When a manufacturer in Thailand or ASEAN evaluates generative AI, the first visible figures are usually a per-user subscription or an input/output token rate. Those figures are useful, but they are not the generative AI implementation cost. A defensible budget also covers data preparation, RAG and integration, identity and security, governance, evaluation, training, monitoring, and support. This guide turns current official prices into a 12-month TCO framework that management, procurement, IT, and operations can use in an internal proposal or RFP.
Prices and availability are current as of 26 August 2026. They can vary by country or region, tax, foreign exchange, reseller, contract terms, and product changes. Every calculation below is an illustrative planning example, not a market quotation or a vendor implementation estimate. Re-quote before contracting.
Generative AI implementation cost: choose the smallest route that fits the workflow
There are three practical buying routes. The right question is not which architecture is the most advanced. It is which smallest route can satisfy the workflow’s acceptance conditions.
| Route | Best fit | Main visible cost | Cost that is often missed |
|---|---|---|---|
| Managed seats | Writing, summarization, translation, research, and broad personal productivity | Users multiplied by monthly subscription | Adoption, training, administration, and information handling |
| Controlled API/RAG workflow | Repeatable work requiring evidence, a defined output, or approval | Model usage plus data and retrieval | Evaluation, observability, connectors, human review, and support |
| System-integrated agent or custom build | Work in which ERP, MES, QMS, or another system must be acted on with an audit trail | Integration and operational engineering | Permission boundaries, exception handling, duplicate prevention, and rollback |
Managed seats are usually the smallest route for broad human-reviewed work. API/RAG is appropriate when the organization needs repeatability, grounding, and controlled outputs. Custom integration should be reserved for cases where system actions and auditability create measurable value. Starting small does not mean ignoring governance; it means applying governance in proportion to the workflow.
If you need to establish benefit before selecting a tool, see our AI implementation impact measurement guide. For the data and retrieval design behind the second route, see enterprise RAG implementation for Thailand.
Why sticker price and 12-month TCO are different
A vendor price page describes what the vendor charges for a defined product. It does not automatically include the work performed by the buyer. In a factory group, that work may be spread across procurement, corporate IT, local IT, information security, legal, HR, quality, operations, and a local support partner. If those efforts remain outside the AI budget, a project can appear inexpensive while consuming substantial internal capacity.
The answer is not to put every unknown into one large contingency. A useful TCO separates cost drivers, owners, quantities, unit prices, assumptions, evidence, and change conditions. Management can then see which amount changes when the number of users grows, when a new data source is added, when retention changes, or when the workflow moves from drafting to system action.
TCO is not a promise that every future cost can be predicted. It is a decision model. It tells the organization what must be re-estimated and re-approved when scope changes. It also prevents a temporary manual task in a pilot from being treated as free labor in production.
The seven ledgers in a defensible 12-month TCO
Seats and accounts
For managed products, record eligible users, contract months, annual or monthly commitment, and what is already included in an existing suite. Use the people who perform the target workflow, not the total headcount merely because it is easy to obtain. Include account provisioning, individual identity, role changes, and deprovisioning.
Seat arithmetic is simple, but value is not created by assigning an account. Track whether the account is used for the agreed workflow and whether its outputs pass the acceptance condition. A raw activity rate can encourage unnecessary usage, so it should not be the only success measure.
API and model consumption
Separate input tokens, output tokens, model, retries, and evaluation calls. Consumption is affected by the context attached to every prompt, retrieved documents, output length, and failed or repeated calls. The model page rate is only one part of the ledger.
Start from workload: eligible users multiplied by tasks per day, working days per month, and months. Then state the assumed input and output tokens per task. Keep the formula editable so that real pilot usage and a new vendor price can be substituted without rebuilding the budget.
Data, RAG, and integration
A grounded assistant requires document collection, deduplication, version control, access inheritance, chunking, indexing, refresh, and deletion. An ERP or MES connection also requires interface specifications, test environments, master-data alignment, failure handling, retries, and audit records.
