When evaluating a chatbot in Thailand, comparing subscription prices alone will not reveal the real chatbot cost. A useful comparison must cover 12-month total cost of ownership (TCO): requirements, Web or LINE channels, RAG over company knowledge, ERP/CRM integration, testing, operations, PDPA controls, and human handoff. This guide separates public price facts from project estimates and turns them into practical RFP, PoC, and acceptance criteria.
Price information was viewed on August 27, 2026. The API, platform, and LINE Official Account prices below are examples of vendor or channel costs. They are not an estimate of the total implementation price from TOMAS TECH or any other provider. Actual costs vary with tax, region, contract, exchange rate, data volume, and development scope.
Chatbot cost in one sentence: compare 12-month TCO, not a monthly fee
A Thai company comparing chatbot options should calculate the same seven cost areas for every proposal:
- Requirements and business-process design
- Channels such as Web, LINE Official Account, and Microsoft Teams
- RAG, search indexes, and knowledge preparation
- ERP, CRM, ticketing, and other business integrations
- Testing, security assessment, and acceptance
- Monitoring, improvement, support, and other operations
- PDPA controls and human handoff
A price per message, per million tokens, or per month represents only part of this picture. In production, people and systems must handle unanswered questions, update documents, review usage logs, enforce permissions, and respond to incorrect answers. Every comparison should therefore state what is included, what is excluded, and the quantity behind each line.
Three prices that are often confused in a chatbot comparison
1. Channel cost: LINE Official Account and similar services
The current Thai LINE Official Account page shows Free with 300 messages per month, Basic at THB 1,280 per month with 15,000 messages, and Pro at THB 1,780 per month with 35,000 messages. Additional messages are THB 0.10 each for Basic and THB 0.06 each for Pro; 7% VAT is additional. This is a LINE channel cost. It does not include AI generation, system integration, operational support, or human-agent labor.
The word “message” may mean different things in LINE billing, an AI platform, and an internal conversation report. In a quotation, define how many channel messages, AI calls, and business API calls one user question is expected to trigger.
2. AI and agent-platform usage
The Microsoft Copilot Studio pricing page displays a USD 200 monthly capacity pack of 25,000 Copilot Credits, with a note that pricing varies by region. Microsoft Learn explains that Copilot Credit consumption depends on the features and processing used, so it does not directly equal the number of end-user conversations or messages.
In pricing announced by OpenAI on July 30, 2026, GPT-5.6 Terra is USD 2 per one million input tokens and USD 12 per one million output tokens. GPT-5.6 Luna is USD 0.20 input and USD 1.20 output per one million tokens. Model selection should also consider answer quality, latency, context length, tool use, and retry rate. A cheaper model does not necessarily reduce TCO if it causes more retries or human escalation.
Google Vertex AI and Amazon Bedrock also vary by model, input and output, caching, batch use, and region. Compare them with the same test set, input/output volume, and quality threshold, then reconfirm official pricing when requesting a quote.
3. Implementation and operations: the less visible cost
Implementation and operations include discovery, conversation design, prompts, RAG, APIs, authentication, audit logs, tests, training, monitoring, and improvement. Thai sites often mix Thai, English, and Japanese, increasing the effort required to test the same intent across languages. PDPA-aware purpose, consent where relevant, retention, deletion, and processor management must also be designed.
The seven components of 12-month chatbot TCO
Requirements: define the boundary of chatbot deployment
Before deciding what the bot should answer, decide what it must not answer and what actions it may perform. FAQ response, order-status lookup, and repair-ticket creation require different identity checks and integrations.
At minimum, document users, languages, service hours, channels, expected questions, prohibited actions, response-time goals, handoff rules, retained logs, and accountable teams. If these are vague, a PoC may look successful while unplanned development accumulates before launch.
Channels: quantify Web, LINE, and Teams
For each channel, estimate monthly active users, conversations, sent and received messages, peak traffic, attachments, and notifications. Separate LINE plans and excess messages, Web hosting and WAF, and Teams licensing or tenant configuration.
Launching several channels at once increases work for UI, authentication, history, handoff, and testing. A practical first phase often starts with the highest-volume channel and a reusable API layer, then expands.
RAG and knowledge: uploading documents is not the finish line
RAG retrieves company policies, product specifications, maintenance procedures, or FAQs and grounds an answer in those sources. Cost depends less on raw document count than on quality and update methods. Duplicate versions, obsolete files, scanned PDFs, and permission differences require preparation and ongoing control.
Include inventory, OCR, chunking, metadata, permissions, retrieval evaluation, update frequency, source links, and deletion procedures in the estimate. For the broader architecture, see our enterprise RAG implementation guide for Thailand.
