Bringing AI into your Thailand office

Staff want to use ChatGPT or Claude for work. Management and IT don't want internal documents leaving the company. Neither side gives, so staff quietly use their personal accounts instead. Nothing unusual for a Japanese-affiliated company in Thailand.

What we've been proposing is placing the AI itself inside the company, reachable only from the internal LAN or via staff accounts. Since conversations never leave the company, staff can feed in documents that contain confidential material as they normally would. The setup is already assembled on our side, and we can run a live on-site demo for companies in Thailand looking at internal AI adoption.

What staff are asking for

Over the last six months, the requests we've heard at client offices look roughly like this:

• Summarising meeting minutes
• Draft translation between Japanese, Thai and English
• Drafting and reviewing code and SQL
• Drafting sales decks and internal proposals
• Q&A against internal manuals ("where's that policy again?")
• First drafts of customer email replies

Every one of these speeds up work. The catch is that the material staff want to feed in often includes things you don't want leaving the company — customer names, unit prices, contract terms, org charts, unannounced initiatives, personal data.

What snags external generative AI at work

If you're using ChatGPT or Claude on a personal or Team plan for work, a few things are worth checking:

• Whether input data feeds future model training (default off on enterprise plans, but behaviour varies with contract and options)
• Conversation logs stay on the vendor side. In an account compromise or misfire, that data is outside the company
• Whether PDPA-covered or NDA-covered material is being sent to a third-party cloud that isn't in scope of your existing contracts
• Personal-account use is invisible — the company can't audit what's being sent where

Corporate plans lower most of these concerns. Even so, there are industries and projects where "customer data leaves for any external cloud at all" is unacceptable — medical, finance, defence-adjacent, or cases where Japan HQ has explicitly banned it.

Internal AI adoption as an option

Open-source AI models are published with their weights available. You can run them on a server (or a GPU-equipped dedicated machine) your company operates. Conversations don't go out, and logs stay on your side.

A few years ago, "open-source AI only works properly in English" was the reality. Recent models have moved Japanese into workable-for-business territory. Thai also reaches a usable level depending on the model, and Japanese-affiliated companies in Thailand can now realistically consider internal AI adoption.

What the setup looks like

We place a GPU-equipped dedicated machine (a workstation-class PC at smaller scales, a server as scale grows) at your office, and expose it only to the internal LAN or to staff accounts. The outbound path to the internet stays closed by default.

• Hardware: GPU-equipped PC or server, chosen against how many people will use it concurrently and how large a model you want to run
• OS: Ubuntu (Linux)
• Internal chat UI: staff use it from the browser with a ChatGPT-like feel
• Internal document search: internal manuals, minutes and past project docs are indexed, and the AI answers with citations
• Access control: internal LAN only, or restricted to staff accounts. Remote use goes through VPN

You don't need a server room or a dedicated rack — a scale that fits in a corner of the office is a fine starting point. Which AI model we pick depends on the intake: mainly Japanese, handling Thai documents, or including code assistance each change the best fit.

Cost range

The number that drives the price is how many people use it at the same time. Even with the same headcount, an environment where 10 people are always on differs from one where only 3 people occasionally are — the GPU class you need shifts.

• Small (5-10 concurrent users, summarisation and translation focus): 200,000-400,000 THB upfront
• Medium (20-30 concurrent users, adds internal Q&A and document drafting): 700,000-1,200,000 THB upfront
• Large (50+ concurrent users, several use cases in parallel): from 2,000,000 THB upfront

The figure is the sum of the dedicated machine, GPU, setup, and tuning effort. Network work, UPS and rack expansion are quoted separately if needed.

Since this is a one-off purchase, monthly running cost from there is just electricity and maintenance. Unlike external AI pay-per-use or enterprise per-seat pricing, adding registered staff doesn't move the monthly bill. What can push cost up is when concurrent users go past what the current hardware supports — in that case we add GPUs. The entry barrier, conversely, is the upfront investment and having a physical spot to put the machine.

When it doesn't fit

There are situations where signing up for an external AI service outright is the better call:

• Uses that consistently need frontier-model accuracy (the top-tier models from ChatGPT or Claude). Open-source releases haven't fully caught up here
• Only a handful of users, a few dozen hours a month. An enterprise plan is clearly cheaper
• No one on-staff or on-retainer who can keep up with model selection, updates and tuning. Once stood up and left alone, it goes stale within a few months

Live demo

We've already assembled the setup and confirmed it running. On-site we can run a PoC / demo shaped like this:

• Short intake on your intended uses (minutes summarisation, internal Q&A, translation and so on)
• Show a setup matched to those uses actually running
• Feed in a batch of your non-confidential documents and check internal-document-search accuracy
• Actually use the chat UI and judge accuracy, response and feel
• Decide whether to push to production, adjust the setup, or fall back to an external AI service

Starting from "try it first, see how much it can do," a couple of weeks is enough to have the decision material in hand.

Related: Seven cybersecurity blind spots for Japanese-affiliated companies in Thailand / Security in Thai system development / Starting a Thai system development project with a PoC

If you're considering internal AI adoption in Thailand and want staff to use AI at work without confidential data leaving the company, please get in touch. Visits to our Bangkok office or on-site demos at your premises are both possible.