Short answer: Estimate an AI feature's monthly cost with cost per call × monthly volume × a 1.5 safety factor; most SMEs land between NT$1,000 and NT$15,000 a month. What gets out of control isn't the API bill but unowned prompt tuning and human review hours.
A 20-person hardware wholesaler paid an agency NT$180,000 for a LINE quoting assistant, and the quote listed only development. Month one brought a NT$9,800 API bill; month two, with more customers, NT$23,000. The agency said it had opened the account and would add a 15% handling fee every month. Only then did the owner realize nobody had asked what the feature would cost to run.
Myths the Industry Gets Wrong
Bottom line: more than half of an AI feature's total cost sits outside the development fee.
- Myth 1: AI is cheap, a few cents a call. Truth: each call is cheap, but customer-facing usage grows with the business, and stuffing a whole catalog into every prompt can multiply cost tenfold.
- Myth 2: build it and forget it. Truth: platforms retire older models on a schedule, such as OpenAI's deprecations, and each change needs retesting.
- Myth 3: the best model is the safest. Truth: most quoting and classification tasks work fine on mid-tier models, often several times cheaper; see OpenAI and Anthropic pricing.
The Core Framework: Monthly AI Cost Formula
Bottom line: measure cost per call, multiply by monthly volume and a 1.5 safety factor, and that's your budget cap.
Monthly budget = cost per call × monthly volume × 1.5
- Cost per call: run 50–100 real questions and average them; don't do mental math from list prices.
- Monthly volume: estimate from the last 3 months of messages, inquiries or documents.
- 1.5: covers peak season, retries and prompts growing longer. Going past 1.5x means checking for waste.
Add one fixed item: 4–8 maintenance hours a month to tune prompts and review wrong answers.
Three Typical Scenarios
Bottom line: same formula; volume and review hours decide the result.
| Scenario | Cost per call (assumed) | Monthly volume | Monthly budget | Main risk |
|---|---|---|---|---|
| Restaurant LINE booking Q&A | NT$0.3 | 3,000 | about NT$1,350 | Weekend peaks double volume |
| B2B quoting assistant | NT$1.5 | 4,000 | about NT$9,000 | Whole catalog stuffed into prompts |
| Clinic patient-education Q&A | NT$0.8 | 2,000 | about NT$2,400 | Review hours far exceed API fees |
Hidden Costs, Fully Listed
Bottom line: beyond the API bill there are at least five ongoing costs.
- Human review: spot-checking 5–10% of conversations, about 4–10 hours a month.
- Prompt maintenance: 2–4 hours a month as products and prices change.
- Model retirement migrations: 1–2 a year, 8–16 hours of retesting each.
- Vector database and hosting: about NT$500–3,000 a month if there's a knowledge base.
- Pass-through markups: vendor-held accounts often add 10–20%.
Vendor KPI Scorecard
Bottom line: score these 8 items 0–2 each; under 11 means ask more questions.
- Provides measured cost-per-call data
- Account opened in the client's name
- Budget caps and alerts configured
- Test set and error-rate metric exist
- Explains how model retirements are handled
- Separates build fees from monthly running costs
- Defines handoff-to-human rules
- States how long conversation data is kept
ScriptWalker's Options and Where We're Not a Fit
Bottom line: we quote AI features as two lines, build and operations, and the account always sits in the client's name.
ScriptWalker (台灣 Laravel／Flutter 客製化開發工作室, a Taiwanese Laravel and Flutter custom development studio) maps its four engagement models like this: project-based covers the build and the test set; monthly maintenance (from NT$8,000) includes prompt tuning and usage reports; advisory (from NT$20,000) reviews AI quotes you already have; full outsourcing suits companies building a website, app and AI together.
Where we're not a fit:
- You want AI to answer medical or legal judgments with no human involved.
- Fewer than 300 conversations a month, where standard LINE auto-replies suffice.
- You can't share real questions to build a test set.
Transition Playbook
Bottom line: in the 90 days around launch, turn estimates into measurements.
- Month 1: open the account in your name, set a cap, measure cost per call on the test set.
- Months 2–3: review usage and wrong answers weekly, tune prompts, test a cheaper model.
- Day 90: compare actuals to the estimate and decide whether to expand or adjust maintenance hours.
Decision Checklist
Bottom line: if more than 3 of these 10 are "no," renegotiate before signing.
- ☐ Do you know the measured cost per call?
- ☐ Have you estimated 12 months of volume?
- ☐ Is the API account in the company's name?
- ☐ Is a budget cap set?
- ☐ Is there a test set of 50+ questions?
- ☐ Is someone responsible for reviewing wrong answers?
- ☐ Does the contract state monthly maintenance hours?
- ☐ Do you know who handles model retirements?
- ☐ Is there a human handoff?
- ☐ Is conversation data retention written down?
FAQ
Who should pay the AI API bill?
The company should pay the AI platform directly under its own account. Vendor-paid billing invites markups and hides your usage data.
Can monthly API spend be capped?
Yes. Major AI platforms offer usage limits or budget alerts; set them before launch and agree what happens when the cap is hit.
Does an AI feature need maintenance fees?
Yes. Models get updated or retired and prompts must follow the business, so budget 4–8 maintenance hours a month.
Next Step
Bottom line: send us your AI quote or the feature you have in mind, and we can estimate a monthly range in 30 minutes.
In a free 30-minute consultation we'll run the formula with you and point out what belongs in the contract:
- Email: [email protected]
- Phone: 0916-224-047
- LINE: @ufv9089p