Document count is not a quality KPI. Conflicting revisions, drafts, expired policies and another country’s rules create confident but operationally wrong answers.
RFP requirements for a comparable quotation
Ask each bidder to answer with standard feature, configuration, custom work, third-party service, assumption, limitation, extra charge and evidence—not just “supported.”
| Requirement group | Question to include | Evidence to request |
|---|---|---|
| Use case | Users, exclusions, peak, hours and countries | Use-case flow and assumptions |
| Languages | Detection, switching, code-switching, glossary, translation ownership and citation language | Live JA/EN/TH/VI test and limitation matrix |
| RAG | Formats, chunking, search, reranking, citations, updates and deletion | Evaluation, retrieval logs and revision demo |
| Identity/access | SSO, RBAC, entity/site boundaries and guests | Access matrix and negative tests |
| Personal data | Purpose, lawful basis as applicable, retention, transfers, deletion and DSAR | Data-flow diagram, processor list and draft DPA |
| AI safety | Prompt injection, disclosure, unsafe content and tool abuse | Threat model, controls and red-team results |
| Integration | LINE, Teams, web, ERP, CRM and ITSM | API, signature, retry and idempotency design |
| Human service | Trigger, summary, queue, SLA and hours | Screens, alerts and fallback path |
| Operations | Monitoring, evaluation, changes, incidents, backup and exit | RACI, runbook and monthly report sample |
| Commercial | Setup, usage, licences, support, overage, currency, tax and termination | Three-year TCO and unit-price sheet |
Evaluate vendors using the same data and scenarios. A free-form demo lets each bidder show only its strongest path. Treat roadmap items separately unless delivery date, plan, region and contractual commitment are explicit.
Knowledge, access control and Thailand PDPA in one data flow
Personal data may appear in the question, conversation log, feedback, human ticket, analytics, backup and model API request—not only in source documents. Map collection point, purpose, controller/processor role, location, recipient, retention, deletion and cross-border transfer.
This article is not legal advice. Thailand PDPA obligations depend on the processing context. Use the Thai PDPC’s GPPC and GPPC PLUS resources as practical references for records of processing, consent management where relevant, breach work and data-subject requests, and have your DPO or counsel confirm roles, legal basis, notices, transfers, retention, processor terms and incident procedures.
Preserve document-level access through the answer
Authentication is not permission to search every internal file. Match user attributes against content metadata and preserve that decision across retrieval results, generated text and citation links. Segregate HR, payroll, customer pricing, unreleased products, quality incidents and audit evidence.
Negative tests should cover another entity or plant, a user without access, a former employee, expired session, shared terminal and guessed URL. Measure how quickly a revoked permission reaches indexes and caches.
Verify model data terms at contract level
OpenAI enterprise privacy states, among other commitments, that business data is not used to train models by default. Retention, logging, eligible services, region, subprocessors and exceptions still need confirmation for the selected product and agreement.
OpenAI’s 19 August 2026 Zero Data Retention update is a current example relevant to buyers focused on retention. Do not assume ZDR automatically covers every API, feature, log or contract. Confirm eligible APIs and features, exclusions, approval, contractual wording and abuse-monitoring treatment. This is a comparison question to ask all candidates, not a recommendation of one vendor.
Connecting LINE, Teams and web
LINE: signature, asynchronous work and duplicate delivery
The LINE webhook receiving guide addresses asynchronous processing and the possibility of redelivery and duplicate events. Follow LINE’s webhook signature verification using the raw request body. Reformatting and reserialising JSON before verification can invalidate the security check.
Separate receive, signature validation, acknowledgement, queueing, deduplication, AI processing and reply. Derive an idempotency key from the event so redelivery cannot create a second ticket or repeat a write. Redelivery is not a guarantee that the business action completed; monitor state and provide a controlled reprocessing queue.
Teams and web use different identity strength
Teams and internal portals may supply organisational identity, group and device conditions. Public web sessions are usually anonymous. Maintain separate policies for anonymous and authenticated users; never elevate access from a URL parameter or self-declared identity.
For a shared backend, normalise conversation, user, channel, entity, locale, consent and human-service state while preserving channel-specific constraints. Operational traceability should not expose unnecessary personal data to analytics users.
Security: combine GenAI risks with standard application controls
The OWASP Top 10 for LLM Applications 2025 includes Prompt Injection and Sensitive Information Disclosure. NIST AI 600-1, the Generative AI Profile, helps organisations place GenAI risks in an AI risk-management process. Translate these references into a threat model for the selected use cases.
Prompt injection is not solved by a warning sentence
Test direct instructions and indirect instructions hidden in retrieved pages, PDFs, email or attachments.
- Distinguish system instructions, user input, retrieved content and tool output.
- Never treat instructions inside retrieved documents as trusted policy.
- Apply least privilege, allowed parameters, limits and confirmation per tool.
