Technology & AI
Tuesday, 3 November 2026, at the University of Applied Sciences Upper Austria in Linz. I was the second of four speakers in the WTUC session “Responsible Science and Artificial Intelligence: Shaping the Future Through Collaboration”, co-chaired by Prof. Dr.-Ing. Klaus Pantke and Prof. Dr. Sven Quadflieg, both are Vice Presidents of University of Applied Sciences Upper Austria. I gave a quick introduction of my own background in aeronautics and mechanical engineering, so code is not foreign to me, but my day job these days is internationalisation, and about half the room that morning worked in international offices too. So the talk put the two halves of my working life together: an AI team, a team of AI agents, for the international office.
The work we all know
The first thing I said was not about AI at all. I described our office, and I could see people nodding, because international offices are much the same everywhere. The work runs on emails, student complaints, signed documents, record keeping, visitor logistics and annual deadlines. It is repetitive, it is seasonal, it is bound by rules, and, because people get tired and calendars get full, it is prone to human error. That last part is exactly what a robot is good at taking off our hands.
A manager and a team
What we built are virtual agents that live in the cloud. I find it easiest to think of them as specialists. Instead of talking to each one separately, I gave them a structure: one manager agent and a set of subordinates. I talk to the manager, and the manager hands the work to whichever team member knows that job, one for partnership agreements (our MOUs and MOAs), one for student mobility opportunities, one that writes travel handbooks for our delegations.
At this point a hand went up. “Sorry, MOA?” Memorandum of Agreement, I explained, the agreement we sign with a partner to say that we want to exchange students and will recognise the credits they earn there. “Sorry about the jargon.” It was a fair reminder that the international office has its own dialect, and I had slipped into it in front of a room where only half of us speak it.
I built all of this with Claude Code, by chatting with it in plain English and describing the task. That is the part people expect to be difficult, but it is in fact the easy part. The difficult part is everything you have to spell out: the specific rules, the instructions, the constraints, what the agent must never do, and exactly which tools, websites and mailboxes it is allowed to touch. The agents live in separate boxes, one for the international office, another for my academic work, another for my other administrative roles. In the internationalization box there is Agent MOU, Agent Mobility, a focus group agent and a few others, all reporting to one supervisor agent.
Demonstration one: OMSIN
Part of our job is student mobility, getting students to go abroad, which mostly means making sure they hear about the scholarships and grants that exist. For years our staff kept a Notion database of outbound programmes by hand: the Taiwan Ministry of Education scholarships, Franco-Thai, WTUN and many more. It was carefully kept and well organised, and the students almost never looked at it. In daily life, we have to compete with TikTok and Instagram or messaging apps for the students’ attention for a share of their screen time.
So I made an agent called OMSIN, an acronym for Outbound Mobility Service Information Node. ออมสิน in Thai means savings, which seemed right for an agent whose job is to find students money.
OMSIN works in two directions. Coming in, it is semi-automated: a staff member can drag and drop a poster into it, or it reads the incoming emails by itself, and either way a person gives the final approval before anything goes into the database. Going out, it pulls programmes from that database and drafts the material: social media posts, reminder messages, even prompts for short video scripts, ready for LINE and WhatsApp, TikTok, Instagram and Facebook.
For the room in Linz I tried to show it live. I asked OMSIN for the exchange programmes in Austria and for the announcement to be written in German, with Gen-Z language. The human stays at the end of the line, and I told the room plainly that full automation without someone checking is not something I would advise. Drafting was never what made this one hard; the hard part is deciding where a person has to look before anything goes out, and accepting that some mornings the robot simply will not answer.
Demonstration two: the MOU records
Our office keeps over a thousand MOU records, and when someone in a meeting asks whether we still have an active agreement with a particular university, finding the answer quickly is harder than it should be.
So we built a chatbot on LINE that queries the underlying spreadsheet in real time. On stage I asked it about Shibaura Institute of Technology, since Prof. Masaomi Kimura of Shibaura was sitting in the room and speaking later in the same session, and this time the answer came back at once: the agreement is active, this is the start date, and yes, it renews automatically.
Behind that answer there are two quieter jobs. When a new agreement arrives, the staff simply forward the email with its attachment to the agent’s own email address. It reads the document, writes a draft entry and asks a staff member to approve it with one simple email message. And at the end of every month it goes through every record and tells the staff which agreements are about to expire, so that a renewal is never missed because nobody happened to look. None of this is clever. It just never forgets the end of the month.
The question from the room
The best question of the morning asked for more than I had shown. Could I ask the agent something strategic, for example to analyse our top 15 percent of mobility partners and give us insights for future decisions? The honest answer was that this is a work in progress, and the reason is not technical. The sensitive part is deciding how much of our data we allow the AI to read. Looking up one agreement is one thing. Handing over the whole history of our partnerships so that a machine can judge them is a decision for people, and not a small one.
What it costs
People always want to know the cost, so I gave the real figures. The Claude monthly subscription that I build with is around USD 100. Running the agents is pay-per-use, and for a small office like ours that comes to roughly USD 30 a month. A bigger office asking more questions will spend more, but it is not the kind of budget that needs a committee.
What I hope the room took home
I closed with two points. The first is to keep humans in the loop. The agents do make mistakes, a misspelled name or an entry approved when it should not have been, and a person checking at the right moment catches them before they matter.
The second is the one I keep coming back to. Building these agents taught me that the technology itself is the easier half. Keeping the data in step, controlling what the agents can reach, connecting them to the chat apps: all of that can be learned. What cannot be downloaded is an intimate knowledge of how an administrative office really works, which rules bend, which deadlines bite, which email always arrives late. The people who already know that are not the engineers. They are the staff in our international offices, many of whom were sitting in that room in Linz, and they are the half that makes the other half worth building.