2026TransGlobe Life
AI agent interface design and build
We designed the conversational interface and workflows of an AI agent for insurance sales, so frontline staff actually use it.
- AI agent
- Conversational UI
- Financial services
AI customer service Enterprise AI chatbots AI answers the common questions,
and hands the rest to people.
We build an enterprise knowledge base from your own documents, FAQs and system data. The AI answers from that data and cites its source; orders, refunds and complaints go to your support team. Start with the busiest support channel, prove the answer rate, then expand.
Free requirements review · Reply within 1 business day · NDA available
JoinX experience
- projects delivered
- 520+
- clients served
- 220+
- years of development
- 10+
- legacy systems taken over
- 100+
What is AI customer service, and how is it different from a chatbot?
A traditional chatbot replies from pre-written keywords and menus, and fails when a question is phrased differently. AI customer service first retrieves the relevant passages from an enterprise knowledge base, then uses generative AI to compose an answer with its source (RAG), so it understands different phrasings. When it needs to look up an order or update data, it calls existing systems through APIs. Success depends less on model size than on accurate content, correct permissions and a smooth handoff to people.
Real projects Conversational AI and search we have built
Conversational AI we have built, not just demos.
These are projects we actually developed, from an AI agent for insurance sales to JoinX GPT, which answers questions on our own site every day.
2024Fubon Life
AI coaching system demo
We built a working demo of an AI coaching system for life insurance in about three weeks, as the basis for an internal evaluation.
- Conversational AI
- Fast validation
- Insurance
2026GoodParty
AI search added to an existing platform
We added AI search to the GoodParty platform, which we built and still maintain, reusing its existing data and admin instead of replacing the system.
- AI search
- Existing system
- Long-term maintenance
In-houseJoinX GPT
The AI assistant on our website
The assistant at the bottom right of joinx.co answers from our services, process and articles, and hands anything else to our sales team.
- Knowledge base
- Human handoff
- Used daily
2020Education company
Facebook page support bot
A support bot for an education company's Facebook page that answers common questions about courses and enrolment.
- Social support
- FAQ
- Auto-reply
* Projects without a client name are described by industry under confidentiality agreements.
What it does What AI customer service can handle
From answering questions to working with your systems.
Most companies start with the first item, then move on to lookups and integration once the answer rate is stable.
Knowledge base answers
Product guides, policies, FAQs and manuals become a knowledge base. Answers cite their source, and updates only need re-indexing.
Human handoff
Order disputes, refunds, complaints and anything the AI is unsure about go to an agent, with the conversation attached so customers never repeat themselves.
One core for web, app and LINE
The answering core is an API, so the website chat, your app and a LINE official account all use the same knowledge base and give the same answer.
Order and member lookups
Order, member or booking data is read through APIs with your existing permissions. Lookups only; changes stay with a person.
Internal AI advisor
The same knowledge base serves staff: sales checks product specs and new hires look up procedures without asking a senior colleague.
Logs and answer rate
Every question, answer and source is logged. Wrong answers and handoffs are flagged so the knowledge base can be improved.
Which one Choosing an approach
Rule-based chatbot, AI customer service or AI agent?
Newer is not always better. It depends on how fixed your questions are and whether system data must change.
| Approach | Best for | Watch out for |
|---|---|---|
| Rule-based chatbot | Few, fixed questions such as opening hours or store locations | Fails when phrasing changes; keyword upkeep takes effort |
| AI customer service (knowledge base + RAG) | Many questions and phrasings, with answers already in documents | Documents must be accurate and current, or the AI repeats outdated content |
| AI agent | Looking up data across systems, creating orders or multi-step tasks | Fine-grained permissions, action logs and human approval for important actions |
Process How we roll it out
Test with real questions, launch only when the answers are right.
Every phase ends with something you can accept. If the proof of concept is not usable, we stop there.
- STEP 01
Questions and data
2–4 weeks
Review past support logs and common questions, find where the answers live, decide what the AI must not see, and define a correct answer.
- STEP 02
Proof of concept
4–8 weeks
Build the knowledge base and answer flow for one channel and a set of real questions, then measure answer accuracy, source accuracy and handoff rate.
- STEP 03
Integration and launch
2–4 months
Connect the website, app or LINE, set up handoff, permissions and monitoring, and integrate order or member systems where needed.
- STEP 04
Operations
Ongoing
Improve the knowledge base from conversation logs, tune the rules, and track cost and answer rate. The same team stays responsible.
Pricing Price ranges
How much does AI customer service cost?
Estimates for custom projects in Taiwan in 2026. The final figure depends on data volume, channels and systems to connect, and acceptance criteria.
| Phase | Price range | Includes | Timeline |
|---|---|---|---|
| Assessment | NT$100k–300k | Question and data review, risk assessment, architecture advice | 2–4 weeks |
| Single-use-case PoC | NT$300k–800k | One channel, limited data, tested under human supervision | 4–8 weeks |
| Production launch | NT$800k–2.5M | Permissions, integration, handoff, monitoring, testing and deployment | 2–4 months |
| Cross-department, multi-system | NT$2.5M and up | Multiple knowledge bases and AI agents, cross-system flows and full governance | 4+ months |
AI model API and cloud hosting costs are billed by actual usage. Post-launch maintenance is a separate agreement.
FAQ Common questions
AI customer service FAQ
How much does AI customer service cost?
For custom projects in Taiwan in 2026: an assessment is about NT$100k–300k, a single-use-case proof of concept about NT$300k–800k, a production launch about NT$800k–2.5M, and cross-department, multi-system work usually starts at NT$2.5M. AI model API and cloud costs are billed by usage.
Will the AI make up answers?
It can be wrong, so we design safeguards: answer only from the knowledge base with sources, say it does not know and hand off when nothing is found, and send sensitive topics straight to an agent. We test accuracy with real questions before launch and keep tuning from the logs afterwards.
Will our data be used to train AI models?
No. We use enterprise AI services that, by contract, do not train on your data. Passwords, personal data and other sensitive content never go into AI tools, and knowledge base access follows user roles.
Can it connect to our LINE official account or app?
Yes. The answering core is an API, so the website chat, your app and a LINE official account all use the same knowledge base. Order or member lookups connect to existing systems through APIs.
What is the difference between AI customer service and an AI agent?
AI customer service mainly answers from a knowledge base. An AI agent can call systems within granted permissions and complete multi-step tasks such as creating an order or changing a booking. Most companies start with AI customer service and consider letting AI change system data once accuracy is stable.
Do we need an AI consultant first?
Not necessarily as a separate engagement. Our first phase is an assessment that reviews your questions and data and explains what suits AI and what does not. You can also start with the AI readiness check on our site.
Get started Start with one support channel
Bring a batch of real support questions, and we will see whether AI can answer them.
No full spec needed. Tell us which channel gets the most volume, what customers ask most, and where the answers live today.
- 1A sales manager replies within 1 business day
- 2Sign an NDA first, then discuss your workflow and data
- 3We discuss workflow and scope first, and quote once the spec is clear