Business integration: what moves chatbot development cost most
ERP, CRM, MES, inventory, and ticketing integrations can change chatbot development cost substantially. Effort depends on whether the bot only reads or also writes, whether an API exists, whether identifiers align, and whether approval is required.
Even “answer inventory” needs definitions for site, warehouse, reserved stock, refresh time, unit, and permission. For write operations, the RFP should specify confirmation, duplicate prevention, authorization, audit logs, and rollback or failure handling.
Testing: measure dangerous failures, not only average accuracy
Average answer accuracy is insufficient for acceptance. Test frequent, high-impact, prohibited, ambiguous, multilingual, misspelled, unauthorized, prompt-injection, and downstream-outage scenarios.
Classify results by correctness, evidence quality, safe refusal, handoff, response time, and API success. NIST AI RMF organizes risk management around GOVERN, MAP, MEASURE, and MANAGE. The NIST Generative AI Profile adds considerations specific to generative AI. These documents do not automatically certify compliance, but they are useful checklists for reducing test gaps.
Operations: budget continuous improvement after launch
After launch, review unanswered and low-rated interactions, handoffs, API errors, latency, and model usage. When products, policies, or prices change, knowledge must change. Teams also need to respond to model and service updates.
An estimate with zero operational budget may become expensive only after a problem occurs. Quantify monthly improvement days, first-line incident support, reporting, backup, and access reviews, and define the intended service level.
PDPA and human handoff: design exceptions from the start
When personal data is involved, define purpose, data minimization, retention, access, deletion, and third-party or processor arrangements. Warn users not to enter unnecessary sensitive data, and design masking and log retention. Legal decisions belong with the company’s responsible people or advisers, not with the chatbot alone.
Human handoff is part of customer experience, not merely failure insurance. Define triggers such as low confidence, repeated clarification, complaints, personal information, contractual judgment, or urgency. Pass conversation history and a summary to the agent, and state out-of-hours behavior and response targets.
Illustrative assumption: a reproducible 12-month TCO calculation
The following is neither a price list nor a proposal. It is an illustrative assumption that shows how to compare options. All project figures remain in THB; USD vendor charges are kept separately and are not converted. Revalidate requirements and official prices for a real project.
Assumed company and usage
- Thai B2B manufacturer serving customer and distributor inquiries
- Two channels: Web and LINE
- Three languages: Thai, English, and Japanese
- 4,000 conversations per month, averaging five exchanges each
- 300 RAG documents, with 30 documents updated monthly
- Read-only CRM access and new ticket creation
- Human handoff during weekday business hours
- LINE Pro for illustration, assuming no more than 35,000 messages per month
Illustrative initial costs
| Item | Quantity × assumed rate | Amount |
|---|---|---|
| Requirements and process design | 15 person-days × THB 18,000 | THB 270,000 |
| Conversation and UX design | 10 person-days × THB 18,000 | THB 180,000 |
| RAG build and document preparation | 20 person-days × THB 18,000 | THB 360,000 |
| Web and LINE connection | 12 person-days × THB 18,000 | THB 216,000 |
| CRM lookup and ticket creation | 20 person-days × THB 18,000 | THB 360,000 |
| Authentication, logging, and PDPA controls | 12 person-days × THB 18,000 | THB 216,000 |
| Multilingual, safety, and load testing | 18 person-days × THB 18,000 | THB 324,000 |
| Training and production transition | 5 person-days × THB 18,000 | THB 90,000 |
| Initial subtotal | 112 person-days | THB 2,016,000 |
The calculation is 15+10+20+12+20+12+18+5=112 person-days, then 112×18,000=THB 2,016,000. The rate is only an illustrative assumption; it is neither a market benchmark nor TOMAS TECH’s quotation rate.
Illustrative monthly costs
| Item | Monthly | 12 months |
|---|---|---|
| Budget allowance for cloud, search, and monitoring | THB 35,000 | THB 420,000 |
| Operational improvement | 4 person-days × THB 18,000 = THB 72,000 | THB 864,000 |
| Support and first-line incident response | THB 25,000 | THB 300,000 |
| LINE Pro | THB 1,780 | THB 21,360 |
| LINE Pro VAT at 7% | THB 124.60 | THB 1,495.20 |
| Monthly subtotal | THB 133,904.60 | THB 1,606,855.20 |
The 12-month TCO is 2,016,000 + 1,606,855.20 = THB 3,622,855.20. The monthly average is 3,622,855.20 ÷ 12 = THB 301,904.60. This illustration excludes USD AI-model usage, additional LINE messages, human-agent salaries, changes to existing systems, and external audits. Add them as separate lines once volume and contract conditions are known.