More documents do not automatically produce a better answer. Old versions, conflicting procedures, and files without clear ownership can reduce quality and weaken access control. Budget data stewardship as an operating responsibility, not merely as a one-time upload.
Identity, security, and governance
This ledger includes single sign-on, permissions, logging, retention, deletion, information classification, acceptable-use rules, and incident response. OpenAI states that Business data is not used for model training by default. That statement is relevant, but it does not by itself prove that a specific contract, configuration, storage path, or connected application meets the buyer’s requirements.
OpenAI announced on 19 August 2026 that eligible frontier-model API customers can use Zero Data Retention. Eligibility and architecture must be confirmed; it should not be assumed for every customer or every component.
ETDA’s Generative AI Governance Guideline for Organizations emphasizes privacy, information security, organizational impact, and governance. NIST AI 600-1 applies Govern, Map, Measure, and Manage to generative AI risk. These are cost-bearing workstreams: people must make decisions, document them, review evidence, monitor outcomes, and correct failures. They are not decorative labels in a policy document.
Evaluation and monitoring
Generative outputs can vary while serving the same intent, so a simple happy-path software test is not enough. Build evaluation cases with expected answers, prohibited answers, required evidence, escalation cases, and the business consequence of an error. Decide which results can proceed and which must return to a person.
In production, monitor consumption, failure, latency, missing evidence, user-reported issues, and changes to models or source data. Budget ownership for updating the evaluation set, stopping a faulty workflow, investigating the cause, and approving a restart.
Change management and training
Training is not limited to button clicks. Users must understand which workflows are allowed, what data must not be entered, how to verify an answer, and who remains accountable for the final output. Where Japanese, Thai, English, and Vietnamese material coexist, localized examples and an agreed terminology map may be required.
Managers also need a way to assess business outcomes rather than praise AI usage in isolation. Front-line teams need a safe route to report errors. Link training to the target workflow and evaluation results instead of treating a single briefing as completion.
Operations and support
After launch, the organization will handle questions, permission changes, source updates, model changes, incidents, invoices, and vendor coordination. Roles between headquarters, the Thailand entity, local IT, business owners, and suppliers must be explicit.
Include first-line intake, technical investigation, business decisions, security decisions, and vendor escalation. Also budget the exit: data deletion, log retention, and the portability of prompts, evaluation cases, and other project assets.
Current managed-service price examples
The following official figures were accessed on 26 August 2026. They illustrate different product structures and should not be treated as directly comparable without normalizing what is included.
| Example | Published price and condition | Comparison rule |
|---|---|---|
| OpenAI Business | USD 20 per user per month with annual billing; USD 25 monthly | API usage is separate; add administration, training, and controls |
| Microsoft 365 Copilot | USD 30 per user per month paid yearly; a qualifying Microsoft 365 plan is required | Confirm the base plan and agent consumption or Azure/capacity arrangements |
| Google Workspace Business Standard | USD 14 per user per month with annual commitment; USD 16.80 flexible at access time; includes Gemini capabilities | Treat it as a workspace bundle, not an AI-only price |
Using the published annual OpenAI Business price, a 30-user, three-month pilot is 30 × 20 × 3 = USD 1,800. This is a seat-cost example before tax, FX, training, administration, and integration. A 100-user, 12-month deployment is 100 × 20 × 12 = USD 24,000, again before all non-license costs.
At the published Microsoft 365 Copilot price, the seat arithmetic for 100 users is 100 × 30 × 12 = USD 36,000. The qualifying Microsoft 365 plan and any agent consumption must also be considered. Do not compare that amount directly with ChatGPT Business or Google Workspace until the included products and eligible population are normalized.
Google Workspace Business Standard includes workspace capabilities as well as Gemini capabilities. A decision to replace or standardize a productivity suite is different from a decision to add an AI-only service. An internal proposal should identify which cost replaces an existing contract and which cost is incremental.
All prices and availability above are current as of 26 August 2026 and may vary with country or region, tax, FX, reseller, contract, and product changes. Obtain a Thailand-specific quotation before approval.