- Keep secrets and irrelevant confidential data outside model context.
- Validate output before using it as SQL, HTML, email or an API argument.
- Re-run attack and regression tests after model, prompt or retrieval changes.
Sensitive disclosure and secrets
Do not embed API keys, webhook secrets, connection strings, personal data or another department’s records in prompt templates. Retrieve secrets at runtime from a proper store and mask logs, errors and analytics. A contractual “no training” term does not mean the application itself stores no logs.
Signature, idempotency and approval
Validate signature, timing, sender and retries for external webhooks. Make write APIs idempotent. For purchases, refunds or customer-data changes, let AI prepare a proposal but require a person or explicit policy engine to approve it and retain an audit record.
Human escalation is an operating process
A transfer button is not a complete process. Define who receives what, when, in which language, and what the bot does after transfer.
Trigger escalation when the user requests a person; evidence is missing or contradictory; the issue concerns personal data, contract, pricing, complaint, safety or quality; repeated attempts show no resolution; defined high-risk language appears; or an integration fails, times out or denies access.
Provide the agent with a necessary and consent-appropriate summary: question, attempted steps, sources, language, relevant customer or machine identifier and urgency. Mask unrelated personal or confidential information. A human answer must pass knowledge approval before becoming permanent bot content.
Display service hours, supported languages, queue, acknowledgement expectations and the out-of-hours alternative. A contract may use example criteria such as “acknowledge within the agreed period in business hours” and “immediately notify the critical queue,” but actual times must be based on staffing, channel, severity and existing SLAs.
Test data and acceptance criteria
Do not accept on one accuracy number
Attach expected facts, mandatory evidence, prohibited content, acceptable variation, expected escalation, access and language to each question.
| Dimension | What is measured | How to write an illustrative contract criterion |
|---|---|---|
| Grounding | Approved evidence supports the answer | Critical questions include every mandatory fact and a source link |
| Retrieval | Correct document, revision and site are found | Measure retrieval within top-k on the buyer’s test set |
| Abstention | The bot avoids unsupported or prohibited answers | Prohibited cases refuse and offer the defined next step |
| Access | No other team or entity data is disclosed | Zero disclosures in agreed negative tests as a mandatory gate |
| Language | Meaning, terms, conditions and negation survive | Local reviewers find zero critical mistranslations in the acceptance set |
| Latency | Users receive an answer within a usable period | Measure p50/p95 by channel and peak window on buyer infrastructure |
| Escalation | The correct cases enter the correct queue | Propose 100% routing for agreed critical cases |
| Security | The app resists attack and tool misuse | Zero critical incidents in the agreed attack set as a gate |
“Zero” and “100%” above are examples of stringent acceptance gates for access, critical translation and critical routing; they are not production performance claims or a guarantee of no future incident. Define the population and risk tolerance, then approve values in your own acceptance plan. General-answer quality and response-time targets should likewise be proposed and calibrated through a baseline measurement.
Build native multilingual test sets
A Japanese set that is merely machine-translated will not reproduce local behaviour. Collect actual Thai and Vietnamese shorthand, typos, polite and colloquial forms, English mixing, model numbers, units and date formats. Assign a concept ID to equivalent questions and compare the conclusion across languages.
Cover normal, ambiguous and multi-intent questions; missing, stale and conflicting evidence; denied access; personal data; prompt injection; long input; attachments; integration failure; duplicate events; and human transfer. Add production failures to the regression set, and rerun it after knowledge, prompt, model, search or connector changes.
TCO and quotation comparison
The cheapest setup can become expensive after data preparation, translation, evaluation, human service and usage charges. Compare a common investment horizon such as three years.
| Cost area | Include | Common omission |
|---|---|---|
| Discovery/design | Process, data flow, threat model and UX | Local business, DPO and legal time |
| Build | Channels, RAG, SSO, integration and administration | Non-standard connectors, network and test environment |
| Data | Inventory, OCR, chunking, metadata and translation | Expired content, tables and scans |
| Licences | User, tenant, channel and administrator | Minimum commitments, environments and guests |
| Usage | Model, embeddings, retrieval, speech and storage | Long threads, retries and evaluation traffic |
| Operations | Monitoring, updates, incidents and support | Multilingual review and out-of-hours handling |
| Security | Testing, SIEM, secrets and audit | Retesting, retention and incident response |
| Exit | Export, deletion evidence, rebuild and training | Portability of indexes and evaluation data |
Give bidders assumptions for monthly conversations, turns, average input/output, document volume and change rate, concurrency, evaluation traffic and human transfers. Confirm currency, tax, exchange-rate treatment, model changes, caching, spending alerts and overage behaviour.
Show differences in assumptions beside price. If one vendor includes content preparation, another makes it the customer’s job and a third excludes translation, total quotations are not equivalent. Evaluate compliant TCO, changeability, data portability and accountable operations—not headline price alone.
