11/02/2026

A 30-Person Firm in Taiwan Needed AI: Three Quotes, One Contract, NT$1.4 Million Apart

Why the cheapest Copilot rollout, the middle RAG pilot and the on-premise box are not versions of the same purchase, priced line by line for 2026

SeaPRwire – The company makes fasteners. Thirty people, a factory in Gangshan and an office that answers quotation requests from German and American buyers in two languages, four thousand active part numbers, and a stack of supplier PDFs that nobody has ever read systematically. The owner wants a system that turns an incoming inquiry into a draft quotation with the right part number, the right tolerance and the right price band, and he wants it before the busy season.

That is the whole brief. It is also, in Taiwan in 2026, the point where three completely different purchases get presented as the same thing.

I spent two weeks building the three cost stacks for this firm, using published licence prices, the government’s own consulting rate card, and the arithmetic that determines whether a seventy-billion-parameter model fits on one machine. The gap between the cheapest and the dearest path, over the three years from 2026 to 2029, is NT$1.41 million. The gap between what each path delivers is larger than that.

The Trouble, in One Purchase Order

Here is the actual work, because the cost model is meaningless without it.

A buyer in Stuttgart sends a PDF drawing and a sentence: can you supply this in A4 stainless, 12,000 pieces, delivered in eight weeks. Someone in Gangshan opens the drawing, reads the tolerance callout, searches a spreadsheet for a matching part number, checks a second spreadsheet for the last price quoted to a similar buyer, checks a third file for the current material surcharge, and types a quotation into a template. If the part is new, they add a line for tooling. If the buyer is new, they add a payment-terms paragraph that only two people in the company know how to word.

Twenty minutes when the part is familiar. Two hours when it is not, and the two hours is where the margin goes, because a slow quote loses to a fast one and a wrong tolerance loses the account six months later.

Everything below is an attempt to cut those two hours. That is the only outcome that matters, and any vendor who cannot tell you how their product touches those two hours is selling something else.

Why the Obvious Fix Does Not Work

The obvious fix is to buy seats. Microsoft 365 Copilot sits inside Outlook and Word, which is where the work already happens, and the licence is a line item rather than a project.

It fails on one detail that nobody mentions in the demo: the drawings are scans. A PDF exported from a CAD system carries text. A PDF that a buyer printed, stamped and photographed does not. Copilot reads what is in Microsoft Graph, and what is in Microsoft Graph is a folder of scanned PDFs plus four thousand rows spread across three Excel files that disagree with each other about part-number formatting.

The second failure is permissions. Copilot answers with whatever the signed-in user can already open. If the quotation template sits in a folder that the whole company can read, Copilot will quote from the version that was superseded in March. Cleaning that up is not an AI project. It is three weeks of someone deciding who owns which folder, and it has to happen before the licences go live, not after.

The third failure is arithmetic. Let me do it, because it is the number that ends the argument.

Table 1: Path one, eight seats of Copilot for a thirty-person firm

Line Basis Year one (NT$) Years two and three (NT$)
Licences at list 8 seats x NT$675 x 12 64,800 64,800
Administration 4 hours a month at NT$350 16,800 16,800
Internal learning 96 hours at NT$350 33,600 0
Total 115,200 81,600
Three-year total 278,400

The published price has three versions in 2026 and they are not close. In the July 2026 Taiwan price list, Copilot Business lists at NT$675 per user per month, is promoted at NT$578.58 on annual billing, and costs NT$810 on monthly billing. The enterprise SKU is NT$965. Eight seats on monthly billing is NT$77,760 a year in 2026 before anyone learns anything, which is more than the annual saving from one avoided wrong quotation.

That is the honest case for path one: it is cheap, it is reversible, and it will make eight people slightly faster at writing emails. It will not read a scanned drawing.

7.4% and 86% Describe the Same Country

Before choosing, it is worth knowing what everyone else chose, and Taiwan’s numbers are unusually good at showing why this decision is confusing.

The Ministry of Economic Affairs’ 2025 white paper on small and medium enterprises, published in 2026, found that 7.4% of Taiwanese SMEs had adopted or were planning to adopt AI. The OECD average cited alongside it in the same 2026 comparison is 39%. A separate survey by the Artificial Intelligence Foundation put the overall AI index for Taiwanese industry at 46.32 in 2026, up from 36.77 in 2025, with information and communications at 52.7 but traditional manufacturing at 43.87 and retail at 44.86. A 2026 Institute for Information Industry MIC survey found 28% of manufacturers had adopted AI.

Then, in October 2026, Meta published a survey conducted by YouGov between 6 August and 2 September 2026 across nine Asia-Pacific markets: 86% of Taiwanese SMEs said they had started using AI, and 94% planned to apply it to marketing and content within twelve months.

Seven point four percent and eighty-six percent. Both are about Taiwanese small and medium enterprises in 2026. Both are probably measured correctly.