API cost formula: do not confuse token spend with implementation budget
On the model page accessed on 26 August 2026, OpenAI lists GPT-5.4 at USD 2.50 per one million input tokens and USD 15 per one million output tokens. GPT-5.4 mini is listed at USD 0.75 input and USD 4.50 output. Prices can change and should be checked again in the RFP and contracting process.
The basic arithmetic is:
Annual model cost = annual input tokens / one million × input rate + annual output tokens / one million × output rate
Consider the explicit example in the research brief: 100 users × 2 tasks per day × 20 working days per month × 12 months = 48,000 tasks per year. Assume 1,500 input tokens and 500 output tokens per task. Annual consumption is 72 million input tokens and 24 million output tokens.
At the GPT-5.4 mini list rates above, 72 × 0.75 + 24 × 4.50 = USD 162 per year. That USD 162 covers model tokens only. It excludes retrieval and storage, connectors, observability, evaluation, human review, support, security, and change management. It must never be presented as the implementation budget for 100 users.
The lesson is not simply that tokens are inexpensive. It is that the work around the model can be a decisive part of TCO. A short, repetitive task may use few tokens while still requiring controlled data, evidence, approvals, exception handling, and operating support.
Hidden costs in a Thailand manufacturing environment
Multilingual data alignment
Japanese headquarters standards, English system fields, Thai shop-floor instructions, and Vietnamese sister-site material can use different terms for the same part, process, defect, or machine. Before retrieval or automation, map approved terms, codes, versions, and owners.
The review must cover more than fluent translation. Item codes, equipment IDs, process names, and quality terms must refer to the same business object. Budget the local reviewers and the update process so that production does not drift away from a successful pilot.
PDPA and data mapping
Where personal or confidential information may be involved, map what is sent, where it is processed or stored, who can access it, and when it is deleted. This article does not offer a legal conclusion. The company should obtain appropriate legal and security review based on its contracts, purpose, classification, and stakeholders.
A statement that data is not used for training does not automatically answer retention, logging, administrative access, or connected-app questions. Confirm each path with settings and evidence.
OT and IT separation
A factory-network or control-system use case cannot be designed like an ordinary chat rollout. Risk changes depending on whether the model only proposes an answer, drafts a transaction for review, or executes an action in MES or ERP.
For system action, use least privilege, approval, input validation, safe failure, and auditability. If the value comes from completing an action rather than drafting text, exception handling and the approval trail become central budget items.
Tax, FX, and procurement
A USD price on a global website may not equal the Thailand entity’s payable amount. Confirm tax, exchange rate, card or invoice arrangements, reseller terms, annual commitment, cancellation, renewal, and price-change terms.
The quotation should state its currency, validity, rate-change notice, usage limits, and treatment of excess consumption. For API workloads, define not only an alert but also the operating response when consumption is abnormal.
Local support and approval evidence
A headquarters-designed rollout can stall if the Thailand operation lacks a local route for questions and exceptions. Identify the business owner, local IT, security contact, and supplier escalation path, including support language and responsibilities.
In quality, purchasing, production, or other approval-sensitive work, preserve the AI proposal, cited evidence, human edits, and final decision. A log is useful only if the organization knows who reviews it and when.
Compare the three routes from the workflow backward
When managed seats are enough
Managed seats fit email drafting, meeting summaries, translation, research, and document structuring where a person reviews the result. They can be distributed quickly, but unstructured experimentation can scatter value. Provide approved workflow examples, restrictions, and a review method together.
Budget seats, administration, training, policy, and support. After the pilot, measure quality in the target workflow and the capacity that was actually redeployed, not merely assigned accounts or generated messages.
When a controlled API/RAG workflow is justified
API/RAG fits question-answer drafting, procedure search, or standardized report generation where evidence and format matter. Model usage is accompanied by permission-aware documents, indexing, evaluation, monitoring, and human review.
