Research Methodology — University of Trento · 32 teams, 92 research questions
The RQs cluster into a few dominant themes. A full third of the class converged on ticket routing, classification, and triage, while pain point discovery and anomaly detection formed the second and third largest groups.
The three largest themes — routing & triage (31), pain points (17), and anomaly detection (15) — map directly to the three stakeholder problems in the exercise (Dana's routing problem, Marco's feedback problem, Priya's early warning problem). This is expected, but the heavy skew toward routing (34% of all RQs) suggests students found the classification problem most concrete and tractable. Fewer teams tackled the harder, more open-ended problems around embeddings, sentiment, or human-AI trust.
The largest theme breaks down into four distinct subclasses, each framing the routing problem from a different angle.
The dominant subclass. Core question: can we automatically determine what a ticket is about and send it to the right specialist?
"How can we automatically figure out what each support ticket is about and send it to the right person as quickly as possible?"
"How can we automatically classify and route incoming support tickets to the correct specialist within 30 seconds?"
"Can we predict the correct specialist from the first 1–2 customer messages, accurately enough to route within 30 seconds and reduce transfers?"
"Is it possible to auto-route tickets to the right specialist within 30 seconds?"
"How can I filter the question to the right agent, basing on keywords in received text?"
"What topic information can be extracted from the conversation corpus, and can it be used to automatically route new tickets to the right specialist?"
A sharper variant: can we classify correctly from minimal initial context (first 1–2 turns), before the full conversation unfolds?
"Can ticket intent be reliably classified from the first 1–2 turns alone, before full context is available — and what routing accuracy is achievable within a 30-second latency constraint?"
"Can we infer the user's underlying problem type within the first 1–3 messages of a conversation?"
"How may I pre-screen help requests, so that human agents don't lose precious time with requests they are not capable to manage?"
"How many questions are needed to get to the core of a problem?"
Goes beyond topic classification to explicitly model agent expertise and match tickets to the best-fit specialist.
"How to find keywords in tickets and connect them with the skills of each agent?"
"domain-expert routing: how can we route similar requests to specialized expert agents for improving service efficiency?"
"If a ticket contains more topics, who is the agent that has to answer that ticket?"
"Given the avg answering time by different agents, what is the fastest way of mapping the right ticket to the right agent?"
"Developing a fast and cost-effective auto routing ticket management system with domain expert selection."
Not just where to route but how urgently — prioritizing by risk, severity, or volume signals.
"Categorize conversations based on risk, to route them accordingly and more quickly or flag them."
"Can context, loudness and volume of request predict the urgency of a feature?"
"How can you notice a stream of similar urgent customer complaints?"
"Properly ranking requests from massive questions database."
This theme splits into three subgroups depending on whether the focus is on identifying, tracking over time, or structuring the feedback.
The longitudinal angle: how do pain points evolve over time? Are we solving problems or cycling through the same ones?
"What are the dominant user problems expressed in support conversations, and how do they evolve over time?"
"The top pain points: which are they and are they changing over time?"
"How can I find significant features in Slack channel conversations to understand what the top pain points are and if they are changing over time?"
"How can unstructured customer feedback be systematically analyzed to identify evolving user pain points over time?"
The snapshot angle: given a corpus of conversations, what are the problems? Focus on extraction and classification at a point in time.
"How to find corresponding pain points across various customer roles?"
"How can you identify reliably the part of the product that gets the most feedback?"
"How to identifying the core problem of users without relying on the post on Slack channel?"
"Can we infer the user's underlying problem type within the first 1–3 messages?"
Meta-level: how to organize feedback into actionable categories — distinguishing bugs from feature requests, deduplicating, building a taxonomy.
"How to detect duplicate feature requests?"
"How to differentiate between issues the customer faces due to bugs versus suggestions which can be added to improve the app/service?"
"How to collect relevant feedback from clients on the features and bugs of our software?"
"A systematic approach for user need identification and priority-based product decision roadmap."
"Is the context clear? Are we really speaking the same language or is there a fundamental mismatch between our knowledge/expectations and needs?"
Unique framing — instead of jumping to a technical solution, this team questioned whether the communication itself is the root problem. Needs sharpening into a testable form, but the instinct is excellent.
"Can ticket intent be reliably classified from the first 1–2 turns alone, before full context is available — and what routing accuracy is achievable within a 30-second latency constraint? Which ticket types are most resistant to early classification, and what properties make them so?"
Concrete input (1–2 turns), performance constraint (30s), and a built-in follow-up about failure modes. Model RQ structure.
Generated from 32 team submissions · Research Methodology, University of Trento