What Students Actually Ask: Demand Structure and Automation Potential in a Hybrid Support System

Synthesis: Gupta et al. (2026) report a nine-week deployment study of an AI-integrated query resolution platform in a large online internship program, covering 4,093 queries from 1,434 participants between 2 May and 6 July 2026. Queries move across four routes: a retrieval-constrained AI assistant answering only from a curated corpus, peer answers, reuse of that corpus, and escalation to administrators. Nearly every query reached a recorded resolution and one in five closed within an hour. The headline is about demand, not technology: reuse of 114 corpus entries absorbed 21.3% of volume, participants self-resolved 12.2%, and 132 participants answered peers at a median of 9 to 15 minutes. Demand was narrow — one process step, certificate submission and the offer letter it triggers, drove 56.8% of corpus-mediated resolutions — and at least 20.4% of queries asked about the status of a pending submission, which a stored answer cannot serve. The authors also treat the resolution record itself as evidence and show where its fields mislead.

Key Findings

1. Over nine weeks the platform handled 4,093 queries from 1,434 participants; nearly every query reached a recorded resolution and one query in five closed within an hour.
2. Corpus reuse absorbed 21.3% of volume across 114 entries, participants self-resolved 12.2%, and 132 peers answered one another at a median of 9 to 15 minutes.
3. Demand was narrow rather than varied: certificate submission and the offer letter it triggers accounted for 56.8% of corpus-mediated resolutions, and no other entry cluster exceeded 9.0%.
4. At least 20.4% of queries were status inquiries about a pending submission rather than requests for information; the retrieval-bound assistant served only 1.3% of them, since a stored answer cannot report a student's current status.
5. Only 8.0% of queries handled by a person duplicated existing corpus content, suggesting the knowledge base was already well used and that further corpus expansion would yield modest returns.
6. Direct administrative answering was concentrated in one Administrators who handled 72.7% of identifiable human responses and 91.0% of the direct-admin path.
7. Repeat inquiry was rare at 0.6%, and 301 queries (7.4%) carried no closure timestamp, so the resolution record is a censored and partly misattributed account.

A narrow demand profile

The study's central move is to classify the query text rather than infer demand from how a query was resolved. Read that way, the demand reaching this online program was strikingly narrow. A single onboarding step — submitting an institutional certificate and receiving the offer letter that follows — accounted for 56.8% of all corpus-mediated resolutions, and reuse across entries was concentrated, with a Gini coefficient of 0.586 and a quarter of entries used exactly once. A second, larger class is not a request for knowledge at all: at least 20.4% of queries asked where a pending submission had reached. Because that answer is specific to one participant and changes daily, the retrieval-grounded assistant served just 1.3% of this class while handling procedural questions readily. That gap is the design working as intended, not failing, and it marks a limit of the Help-Seeking model that assumes a query is a request for knowledge the asker lacks.

Dispatch, cost, and the peer layer

Volume and time split sharply by route. Scripted bulk action (31.6%) and direct administrative answering (27.4%) carried most of the load; the assistant handled 3.6% and the peer layer 7.0%. But speed inverted the volume picture: 70.3% of peer-answered queries closed within an hour, against 1.9% of direct administrative answers and 14.8% of scripted closures. Peer resolutions came from 132 participants, 101 of whom also raised queries, with a Gini of 0.414 — modest in volume yet broadly shared. Because most administrative closures occur in synchronous bulk events, raw latency conflates routine turnaround with periodic queue clearing; separating the two gives a routine median of 6.1 hours against the 28.9-hour aggregate. The 8.0% duplication figure, falling to 6.3% once status inquiries are set aside, is a conservative floor, since the lexical TF-IDF comparison misses paraphrase.

Reading the resolution record as evidence

The paper's third contribution is methodological: it treats the operational record as an object to be tested. Three field-level properties qualify any reading. The resolver-email field denotes the author on administrative paths but the approver on peer paths; the acceptance flag encodes only whether the resolver differed from the asker, not quality; and closure timestamps bunch into bulk events. Attribution checks were reassuring in one direction — 99.7% of the top administrator's 1,024 records were direct answers and peer identity is held separately — yet human oversight here means approval routing, not judgment of answer quality. The authors are explicit that no finding evaluates resolution quality, and that repeat inquiry at 0.6% is too rare to stand in for it. For anyone applying learning analytics to support tickets, establishing workflow semantics before interpreting contributor distributions is the transferable lesson.

What this means for practice

• Instructors. Stop treating volume as an answering problem. The largest demand class is "where is my submission?", so surfacing process state to students removes load without anyone writing a better answer.
• Administrators. Record authorship and approval in separate fields on every path, as the peer paths already do, so contribution becomes measurable system-wide rather than only where approval routing happens to capture it.
• Educational technology developers. Scope a grounded assistant to procedure and route status questions to a live system of record; a retrieval-bound agent cannot answer per-student state, and asking it to try invites unsupported output.
• Researchers. Treat support-ticket records as censored and semantically loaded. Check closure-timestamp clustering and what an acceptance field actually encodes before reading contributor or latency figures as evidence.

Limitations

• The data cover a single program across nine weeks of onboarding, with volume peaking in week 3, so they describe an onboarding period rather than a steady state and the authors claim no generality across programs.
• Demand categories were defined by one author using a rule-based classifier at 75.0% precision with no second coder, so category volumes are lower bounds and the two least precise categories are indicative only.
• The coverage estimate compares query text against a proxy corpus built from corpus-linked queries because the export holds no answer text, so it shows comparable content existed rather than that the assistant would have answered correctly.
• The authors operate the system under study, which gives both direct workflow knowledge and an interest in how the system is described.

Citation

Gupta, J., Ayinampudi, P., Aditya, B. M. V., Hegade, P., Sharma, R., Sharma, S., et al. (2026). What Students Actually Ask: Demand Structure and Automation Potential in a Hybrid Support System. arXiv preprint.