RAG is not a guarantee of truth. Acceptance conditions must cover a missing document, an outdated source, conflicting sources, and a query that should be escalated. Operations must keep source ownership and refresh under control.
When system integration or a custom build is justified
Integration fits work in ERP, MES, QMS, maintenance, or another system where the AI changes business state. The design must address interface contracts, permissions, approval, duplicate execution, reversal, audit, and incidents.
Custom does not merely mean a branded screen. It is an investment in safely embedding a controlled action into existing operations. If that value is not yet demonstrated, a managed product or a read-only RAG pilot can test the workflow with less complexity.
For governance, permissions, and adoption across subsidiaries, read our overseas subsidiary generative AI implementation roadmap.
A 30/60/90-day pilot with acceptance gates
Day 30: freeze the workflow and baseline
Define eligible users, workflow, input, output, and the person accountable for the final result. Select which current time, external spend, overtime, rework, or quality issue will be measured. A narrow workflow with an explainable acceptance condition is more useful than an open invitation to test everything.
Prepare prohibited data rules, permissions, retention, user guidance, and an issue route. Build evaluation examples that include expected answers, prohibited answers, and cases that must return to a person.
Day 60: evaluate real data and update cost assumptions
Replace planning assumptions with actual users, tasks, input/output consumption, retrieval volume, and review effort. For seats, investigate non-use. For API, inspect failures and retries. For RAG, check whether the right evidence was retrieved.
Separate problems that require a different model from problems that require better data or workflow design. A larger model should not be the only remedy; input quality, retrieval, output constraints, and approval can be more relevant.
Day 90: decide the production gate
Do not promote the pilot solely because users liked it. Check the agreed quality, security, audit, operations, cost, and benefit conditions. If a condition is not met, retain the options to narrow scope, remediate, or stop.
Recalculate the 12-month TCO with production users, data growth, new connections, support, and renewal assumptions. Convert every temporary manual pilot activity into an owned production task or remove the dependency.
RFP checklist for generative AI implementation support
| Area | Question for the RFP | Evidence to request |
|---|---|---|
| Scope | Workflow, exclusions, input, output, and human approval points | Process map and responsibility boundary |
| Data | Location, retention, deletion, permissions, training use, and data path | Data flow, settings, and contract terms |
| Non-functional | Response, availability, capacity, incident, and change notice | Service terms and operating procedure |
| Security | Identity, least privilege, logs, classification, and incident response | Administrative controls and response model |
| Evaluation | Evaluation set, pass condition, evidence, error, and re-evaluation | Test plan and result format |
| Cost | Initial, recurring, usage, excess, renewal, tax, FX, and contract | Quotation with assumptions |
| Exit | Deletion, asset transfer, logs, and portability | Termination procedure and deliverables |
| Support | Language, intake, priority, and escalation | Role matrix and contact route |
Ask not only for a price, but also for the assumption that drives each price. Every claim that an item is included should state its scope, limit, and exclusion. Confirm whether the organization receives the evaluation assets, operational documentation, and source artifacts needed to avoid lock-in.
When comparing generative AI consulting, separate workflow selection, TCO, data mapping, governance, PoC evaluation, RFP support, implementation, adoption, and handover. The goal is not the installation of a named product. The goal is a workflow that meets an agreed business acceptance condition.
ROI and payback without double counting
Use the formulas in the brief:
Annual gross benefit = eligible users × tasks per month × minutes saved × adoption rate × loaded labor cost per minute × 12 × realized-quality factor
Net annual benefit = gross benefit − annual run cost
Payback months = initial implementation cost ÷ monthly net benefit
Time saved is not automatically cash. Confirm whether capacity is redeployed to valuable work or whether external spend, headcount, or overtime actually changes. Do not count the same saved time as both labor reduction and productivity benefit.
Quality also needs an explicit measurement method. A faster draft that requires more review and correction may lower net benefit. Keeping adoption rate and realized-quality factor separate distinguishes a well-performing tool that is not used from a widely used tool that does not meet quality requirements.