Why API cost should remain a separate table
If estimating OpenAI usage, multiply projected input and output token volumes by the corresponding rates and keep USD in USD. As an illustration, assume 100 million input tokens and 20 million output tokens in one month. GPT-5.6 Terra would be 100×2 + 20×12 = USD 440/month, while Luna would be 100×0.20 + 20×1.20 = USD 44/month. Quantities are expressed in units of one million tokens.
Do not select a model on this difference alone. If Luna increases retries or escalations on complex questions, operations may reverse the apparent saving. The PoC should measure quality and usage and compare routing—for example, a low-cost model for classification and a stronger model for critical answers. These amounts remain vendor-cost illustrations, not a total implementation estimate.
Required columns in a chatbot comparison table
A table containing only product names and monthly fees overvalues the cheapest-looking option. Include at least the following:
| Comparison area | What to confirm |
|---|---|
| Business scope | FAQ, search, lookup, registration, and approval boundaries |
| Languages | Whether UI, retrieval, answers, and evaluation are multilingual |
| Channels | Web, LINE, Teams, their costs and limitations |
| AI usage | Input, output, messages, and tool-call units |
| RAG | Documents, formats, updates, permissions, and evidence display |
| Integration | Systems, read/write scope, APIs, and exceptions |
| Security | Authentication, encryption, logs, environment separation, vulnerabilities |
| PDPA | Purpose, consent, retention, deletion, processors, and cross-border review |
| Human handoff | Triggers, context transfer, hours, and response target |
| Testing | Cases, languages, high-risk scenarios, and load |
| Operations | Monitoring, improvement time, SLA, and knowledge updates |
| 12-month TCO | Initial, fixed, usage-based, internal labor, and contingency |
Give every vendor the same usage scenario and ask for a sensitivity analysis at twice the volume. When a feature is “included,” confirm whether configuration, testing, and support are included too.
What to include in a chatbot implementation RFP
Objectives and KPIs
Replace “introduce AI” with measurable outcomes such as first-response time, self-service rate, human handling time, and answer quality. Do not rely only on revenue, which has many causes; use leading indicators that operations can measure directly.
Scope and assumptions
State teams, users, channels, languages, documents, integrations, service hours, and volume. State exclusions and request a change-control method when assumptions change.
Data and security
Ask about classification, location, access, logs, encryption, backup, deletion, subprocessors, and incident notification. Data sent to model providers, training-use terms, and retention vary by service and contract, so require current documentation with the proposal.
Deliverables and acceptance conditions
Define design documents, data-processing specifications, test evidence, operating procedures, administrator training, and the handover scope for source code or configuration. Acceptance should use a pre-agreed test set and thresholds, not merely a working demonstration.
Pricing response format
Separate initial fees, fixed monthly fees, usage charges, third-party licenses, channel charges, maintenance, additional-development rates, and internal work. Require tax treatment, currency, validity period, minimum term, termination, and data-return terms in the same response.
What a PoC should prove: build small, evaluate against production needs
A PoC reduces uncertainty; it is not just a screen demonstration. Limit it to one business process, one primary channel, and representative documents, while deliberately including difficult questions and failure conditions.
Recommended PoC sequence
- Agree success criteria with the business owner
- Build an evaluation set from real inquiries
- Remove or appropriately control personal data
- Implement minimal RAG, permissions, and handoff
- Measure the baseline
- Record the number and time of improvement cycles
- Estimate production volume and operational effort
- Decide Go, Conditional Go, or No-Go
If PoC spending is isolated too aggressively from production, more work will be discarded. Design data structures, evaluation sets, and logging for reuse. Conversely, a PoC does not need full production high availability or every integration.
Sample numerical acceptance criteria
These are also illustrative and should be adjusted to business risk.
| Metric | Example acceptance threshold |
|---|---|
| Accuracy on critical FAQs | At least 95% |
| Accuracy on general FAQs | At least 85% |
| Relevance of evidence links | At least 95% |
| Unauthorized information exposure | 0 cases |
| Safe refusal of prohibited questions | 100% |
| Successful human handoff | At least 98% |
| Business API success | At least 99% when downstream is healthy |
| Response time | 95th percentile within 8 seconds |
| Material quality gaps among Thai, English, and Japanese | Agree tolerance per language |
Freeze the denominator, partial-credit rules, and evaluation-set version. In addition to averages, define critical failures that trigger No-Go even once. If a threshold is missed, a Conditional Go may narrow the scope or require human approval.
Mistakes to avoid when requesting a chatbot development quote
Sending only “about 300 FAQs”
A count does not reveal file format, duplicates, languages, update frequency, or permissions. Share representative samples and known quality issues, and clarify responsibility for preparation.
