HelsinkiUni Help has answered questions from students and staff at the University of Helsinki since November 2025. Much of our work after launch has gone into the tools for the people who run it: finding problems, deciding who should look at them, and checking whether a correction helped.
Those tools live in Drupal: review pages, roles, reports and switches. The service that writes the answers runs separately. This session is about the work around it.
Each night, automated jobs check the content imports, ask again the questions the service should decline, and group the questions it could not answer. An AI model sorts each group: a possible content gap, an unclear question, or noise. The service owner first sorted a sample by hand, without seeing what the models said. On that sample, the older of two models agreed with the owner more often, so that is the one we run.
The owner gets a daily Slack summary of what needs attention. Each line names who acts, or says plainly that nobody has been named yet. We also learned to check that the summary itself arrives: one morning it didn't, and the silence looked like good news.
Most of the demo follows one content problem through Drupal. An editor opens a possible content gap. Where an AI model guesses why the answer failed, the screen says "AI suggestion, often wrong". We tested that guess before showing it: it was good at spotting missing content, but often missed cases where the answer was already on the page. So the editor opens the pages it read and decides. The gap then reaches the team responsible for the page, on their own site. They change the page and ask for another check.
The check runs the next night. The service is asked the same question again, as an anonymous student. If it now answers, the fix is marked confirmed. If not, the gap reopens. The check shows that an answer now exists, not that it is right, so a person still reads it.
I'll also cover the reports we removed because people weren't using them. And a feature the owner wanted even after our test showed too many false alarms: an early warning when a topic suddenly rises. We changed what it promises. It is now a weekly hint, not an alert, and the screen says how often it was wrong in our test, next to every result.
At present, people make the content changes. We'll finish with where AI-written corrections might help, and what we would need to check before allowing them.
The demo uses demo data. No prior AI experience is needed.
Learning objectives
- Build a Drupal workflow for reviewing unanswered questions, assigning content problems and re-checking corrections.
- Create a daily report that makes problems and unassigned work visible to the service owner.
- Judge automated suggestions and checks: what they can show, where they go wrong, and where a person must decide.
- Find a first use case in your own site's unanswered questions or searches that return no results.