{"resourceId":"jeju-national-university-korus-ai-clerk-onprem-llm-20261028","versions":[{"version":"external-896eed60c284e89b5c4a641419f72ebf5ec4753b32a1a61d6be8cd678559ee9d","resource":{"id":"jeju-national-university-korus-ai-clerk-onprem-llm-20261028","title":"Jeju National University pilots a staff-built AI clerk on Korea's shared university ERP, running on its own GPU-hosted LLM","organization":"Jeju National University","sector":"Higher education administration (ERP, finance, HR)","geography":"Jeju, South Korea","publishedAt":"28 October 2026 (Jejudomin Ilbo)","publicationDate":"2026-10-28","eventDate":"2026-10-28","sourceName":"Jejudomin Ilbo","sourceLabel":"Regional news reporting (Korean)","sourceUrl":"https://www.jejudomin.co.kr/news/articleView.html?idxno=401734","evidenceClass":"independent-reporting","outcomeClass":"emerging","topics":["knowledge-work","infrastructure","data-security","developers-agents"],"finding":"Jeju National University began piloting an AI assistant layered onto KORUS, the shared resource management system for Korean national and public universities. Administrative staff built it alongside regular duties with no separate budget, adding features over the user interface instead of modifying the shared system. It summarizes, proofreads and drafts replies in the e-approval window, triages incoming documents by urgency, translates, and offers HR, payroll and finance chatbots grounded in manuals and laws. Generative AI runs on a university-owned GPU server rather than external services. Wider rollout to 36 universities is under consideration.","sledRelevance":"U.S. campuses on shared or system-wide ERPs face the same constraint: they cannot change the core system quickly. An overlay approach with self-hosted models is a transferable pattern, though Korean public-sector context and data rules differ.","evidence":"Newspaper report of the university announcement and 27 October staff demonstration; no measured results.","architectureImplications":"UI overlay on a shared ERP rather than core changes; on-premises GPU LLM; local translation on the PC.","governanceImplications":"Staff-built tools on a shared national system raise questions about review, support and change control before wider rollout.","securityPrivacyImplications":"Sensitive information removed from chatbot manuals; translation kept on device; no external AI services.","caveats":"Pilot with no performance data; self-reported by the university; Korean regulatory and shared-system context limits transfer.","streamIds":["campus-operations"],"roles":{"sales":"Sales framing: a public university added AI to a shared ERP it could not change, using its own GPU-hosted model and staff time. Interpretation: system offices and campuses on shared ERPs may want overlay assistants for document drafting, triage and policy questions without sending data to outside services. A bounded engagement is an overlay feasibility study on one workflow. Do not claim time savings; the pilot reports none.","engineering":"Engineering framing: the tool sits on top of the KORUS interface, runs generative AI on a university GPU server and translates on the PC. Interpretation: pre-sales should assess whether a campus ERP's interface allows a safe overlay, what self-hosted model capacity exists, and how chatbots ground answers in manuals and law. Overlay tools can break when the underlying system changes, which the source does not address.","delivery":"Delivery framing: staff built this alongside their jobs with no budget, which works for a pilot but not for 36 universities. Interpretation: delivery teams should plan ownership, support, testing against ERP updates and feedback loops before scaling. Acceptance should cover answer accuracy against manuals and triage agreement with staff judgment, since the pilot publishes neither."},"roleDetails":{"sales":{"customerProblem":"Shared ERP systems are hard to change, leaving staff with manual lookups and slow document handling.","stakeholders":["System office IT","Campus administrative leaders","HR and payroll","Finance and accounting","Information security"],"discoveryQuestions":["Which shared systems can you not modify?","Do you have self-hosted model capacity?","Which document workflows consume most staff time?"],"valueHypothesis":"Interpretation: an overlay assistant could reduce manual lookups without core ERP changes; no gains are measured.","potentialEngagement":"A bounded overlay feasibility study on one administrative workflow.","unsupportedClaims":["Time savings","Answer accuracy","Readiness for multi-campus rollout"]},"engineering":{"fit":"Fits institutions on shared or legacy ERPs with access to GPU capacity.","architectureIntegration":"Browser or desktop overlay on the ERP interface, self-hosted LLM on a university GPU server, chatbots grounded in manuals and law, on-device translation.","prerequisites":["GPU server and model hosting","Curated manuals with sensitive data removed","Permission to overlay the ERP interface"],"constraints":["Overlay may break on ERP updates","Korean data rules differ","Self-hosted model quality unknown"],"security":"No external AI services; manuals scrubbed of sensitive data; translation stays on the PC.","proofOfValue":["Test answer accuracy against manuals","Measure triage agreement with staff","Test overlay after an ERP update"]},"delivery":{"work":"Pilot and then scale an ERP overlay assistant with support and change control.","dependencies":["GPU capacity","Shared-system operator agreement","Staff feedback process"],"ownership":"Campus administrative IT, with the shared-system operator if expanded.","skillsAdoption":"Staff-led development; demonstrations to all staff before pilot.","governanceCheckpoints":["Security review of overlay","Accuracy review of chatbot answers","Decision gate before multi-campus rollout"],"acceptanceCriteria":["Proposed: chatbot answers match manuals in a sampled review","Proposed: named owner and support process before expansion"],"risks":["Overlay fragility","Unsupported staff-built code","Model errors on legal or payroll questions"]}},"retrievedAt":"2026-10-28T21:48:45-05:00","enrichedAt":"2026-10-28T21:52:11-05:00","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Built by administrative staff themselves; aims to reduce manual lookups and phone queries to other offices.","procurementImplications":"No separate budget reported; relies on existing GPU capacity and staff time.","operatingModelImplications":"Shows a staff-led development model for administrative AI that may later need central ownership if shared across 36 universities.","sourceVerification":{"openedUrl":"https://www.jejudomin.co.kr/news/articleView.html?idxno=401734","referenceExcerpt":"생성형 AI 역시 오픈AI 등 외부 기업 서비스를 쓰지 않고, 대학이 보유한 그래픽처리장치(GPU) 서버의 대규모언어모델(LLM)로 구동한다.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}