Short answer: When a vendor uses AI to write code, a fair overall discount is usually only 10–20%, not two-thirds off. Only about 40% of project hours (coding, testing, documentation) speed up noticeably with AI, while requirements, design, communication and launch barely change. A quote half the price of others more often means testing was cut.
A 40-person bakery chain wants a membership and booking system. Vendor A quotes NT$600,000 and vendor B NT$320,000, saying "we use AI for all our coding, so we're cheaper." The general manager asks: if AI really is three times faster, is A overcharging? And where exactly do B's NT$280,000 in savings come from?
Myths, busted
The three most common beliefs all mistake coding for the whole project.
- Myth 1: AI is 3x faster, so the quote should be a third. Truth: coding is only about 30% of project hours. In GitHub's 2022 experiment, developers using Copilot finished one task 55% faster, but that was a small, well-defined task, not a whole project.
- Myth 2: teams using AI are always faster. Truth: METR's 2025 randomized study found experienced developers working on large codebases they knew well took 19% longer with AI tools. Speed-ups depend on the task and the cost of review.
- Myth 3: AI-written code is just as maintainable. Truth: code nobody reviews line by line often has inconsistent structure and naming, which raises the hours for the next change.
The framework: the AI-accelerable hours formula
Fair discount = Σ (phase share of hours × AI savings rate for that phase). Ask the vendor to split the quote into seven phases, then apply this table:
| Phase | Typical share of hours | AI savings rate | Effect on total |
|---|---|---|---|
| Requirements and planning | 15% | 0–5% | 0–0.75% |
| UI design | 15% | 0–5% | 0–0.75% |
| Coding | 30% | 30–45% | 9–13.5% |
| Testing | 10% | 15–30% | 1.5–3% |
| Documentation and handover | 5% | 40% | 2% |
| Project management and communication | 15% | 0% | 0% |
| Deployment and launch | 10% | 0–5% | 0–0.5% |
| Total | 100% | about 12.5–20.5% |
Coding, testing and documentation make up about 45% of hours, and roughly 40% is noticeably accelerated. The savings rates are ScriptWalker's estimates from its own projects; substitute your vendor's figures.
Three typical scenarios
The clearer and more repetitive the requirements, the more AI can save.
- Brochure website (10 pages, NT$80,000): mostly layout and content; fair discount about 5–10%, or NT$4,000–8,000.
- Membership and booking system (NT$600,000): many forms and APIs; fair discount about 12–18%, or NT$70,000–110,000. Vendor B's quote is 47% lower, far beyond what AI explains.
- ERP integration rework (NT$1.2M): understanding the old system and repeated checks with accounting; fair discount about 8–12%.
Hidden costs
AI saves hours but adds new costs, and a quote should show them.
- Code review: each hour of AI output needs about 0.3–0.5 hours of human review.
- Tools: about NT$600–6,000 per engineer per month in AI subscriptions.
- Security and confidentiality: if source code or client data goes to outside services, it needs extra settings and contract terms.
- Rework risk: quotes that skip testing often add 20–30% more hours in post-launch fixes.
- Your own time: requirements and acceptance hours do not shrink with AI; your internal contact still needs 4–8 hours a week.
Vendor KPI scorecard
Score eight items 0–2 each; talk price only at 12 or above.
- Quote broken down by phase hours
- Can say which phases use AI
- A fixed code review process
- Test scope written into the contract
- States whether code and data go to outside services
- A specific list of handover documents
- Clear warranty period and defect definitions
- Willing to run a small paid pilot
ScriptWalker's options, and who we are not for
We use AI for coding, testing and documentation, and our quotes list hours by phase so you can check the math. Engagement models: one-off project, monthly maintenance, advisory and small pilot. We are not a good fit if you:
- only want the lowest price and don't care about testing or review
- want a system with undefined requirements done in two weeks
- require all code to be AI-generated with no human review
- have no one internally to take part in acceptance
Start-up playbook
Use 90 days to test whether the quote holds.
- Month 1: run a paid 4–6 week pilot on one small module and require weekly time logs.
- Months 2–3: compare actual hours with the quote and confirm review and testing happened.
- Day 90: decide on the full project based on the pilot; a discount within the formula's range is reasonable.
Decision checklist
If more than three answers are "no," don't sign yet.
- ☐ The quote is split into seven phases
- ☐ The discount falls within the formula's range
- ☐ You know who reviews AI output
- ☐ Test scope is in the contract
- ☐ You know whether data goes to outside services
- ☐ Source code delivery and ownership are clear
- ☐ Warranty and defect definitions exist
- ☐ There is an internal acceptance contact
- ☐ A small pilot is possible first
- ☐ The scorecard reaches 12 or more
Frequently Asked Questions
If a vendor uses AI, what discount is fair?
About 10–20% for most projects. Coding and testing speed up; requirements, design, communication and launch barely change.
A quote is half the price of others. Is that AI?
Not necessarily; often testing and documentation were cut. Ask for hours by phase and see which part is missing.
Is AI-written code cheaper to maintain?
Not necessarily. Unreviewed AI code can make later changes slower. It depends on the review process.
How do I check that a vendor uses AI well?
Ask three things: who reviews AI output, how testing works, and whether source code or data is sent to outside services.
Next step
Holding two quotes that are far apart? Send them to us. ScriptWalker (a Taiwanese Laravel/Flutter custom development studio) will break them down with this formula for free and walk you through the differences in a 30-minute online call.
- Email: [email protected]
- Phone: 0916-224-047
- LINE: @ufv9089p