Fixing the cheapest model in advance
Model prices change, and different work needs different quality. Require a replaceable model layer, an evaluation set, and usage monitoring so options can be compared after price changes. Our generative AI implementation cost guide for Thailand explains the broader cost structure.
Excluding human support as “existing operations”
Reviewing history, routing to a team, managing after-hours queues, and following up all require time. Measure current handling time and expected remaining work, then calculate both savings and additions.
Treating multilingual delivery as translation only
Product names, politeness, Thai tokenization, abbreviations, mixed English, and Japanese internal terms affect retrieval. Real users should evaluate each language, with critical questions tested separately.
Having no owner after the PoC passes
Assign owners for answer quality, knowledge updates, systems, PDPA, and human queues. A monthly review should prioritize unresolved questions and improvements.
Practical ways to reduce chatbot cost
Cost reduction is not only price negotiation. The strongest levers are narrowing the first scope, measuring quality, and reserving expensive processing for where it adds value.
- Start with the top 20–30 inquiry topics
- Remove duplicates and obsolete documents before RAG ingestion
- Use different models for classification, retrieval, and generation
- Cache repeated questions appropriately
- Begin integrations with read-only use cases
- Escalate low-confidence responses instead of guessing
- Stop unused features based on logs
- Set usage caps and alerts
Do not cut access controls, audit logs, deletion, or testing merely to reduce the initial number; post-incident remediation may cost more. TCO reduction means removing low-value processing and rework, not moving risk to another team.
FAQ: chatbot cost, comparison, deployment, and development
How much does a chatbot cost?
It varies greatly by scope. Compare 12-month TCO including requirements, RAG, integrations, testing, PDPA, human handoff, and operations—not only platform or API fees. The THB 3,622,855.20 figure in this article is an illustrative calculation, not a quotation.
Is a free or low-cost chatbot comparison enough?
It may be enough for fixed FAQs or a small internal experiment. Permissions, system updates, audits, multilingual service, or an SLA require more design. Test features in a free tier, then recalculate TCO under production conditions.
Where should a chatbot deployment start?
Analyze inquiry logs and select one high-volume process with clear sources and manageable risk. Define KPIs, exclusions, evaluation cases, human handoff, and an operational owner before the PoC.
What matters most in a chatbot development RFP?
There is no single universal item, but comparable proposals require scope, volume, data, integration, acceptance criteria, and a common pricing format. Clear exclusions and usage units make later additions easier to assess.
Can implementation cost be calculated from LINE pricing alone?
No. LINE Official Account pricing is a channel cost. AI models, RAG, integrations, development, testing, monitoring, and human support require separate evaluation, and “message” definitions differ among services.
Does the cheapest API model always reduce TCO?
No. Answer quality, retries, long inputs, tool calls, human escalation, and development work also matter. Measure quality and usage on the same test set and compare routing models by business task.
Can PDPA compliance be delegated entirely to a chatbot developer?
A provider can support technical controls, but the deploying company should participate in decisions about purpose, legal basis, retention, internal accountability, and data-subject handling. Involve legal counsel, a DPO, or another specialist when needed.
Conclusion: use price lists as inputs to an acceptable TCO
Monthly fees and API rates are only the starting point for a chatbot cost comparison. Quantify requirements, channels, RAG, business integrations, testing, operations, PDPA, and human handoff over the same 12 months. Align assumptions in an RFP, measure quality and volume in a PoC, and decide Go or No-Go against acceptance criteria. Keeping public price facts separate from illustrative calculations supports a decision based on business value and risk—not apparent cheapness.
TOMAS TECH can discuss chatbot comparison, RFP preparation, PoC design, and existing-system integration even at an early planning stage. Contact TOMAS TECH with your target process, languages, channels, and approximate volume.
References
- OpenAI API Models (accessed August 27, 2026; recheck model availability and specifications at implementation)
- OpenAI: Advancing the price-performance frontier with GPT-5.6 (published July 30, 2026; accessed August 27, 2026)
- Microsoft Copilot Studio pricing (accessed August 27, 2026; pricing varies by region)
- Microsoft Learn: Billing rates and management (accessed August 27, 2026; Copilot Credit consumption definitions)
- Google Cloud Vertex AI generative AI pricing (accessed August 27, 2026)
- Amazon Bedrock Pricing (accessed August 27, 2026)
- NIST AI Risk Management Framework (accessed August 27, 2026)
- NIST AI 600-1: Generative Artificial Intelligence Profile (published July 2024; accessed August 27, 2026)
- LINE for Business: LINE Official Account (accessed August 27, 2026; current Thai display, excluding 7% VAT)