The difference is the verb. “Using AI” means at least one person on at least one device has opened a chatbot. “Adopted” means a process changed. A factory owner who asks ChatGPT to polish a German email is using AI. A factory whose quotation process runs through a retrieval system is adopted. Taiwan has a lot of the first and very little of the second, and the 12-fold gap between the two numbers is the most useful statistic in the country for anyone selling into this market, because it tells you where the purchase actually is. It is not in the licence. It is in the process.

Path One: Buy Seats

What you get for NT$278,400 over three years is eight people who draft faster. What you do not get is a system that knows your part numbers.

The failure is not subtle and it is not the model’s fault. Copilot’s answer quality is bounded by the quality of the corpus it is allowed to read. A firm with 4,000 part numbers in three spreadsheets and a folder of scans has a corpus problem, and a seat licence does not solve corpus problems. It solves writing problems.

There is a version of path one that works, and it is worth naming because it is what most Taiwanese SMEs actually do: buy three seats, give them to the people who write the most English, and accept that the quotation logic stays manual. That costs NT$104,000 over the three years to 2029 including learning time, and it delivers a real but small return. The mistake is buying thirty seats and expecting the quotation problem to disappear.

Path Two: Buy a Working System

The middle path is to have someone build the retrieval layer: parse the drawings, normalise the part numbers into one index, encode the pricing rules, and put a chat interface on top that cites which document each answer came from.

The cost driver here is not the model. It is the person-days, and Taiwan publishes a number for those. The Republic of China Information Service Industry Association’s 2026 cost-estimation standard for outsourced information services lists a project manager including management consulting at NT$223,083 per person-month for the first category, and NT$289,641 for the third. Divide the 2026 rate by 22 working days and a person-day is about NT$10,140 before expenses and tax.

A three-month pilot at four person-days a month is twelve days, NT$121,680 in 2026 money, plus a project fee. Market quotes for small integration work in Taiwan start around NT$50,000 and go up with scope. Add sixty internal hours at NT$350 for the people who have to explain the pricing rules, and year one lands at NT$192,700.

Table 2: Path two, a three-month retrieval pilot with an integrator

Line Basis Year one (NT$) Years two and three (NT$)
Integrator days 12 days at NT$10,140 121,680 81,120
Project fee Market floor for small integration 50,000 30,000
Internal hours 60 hours at NT$350 21,000 0
Total 192,700 111,100
Three-year total 414,900

This is the only path that produces the thing the owner asked for: a draft quotation with a cited source. It is also the only path with a dependency, because the integrator owns the pipeline and the firm owns the data, and the handover document is the difference between an asset and a subscription.

What it does not buy is a model that speaks good German. A retrieval system built on a mid-size open model will quote the right part number and write a clumsy sentence. For a quotation, that is the correct trade. Nobody loses an order because the email reads awkwardly. They lose it because the tolerance is wrong.

Path Three: Buy a Machine

The third quote will be the largest and it will arrive with the words “data sovereignty” in the covering email. It is a two-GPU server running an open-weight model inside the factory, and the arithmetic of whether it fits is public.

A model’s weight memory is parameters times bits per parameter divided by eight. In the 2026 open-weight catalogue, seventy billion parameters at sixteen bits is 140 gigabytes, which does not fit on any single accelerator a small firm can buy. At four bits it is 35 gigabytes, which fits on one 48-gigabyte card with room for a small context, and at eight bits it is 70 gigabytes, which needs two. Add the KV cache, the operating system, the vector index and the headroom that keeps latency tolerable under three concurrent users, and the honest configuration for a 70B model is two 48-gigabyte cards in a chassis that can cool them.

In Taipei channel pricing in October 2026 that chassis, with the CPU, memory and NVMe to feed it, sat around NT$1.25 million. That is an estimate from component prices, not a quotation, and it excludes the room, the UPS and the person who will be on call.

Table 3: Path three, a 70B-class model on premises

Line Basis Year one (NT$) Years two and three (NT$)
Server with two 48GB accelerators Channel estimate, October 2026 1,250,000 0
Integration 25 days at NT$10,140 253,500 0
Internal hours 120 hours at NT$350 42,000 0
Operations Power, part-time admin 48,000 48,000
Total 1,593,500 48,000
Three-year total 1,689,500

Three years of path three costs NT$1,689,500. Three years of path one, from 2026 to 2029, costs NT$278,400. The difference is NT$1,411,100, which is the number in the headline and roughly one year of gross profit for a firm this size.

There is a fourth thing in the quote that is not a line item: the model choice is constrained before the price is. A Taipei deployment consultancy’s 2026 guide notes that open-weight models from Chinese vendors are typically excluded at procurement review in Taiwan’s public sector and regulated industries, even when run entirely offline, because the constraint is supply-chain security rather than data egress. For a private fastener exporter that constraint does not bind. For anyone selling to a Taiwanese state-owned enterprise or a bank, it removes the cheapest capable models from the list and pushes the hardware requirement up a tier. The procurement rule, not the benchmark, decides the invoice.

