{"resourceId":"jeonbuk-kr-sovereign-staff-ai-9300-users-dlp-20261022","versions":[{"version":"external-0701ed5f34db0762c8c322f60e62826eb3c89d6b4e8e146707cd6c8c8af6c321","resource":{"id":"jeonbuk-kr-sovereign-staff-ai-9300-users-dlp-20261022","title":"Jeonbuk's in-house staff AI reaches 9,300 accounts and 440,000 questions on a 300 million won budget, with DLP routing sensitive prompts inside the network","organization":"Jeonbuk State Government and Daegu Suseong-gu (presenting at the Korea Local Information Research and Development Institute forum)","sector":"Regional and local government internal staff AI","geography":"Jeonbuk State and Daegu Suseong-gu, South Korea","publishedAt":"22 October 2026, 16:37 KST (Digital Daily)","publicationDate":"2026-10-22","eventDate":"2026-10-22","sourceName":"Digital Daily (디지털데일리)","sourceLabel":"Trade press report of conference presentations","sourceUrl":"https://www.ddaily.co.kr/page/view/2026102216374535387","evidenceClass":"independent-reporting","outcomeClass":"emerging","topics":["knowledge-work","infrastructure","data-security","developers-agents","accessibility-workforce"],"finding":"Jeonbuk State built an in-house staff AI for 300 million won rather than accept a 5 billion won private proposal or pay an estimated 2.57 billion won a year for commercial accounts. Hosted inside the administrative network on AMD MI300X servers after a national security review, it has 9,300 accounts, 440,000 questions and 99 features. External models are reached by API, with DLP rerouting prompts containing personal or sensitive data to the internal model. Staff-built tools include a treasury deposit optimiser projected to add about 1.6 billion won a year. Expansion to 14 cities and counties is planned by year end.","sledRelevance":"Why now: presented 22 October and not previously archived. A detailed cost, hosting and data-control comparison for government-run staff AI, relevant to U.S. counties weighing commercial copilots against hosted models; Korean security review and procurement rules limit transferability.","evidence":"Presenters' self-reported usage, cost and design figures; projections for interest gains; no evaluation report.","architectureImplications":"On-premises inference inside the government network with API access to external models and a DLP gateway; models swapped from DeepSeek V3.1 to GLM-family without vendor involvement.","governanceImplications":"National Intelligence Service security review preceded launch; staff-built tools are encouraged through a research club, with a proposal to separate field development from expert security review.","securityPrivacyImplications":"DLP detects sensitive or personal data and redirects prompts away from external AI; internal questions stay on the government network.","caveats":"Self-reported by the builders at a conference; no measured productivity or accuracy data; projected interest gain; Korea-specific security and budget context.","streamIds":["local-government"],"roles":{"sales":"Framing: cost and data location drove a build decision. Jeonbuk compared a 2.57 billion won annual subscription and a 5 billion won proposal with a 300 million won in-house build, citing overseas data transfer. Interpretation: this supports bounded conversations with counties about total cost and data residency for staff AI, but the figures are self-reported, productivity was not measured, and the projected interest gain is not a realised saving, so it cannot back claims that in-house builds are cheaper everywhere.","engineering":"Framing: a hybrid design with a DLP gate. Internal inference runs on AMD MI300X servers inside the government network, external models connect by API, and prompts with sensitive data are rerouted internally. The team swapped open models and reports 3.3x performance on the same hardware. Interpretation: this is a useful reference architecture for hybrid staff AI, but U.S. teams would need their own security review, model licensing checks and monitoring; operations depend on one part-time engineer.","delivery":"Framing: adoption was built through staff who build. A 13-member club produced 13 tools, and training reached 1,800 staff, with expansion to 15,000 users planned. Interpretation: delivery plans for similar programs should budget for operations staffing beyond one person, separate citizen-facing development from security review as presenters proposed, and measure task time and accuracy before scaling to cities and counties."},"roleDetails":{"sales":{"customerProblem":"Staff are already using free external AI with government data, and commercial licences for all staff are costly.","stakeholders":["CIO or administrative information division","Security office","Finance and treasury","Staff innovation groups","Partner university"],"discoveryQuestions":["How many staff use unapproved AI tools now?","What would full commercial licensing cost per year?","Where must prompts and data reside?","Who would operate a hosted model?"],"valueHypothesis":"Bounded: a hosted staff AI with DLP may lower licence cost and keep data in-network, but no measured productivity exists in this source.","potentialEngagement":"A total-cost and data-residency options study comparing commercial copilots with a hosted model and DLP gateway.","unsupportedClaims":["Measured time savings","Realised 1.6 billion won interest gain","Lower cost in a U.S. context","Accuracy of outputs"]},"engineering":{"fit":"Regional or county governments with security requirements that rule out sending data to external AI services.","architectureIntegration":"On-network GPU inference, API connections to external models, DLP routing, document generation and meeting transcription features.","prerequisites":["Security review","GPU hardware","Open-model licensing review","Monitoring automation"],"constraints":["Korean security review regime","Model provenance scrutiny for open models","Single-operator staffing"],"security":"Keep internal prompts on the government network and enforce DLP before any external model call.","proofOfValue":["Measure DLP detection on a test set of sensitive prompts","Benchmark task accuracy on common document types","Load-test for the planned 15,000 users"]},"delivery":{"work":"Run and scale a hosted staff AI platform and a staff-built tool program.","dependencies":["Hardware budget","University partnership for staff","Security approvals for each new tool"],"ownership":"Jeonbuk administrative information division (one officer also covering cloud work).","skillsAdoption":"Monthly no-code AI building sessions and two-day training for 1,800 staff.","governanceCheckpoints":["Security review before each new external model","Review of staff-built tools before citizen use","Pre-expansion review before municipal rollout"],"acceptanceCriteria":["Proposed: measured task-time change on a sample of document workflows","Proposed: DLP false-negative rate below an agreed threshold","Proposed: second trained operator in place before expansion"],"risks":["Key-person dependency","Unverified output quality","Model supply-chain concerns"]}},"retrievedAt":"2026-10-25T22:18:28-05:00","enrichedAt":"2026-10-25T22:24:37-05:00","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Monthly 'vibe coding' training and two-day courses for 1,800 staff; one part-time administrator runs the platform with heavy automation.","procurementImplications":"In-house build cost 300 million won versus a 5 billion won proposal; a Rubin-class server is in next year's budget; plan to open-source the system.","operatingModelImplications":"Single-operator platform relies on 32 dashboards and 200+ nightly automated checks; key-person dependency is a risk.","sourceVerification":{"openedUrl":"https://www.ddaily.co.kr/page/view/2026102216374535387","referenceExcerpt":"계정 등록자는 9300명, 누적 질문은 44만건이다. 직원 요청에 따라 늘어난 기능은 99종에 이른다.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}