Separate initial implementation from annual run cost for decision clarity. Data preparation, first integration, initial evaluation, and training design may be initial work; seats, API, monitoring, refresh, support, and governance continue. The appropriate accounting treatment should be confirmed under the company’s own rules.
How to build the internal 12-month TCO proposal
Begin with one sentence defining the workflow. “Introduce generative AI” is too broad. “Draft evidence-linked answers from Thailand plant maintenance procedures for an engineer to approve” identifies input, output, user, and reviewer.
Choose the smallest of the three routes. Do not build system action if a managed-seat pilot can validate the hypothesis. Use API/RAG when evidence and controlled format are required. Compare integration when the business value depends on action and auditability.
Fill the seven ledgers. Even when an amount is unknown, record the quantity, pricing rule, owner, evidence date, and event that requires a new quotation. An explicit uncertainty is more defensible than a blank cost presented as zero.
Finally, attach the 30/60/90-day acceptance gates and stop conditions. Decide in advance what happens if information controls fail, quality misses the threshold, an operating owner is unavailable, or net benefit cannot be supported.
Frequently asked questions
What does generative AI implementation cost include?
It includes more than a subscription or token rate. Use seven ledgers: seats, API/model usage, data/RAG and integrations, identity/security/governance, evaluation/monitoring, change management/training, and operations/support. The USD 1,800, USD 24,000, USD 36,000, and USD 162 figures in this article are explicit arithmetic examples, not total implementation quotations.
How should a company implement generative AI?
Start with a defined workflow, input, output, eligible users, accountable reviewer, and acceptance condition. Map data and restrictions, select the smallest route, and update the TCO through the 30/60/90-day gates. Product comparison comes after the company knows what it must accept.
Should a company choose generative AI seats or an API?
Use managed seats for broad human-reviewed productivity work. Use API/RAG for repeatable, evidence-linked, controlled outputs. Consider integration when system action and auditability create measurable value. The smallest route that meets the acceptance conditions is usually the soundest starting point.
What should be checked in a generative AI implementation support quote?
Separate initial, recurring, usage, evaluation, data, integration, security, training, support, and exit costs. Confirm included work, limits, exclusions, rate-change conditions, ownership of deliverables, and local-language support.
What should generative AI consulting cover?
Providers may offer different scopes. Ask separately about workflow selection, TCO, data mapping, governance, PoC evaluation, RFP design, implementation, adoption, and operating handover. A strong engagement defines acceptance and stop conditions, not only a product configuration.
If API token spend is low, is the implementation inexpensive?
Not necessarily. In the 48,000-task example, GPT-5.4 mini model tokens alone total USD 162 per year. Retrieval, storage, connectors, observability, evaluation, human review, support, security, and change management remain outside that figure.
Are the published prices the final prices for Thailand?
Not necessarily. The prices and availability here were accessed on 26 August 2026. Country or region, tax, FX, reseller, contract, and product changes can alter the payable amount. Obtain a current local quotation.
Summary: budget for an accepted workflow, not a price page
Generative AI implementation cost is the sum of seven controlled ledgers, not a single subscription or token line. Compare managed seats, API/RAG, and system integration from the workflow backward. Select the smallest route that meets quality, information, audit, operational, cost, and benefit conditions, then update the 12-month TCO at each acceptance gate.
TOMAS TECH can help while you are still defining a Thailand or ASEAN 12-month TCO and RFP. You can contact us before selecting a product to structure the workflow, data path, evaluation gates, and operating ownership.
Sources
- https://openai.com/business/pricing/
- https://openai.com/business-data/
- https://openai.com/index/offering-zero-data-retention-for-frontier-models/
- https://developers.openai.com/api/docs/models/gpt-5.4
- https://www.microsoft.com/en-us/microsoft-365/copilot/pricing/enterprise
- https://workspace.google.com/solutions/ai/
- https://www.etda.or.th/th/pr-news/AI_Gov_Anual.aspx
- https://www.etda.or.th/th/pr-news/aigc_Driving-Trust_AI_Governance.aspx
- https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence
