AI Projects Come With Two Bills: Development Cost and the Monthly Model Bill

AI project pricing has a trait no other software project has: there are two bills. One is the one-time development cost; the other is the monthly model-usage bill. Clients who sign after seeing only the first bill are often shocked by the cloud invoice in month three. This piece lays both bills on the table.

Bill one: development cost

Our AI PoCs (proof of concept) start at NT$100,000 (≈US$3,100). Why does an AI project need a PoC first? Because until it's built, anyone who guarantees an AI's performance is lying to you. That NT$100,000 buys a two-to-four-week experiment that shows its cards: run on your real data, in your real scenario, producing concrete answers — what's the accuracy, what does it cost, is it worth scaling.

Once the PoC passes, full development is priced by scope. Common ranges run roughly NT$400K to NT$2M:

• Workflow-embedded (customer-support assist, document summarization, report interpretation): NT$400K–1M
• System-integrated (connecting internal systems, multi-step agents, knowledge-base retrieval): NT$1M–2M
• Product-grade (public-facing service, high concurrency, full hardening): NT$2M and up

Bill two: model usage

This is the AI project's monthly fee, determined by three variables: number of calls × tokens per call × model unit price. Our own AI product lines pay this bill every day, so we know how to keep it down:

• Model tiering: cheap models for simple tasks, flagships only for the hard ones. Eighty percent of requests don't actually need the strongest model.
• Caching and batching: cache repeated system prompts, route non-urgent tasks through batch processing — unit costs get cut in half outright.
• Prompt trimming: every 1,000 tokens saved, multiplied by a million calls a month, is real money.

A well-designed mid-size application might run a few thousand to a few tens of thousands of NT dollars per month in model fees; badly designed, the same functionality can burn ten times that. A vendor who doesn't spell out usage estimates in the quote either has never operated an AI product, or doesn't intend for you to know in advance.

The most expensive thing in an AI project isn't the model — it's getting the order wrong: building big before validating.

When the right answer is to stop

Here's the honest part: roughly thirty percent of PoCs conclude "not worth doing" — the accuracy can't support the business requirement, or the labor saved doesn't cover the monthly model bill. That's not failure; that's spending NT$100K to dodge a seven-figure tuition fee. Our PoC design method is written up in why AI projects should start with a PoC — read it before setting a budget.

Three questions to ask before signing

1. What's the basis for the monthly model-fee estimate? How much does peak usage change it?
2. If the model's price rises or the model is revised, who absorbs it — and what's the response plan?
3. Where's the stop-loss point if results fall short — and is it written into the contract?

Our own subscription AI tools platform pays a model bill every day, so the numbers we estimate for clients are calibrated against our own P&L. To evaluate whether your scenario fits AI and how to allocate the budget, book a free 30-minute consultation — we'll say "not worth doing" out loud when that's the answer.