Illustrative 90-day implementation plan
Days 1–15: requirements and data boundary
Appoint sponsor, process owner, information owner, DPO/legal, IT and local-language reviewers. Agree use cases, exclusions, channels, languages and KPIs. Map data, access, retention and transfers. Inventory sources with version, owner and classification. Build the first test set from real inquiries.
Days 16–35: RFP, selection and design
Compare candidates using identical scenarios. Demonstrate RAG citations, access denial, escalation and logs. Negotiate SLA, security, data terms, TCO and exit. Make acceptance and change control contractual deliverables.
Days 36–65: build and evaluate
Limit the build to representative content, users and channels. Configure JA/EN/TH/VI terminology and answer style. Evaluate normal, refusal, access, attack, connector failure and escalation cases. Classify each defect as knowledge, retrieval, access, model, UX or operation rather than blaming the model by default.
Days 66–80: controlled pilot
Limit departments and sites. Explain AI limitations and escalation. Review samples daily for unsupported answers, retrieval failure, unresolved requests and language issues. Keep high-risk features behind a feature flag. Confirm that local staff can operate the queue and knowledge approval.
Days 81–90: accept and decide
Run the fixed acceptance set and analyse pilot data. Report residual risks, exceptions, cost and unresolved items. Decide go, conditional go, extension or stop. If proceeding, release additional countries, teams, channels and write permissions in stages.
Common failures
- Mistaking fluent demos for accuracy: verify source, revision, access and refusal.
- Uploading every PDF: control ownership, validity, classification and deletion.
- Delegating Thai entirely to automatic translation: test local abbreviations, code-switching and critical instructions with local owners.
- Providing no exit when AI cannot answer: connect refusal to a staffed workflow.
- Using harmless PoC data but sensitive production data: reassess the production data flow and permissions.
- Omitting usage and human cost: include evaluation, monitoring, translation, updates and reprocessing.
- Changing models without change control: version, approve, regression-test and retain rollback.
Chatbot implementation FAQ
How should chatbot implementation costs be compared?
Use the same workload assumptions and include setup, licences, model/search usage, data preparation, translation, security, monitoring, knowledge maintenance, human handling and exit. Convert exclusions and customer work into cost or effort.
Should chatbot development start with RAG?
RAG is suitable when approved documents contain the answer and questions cross multiple sources. A fixed FAQ may be sufficient for stable guidance. Actions require more than RAG: API authorization, confirmation, idempotency and audit.
Is one model enough for a multilingual chatbot?
Model-language support is only one layer. Test retrieval, embeddings, OCR, voice, topics, UI, analytics and human queues in each language. Local business reviewers must accept terminology and critical procedures.
What matters in Thai chatbot acceptance testing?
Include native Thai, English acronyms and model numbers, variants, colloquial and polite wording, short queries and typos. Confirm that negation, conditions, deadlines, actors and units remain correct and that sources belong to the right Thai entity, site and revision.
How should an accuracy percentage be set?
Do not use one universal number. Separate critical questions, general FAQ, retrieval, refusal, access and language. Set proposed thresholds after measuring your test set. Zero leakage in agreed access-denial tests can be a mandatory acceptance gate, but it is not a claim that future production incidents are impossible.
Is a cloud security certification enough for Thailand PDPA?
No single certificate answers the full processing context. Have the DPO or legal team confirm purpose, roles, notices, transfers, retention, deletion, DSAR, breach response and processor terms across input, logs, knowledge and integrations.
How are duplicate LINE actions prevented?
Verify the signature over the raw request, acknowledge promptly, process asynchronously and persist an idempotency key. A repeated event must not create a second ticket or repeat a business update.
Conclusion
Successful chatbot implementation in Thailand and ASEAN begins with business scope, authoritative knowledge, language ownership, access, PDPA data flow, human escalation and measurable acceptance—not model selection. Compare bidders through one RFP, one test set and one TCO basis. Treat RAG as an evidence-and-permission system, and release agent actions only with least privilege, approval, idempotency and audit.
TOMAS TECH can support process scoping, RFP requirements, multilingual RAG, LINE/Teams/web integration and acceptance design for Thailand and ASEAN—even before a product is selected. If you want to clarify your data boundary or compare quotations on a common basis, contact our team.
References
- Microsoft Copilot Studio: Language support
- Microsoft Copilot Studio: Configure multilingual agents
- NIST AI 600-1: Generative Artificial Intelligence Profile
- OWASP Top 10 for LLM Applications 2025
- OpenAI Enterprise Privacy
- OpenAI: Our commitment to Zero Data Retention
- LINE Developers: Receiving webhook events
- LINE Developers: Verify webhook signature
- Thailand PDPC: GPPC
- Thailand PDPC: GPPC PLUS
