2026-09-15
56
Every fourth quarter, Taiwanese-owned plants in Southeast Asia budget for next year, and this year there is a new line: AI. Gartner forecasts worldwide AI spending of US$2.59 trillion in 2026, up 47%, yet most enterprises prefer to adopt AI incrementally inside software they already use. This article lays out three ways to budget for AI (subscriptions, projects, foundation upgrade), four hidden costs usually missed, and why upgrading your existing ERP to host AI is an investment, not an expense. #ITBudget #AIInvestment #SubsidiaryFinance #ERPUpgrade #AiGP
For manufacturers that have run plants in Vietnam or Thailand for years, the fourth quarter has a fixed task: next year's budget. Labor, equipment, rent, tax advisors; each line has years of reference data. In the 2026 budget meeting there is one line with no precedent: AI.
The message from headquarters in Taiwan is clear: next year we use AI. The finance manager's questions are practical: which account? How much is reasonable? If staff simply subscribe to a few AI tools on their own, does that even belong in the budget?
Start with the bigger picture. Gartner's forecast published on May 19, 2026 puts worldwide AI spending at US$2.59 trillion in 2026, up 47%, with infrastructure accounting for more than half. The same release notes that most organizations show limited appetite for using AI to drive disruptive change and instead favor tactical initiatives with incremental gains in efficiency and productivity, expanding their use of generative AI already embedded in existing applications rather than buying entirely new solutions (source: Gartner press release, May 19, 2026).
That observation matters for a Taiwanese plant: you do not need to build something new for AI. The point is to get AI working inside the systems you already run.
In practice, AI spending in Southeast Asian plants falls into three patterns, each with different expectations and risks.
The first is tool subscriptions. A few tens of dollars per person per month lets sales draft English emails, finance tidy spreadsheets and engineers look up specifications. Lowest barrier, fastest results, but what you buy is personal productivity, not company capability. Worse, when every department subscribes separately and the bills sit on different credit cards, the year-end total is usually higher than expected, and nobody knows where the data went.
The second is project fees. An outside team builds a custom AI application for one scenario, say reading customer orders automatically or generating reports. It is targeted, but every scenario is a new project: data has to be connected again, permissions set again, and maintenance and model updates after the project often have no owner.
The third is a foundation upgrade. Upgrade the existing ERP to a version that can host AI, so AI assistants read documents, review reports and answer questions inside the system you already use. The initial outlay is larger, but what you buy is a base on which no future scenario starts from zero, and data, permissions and audit trails follow the ERP's existing framework.
None of the three is right or wrong in the abstract. The question is whether you want AI to stay at the individual level or become a company capability.
Whichever pattern you choose, the budget errors are rarely in the visible license fees. They are in four hidden costs.
1. Training. Tools get bought, staff do not know how to use them, are afraid to, or use them wrongly. That is the most common cause of failure, and with every departure the training restarts, which hurts most in high-turnover plants.
2. API and usage. Generative AI is mostly priced by usage. Individual subscriptions look cheap, but once AI is wired into daily processes with thousands of calls a day, cost gets out of hand, and most companies have nobody managing token usage and allocation.
3. Trial and error. The largest hidden cost of assembling AI in-house is starting from nothing: data formats that do not match, the wrong scenario, permissions that cannot be controlled halfway through. This time and effort never appears on an invoice, but it is real spending.
4. Security and compliance. Staff pasting customer data, contracts and cost sheets into public models is a risk many owners have not registered. Vietnam's Cybersecurity Law implementing decree (Decree 53/2022/ND-CP) requires cross-border service providers to establish a local presence and store data within Vietnam (subject to the latest official guidance); where data lives, who can access it and whether it can be audited all come back as compliance cost.
Add those four in and "subscriptions are cheapest" often stops being true.
From a finance viewpoint, AI tool subscriptions are classic operating expense: pay monthly, stop anytime, and when you stop there is nothing left. A foundation upgrade is different. It turns the orders, inventory, costs and customer records accumulated in ERP over fifteen years into assets AI can call directly, and that capability stays with the company regardless of staff turnover.
As our earlier article on the AI upgrade path noted, most established manufacturers still run Workflow GP3 or GP4, which are 32-bit architectures; the upgrade to AiGP moves them to 64-bit. For AI to read large volumes of data and compute in real time, the foundation must keep pace. That is a security and performance matter as well as an AI matter, so the upgrade's value should not be measured only by how many people AI saves.
Compare in-house adoption with a vendor upgrade. In-house, you first have to ensure data accuracy and structure, guess at use cases yourself and control API traffic and fees on your own. With a vendor upgrade, the vendor handles AI integration without changing existing processes or data structures, the model source is reliable, confidential data stays inside, and token usage is managed. None of this shows directly on a budget sheet, yet it decides whether next year's money becomes a capability or a pile of unconnected subscriptions.
Concrete benefit still comes back to scenarios. Take order entry: if a plant receives 50 orders a day and manual entry takes about 3 minutes each, that is 2.5 hours a day; with AI reading the image or email and generating the document, each order takes about 6 seconds and the day is done in 5 minutes (based on advisory experience; actual figures vary by industry and company, for reference only). Numbers like these, not visions, are what to set against the investment when budgeting.
1. Take stock before choosing a pattern. List every AI subscription and paid account the company currently holds, total them, then ask: has this spending become company capability? Most companies find the scattered subscriptions already fund a more structured approach.
2. Estimate benefit by scenario, not by vision. Pick one or two high-frequency, high-load scenarios, estimate "how many more orders the same people can handle," and include training, usage, trial-and-error and compliance in the cost.
3. Make data sovereignty part of the budget. Whatever pattern you choose, ask up front how models are connected, where data is stored, and how permissions and audit trails work. For plants operating in Vietnam and Thailand this question cannot be skipped.
How you fill in the AI line next year depends on what you want the company to have three years from now. If you are preparing the 2027 budget and want to assess the investment and return of upgrading your existing ERP to AI capability, submit a request form and a consultant familiar with your industry and country will assess it with you.
A1: Yes, and they should be managed centrally. Scattered subscriptions are not only a cost issue but a data-flow and security issue; putting them in the budget makes the full picture visible and lets you judge whether a more structured approach is warranted.
A2: Look at three years, not one. Project fees require new data integration and permission setup for every scenario; a foundation upgrade is a one-time outlay after which later scenarios are layered on the same framework. Include the four hidden costs of training, usage, trial and error and compliance so the comparison is fair.
A3: One high-frequency scenario with the most complete data, such as order entry or report analysis, and make sure it runs inside your existing system rather than as another island. Small but structured beats broad but fragmented for showing results next year.
鼎新數智購
6 Followers
延伸閱讀
鼎新數智購
6 Followers