A training center goes digital, part 5: once the portal is stable, AI plugs in over an API — test generation and online Q&A

This series follows one training center through several phases of going digital. The journey is built from real implementation experience in this industry — the stages and challenges should feel familiar if you run this kind of school. Each part covers what was built, what was deliberately not built, and what evidence justified moving to the next phase.

Why AI appears only now

If you have read along this far, you have noticed something: AI appears very late in this story. That is not an oversight. It is deliberate.

The first four phases set up AI’s two prerequisites:

1. A unified entry point and clean data — enrollment, students and classes are structured, so AI has context to lean on;
2. A portal that is genuinely used — students log in, teachers post homework and grades; the portal is not a shell.

Only now does the center have the conditions to talk about AI. And the AI it talks about is not desktop software on a teacher’s computer — it is a capability the portal plugs in over an API that students use by logging in, and the school controls how and when it appears.

Over an API means one program calls another program’s capability through a defined interface. In this story the AI is not a desktop application on a teacher’s computer — it is a capability the school’s portal plugs in over an API, and students use it inside the portal.

The first AI capability: drafting test questions

The center’s pain was real: every term’s test was written by hand, different teachers had inconsistent styles, and revising was time-consuming. AI’s first step went here:

• The teacher feeds AI this textbook, this lesson’s key points, and last test’s question types;
• AI drafts a first pass at a test: questions, options, reference answers;
• The teacher reviews question by question: fixes errors, cuts anything off-style, adds their own;
• Once confirmed, the questions appear in the student portal’s test.

The point is not “AI produces a great test in one go,” but AI turns a blank page into a draft the teacher can edit. Writing time dropped from two or three hours to under half an hour, while teacher review kept the quality. The school saved repetitive labor, not teacher judgment.

The second AI capability: online Q&A

After-class student questions were still scattered — whoever was online answered, responses were slow, and the same kinds of questions recurred: make-up, homework submission, course rules.

• The portal gives students a question entry point, and questions first pass through AI:
  • rule-type questions (schedule, leave, make-up, homework submission) — AI answers directly from the school’s FAQ and course material;
  • questions beyond the rules, needing teacher judgment — AI does not force an answer, it routes to a teacher;
• Teachers get a pending queue and handle only what genuinely needs a human;
• Every answer is recorded, so the school can see where students are actually stuck.

Response speed went from “whoever happens to be online” to “almost immediate,” and teachers went from “answering every question” to “answering only the ones that need judgment.” The staff-hours saved could be counted: the time admin and teachers spent on repeat questions each week dropped visibly after the AI plug-in.

What the school decided before letting AI go live

• Pilot, then expand. Try it in one course for one term, watch real usage, then decide whether to widen it.
• All AI-generated content passes through teachers. Tests need a human pass; Q&A routes to a human at the rule boundary. AI assists; it does not replace.
• No promise that AI improves on its own. The school was explicit with students: this is a supporting tool; learning judgment still comes from teachers.
• Respect personal data. Students’ questions and grades stay inside the portal and do not leak out carelessly.

What this phase deliberately was not

• Not fully automated teaching. No AI lectures, no AI grading replacing teachers, no AI deciding who levels up.
• Not a chatbot open to the public. AI Q&A lives inside the learner portal and serves enrolled students — it is not a widget on the public website answering anyone.
• Not a general-purpose AI assistant. It runs on the school’s material, FAQ and course rules — not an anything-goes tool.
• No extra charge just because AI exists. It is an independent capability of the portal, not an automatic “value-add” that ships with a website package.

What evidence decided whether to expand

After one pilot term, the center looked at only these things:

• whether the teacher hours from test draft to live test actually dropped;
• whether response time for student questions and the share routed to a human measurably improved;
• whether teachers refused to review — if teachers felt AI output was unusable, the pilot failed;
• whether students actually used these features in the portal instead of going back to asking teachers on LINE.

If those numbers held, AI moved from “pilot” to “regular capability.” If they did not, the school went back to human — the portal and manual flow could still hold on their own, and AI was a bonus, not a lifeline.

Part 6 looks back across the whole road: entry point, data, portal, AI — and how each step was decided by the evidence of the step before.

---

The boundary matters: AI test generation and online Q&A plugged into the learner portal over an API are the school’s own independent system project — not part of a standard website package, and not an automatic value-add that ships with a website. AI output must be reviewed by teachers and does not replace teacher judgment; student data is used inside the portal and respects personal-data protection requirements.