The Number That Decides It: Scanned Pages

Between paths two and three there is a factual question that settles most of the argument, and it is not about model quality. It is about how many of the 4,000 part drawings in the 2026 archive are scans.

A text-layer PDF can be parsed in milliseconds by open-source tools. A scan needs optical character recognition, and OCR on a stamped engineering drawing with a tolerance callout in a rotated annotation is where accuracy falls apart. If 80% of the 2026 corpus is scans, the retrieval system will cite the wrong revision and the project fails for a reason that has nothing to do with the model. If 80% is CAD-exported text, the same system works.

So the first deliverable of any pilot is not a chatbot. It is a count. Take 200 drawings at random in the first week of the project, run them through a text extractor, and record how many return usable text. That exercise costs one person two days, NT$20,280 at the 2026 rate card, and it tells you more about feasibility than any benchmark table.

Which Path, Under Which Condition

The decision rule I would actually use, written as instructions rather than principles.

  • If fewer than 30% of the drawings carry a text layer, stop. Neither path two nor path three will work, and the money belongs in digitising the archive first. That is a scanning and indexing project, not an AI project, and it costs less.
  • If the firm already lives in Microsoft 365, has clean permissions, and the quotation logic is simple enough to fit in a paragraph, buy eight seats and stop there. NT$278,400 over three years is a fair price for a 10% speed gain on writing.
  • If the drawings are mostly text, the part numbers can be normalised into one index, and someone in the firm can own the rules, buy the pilot. NT$414,900 over three years is the cheapest path that changes the process.
  • If a customer contract requires that data never leaves the building, the on-premise box is not optional and the NT$1.69 million is a compliance cost rather than an AI cost. Budget it as such, and put it next to the ISO audit fee, not next to the software line.

Two of those four conditions are about documents and one is about contracts. None of them is about which model scores higher on a leaderboard, which is the thing every vendor wants to talk about.

What I Would Sign

For this firm, in this order: pay for the 200-drawing count first, then the pilot, then decide whether the box is needed.

The reason is not that the middle path is cheaper than the box, although it is. The reason is that the pilot, run in 2026, produces the artefact that makes every later decision cheaper: a normalised index of 4,000 part numbers with the source document attached to each row. That index is worth having even if the chatbot is abandoned, because it is the thing that lets a new salesperson answer a tolerance question without walking to the factory floor. The box does not produce it. The seats do not produce it. The integrator produces it, and it stays.

Taiwan’s own numbers say most firms will buy the seats, because 86% are already “using AI” and a licence is a decision one person can make on a credit card. The 7.4% who changed a process did something harder, and the gap between those two groups is where the next five years of competitive difference in Kaohsiung’s fastener cluster will be decided. It will not be decided by which model they picked. It will be decided by whether anyone counted the scans.

Data Basis and the Four Costs the Quote Leaves Out

The three stacks use published prices wherever a published price exists and state the assumption wherever one does not.

  • Copilot Business at NT$675 list, NT$578.58 annual promotional and NT$810 monthly, and the enterprise SKU at NT$965, are from Microsoft Taiwan’s pricing page as checked on 29 July 2026, excluding 5% business tax.
  • The NT$223,083 first-category person-month rate is from the Information Service Industry Association’s 2026 cost-estimation standard for outsourced services, inclusive of salary, management fee and public fee, exclusive of other direct costs and tax. Dividing the 2026 rate by 22 working days gives NT$10,140.
  • NT$350 per internal hour is an assumption: a NT$56,000 monthly salary plus 19% employer burden, divided by 176 hours, rounded down. Replace it with your own number and every table moves proportionally.
  • The 70B memory arithmetic uses parameters times bits divided by eight. The NT$1.25 million chassis is an October 2026 component-price estimate, not a vendor quotation, and it is the softest number in this piece.
  • The 7.4% SME adoption rate is from the Ministry of Economic Affairs 2025 white paper published in 2026; the 86% figure is from a Meta-commissioned YouGov survey of 3,983 SME owners across nine Asia-Pacific markets, fielded 6 August to 2 September 2026; the 46.32 industry AI index is from the Artificial Intelligence Foundation’s 2026 survey.

Four costs are missing from all three quotes and they are the ones that decide the outcome. The permission cleanup before any seat goes live, which is three weeks of someone senior. The OCR pass over the archive, which is priced per thousand pages and quoted by nobody. The person who owns the pricing rules and will be interrupted forty times in the first month. And the cost of the wrong quotation that ships before the system is trusted, which is the only line in the model that has a customer’s name on it.

Sources

This article was prepared by the expert team at Global Infinity Labs, drawing on insights and data provided through its MCP data services.

About Global Infinity Labs

Global Infinity Labs is an AI innovation company specializing in MCP data services.

Email: support@globalinfinitylabs.com



source https://newsroom.seaprwire.com/press-releases/finance/a-30-person-firm-in-taiwan-needed-ai-three-quotes-one-contract-nt1-4-million-apart/
at
Labels: , ,