Public-sector AI case session 2: is AI only the AI unit's job? ([2회차] AI 챔피언 공공 AI전환 사례공유_AI는 전담부서만의 일일까요?)
B · Useful, limited First-hand operating account from Ahn Jin-hee of the intelligence and data unit at Korea's Anti-Corruption and Civil Rights Commission on running a distributed AI-builder network; specific and candid, but self-reported with no outcome metrics, in a session hosted by the programme that trains these champions.
Summary based on the full transcript·Original: about 55 minutes to watch
In short
A 55-minute Korean session (uploaded 3 November 2026) in the public-sector AI Champion series hosted by NIA. Ahn Jin-hee, an administrative-track civil servant in the Anti-Corruption and Civil Rights Commission's (ACRC) new intelligence and data unit, explains why the commission spread AI building across 20 "innovation leaders" instead of keeping it in her unit: a central team that builds every tool also has to maintain every tool. She describes how work was found, how accountability is kept with people, and how a complaint screener that failed on the internal network was rescued by encoding a veteran officer's rules.
Key numbers
- 20AI innovation leaders spread across the commission's bureaus, with tool subscriptions and API keys paid centrally
- 27candidate tasks surfaced by an app that asked staff to describe their processes
- 498statutes covered by the whistleblower protection act, which makes complaint-vs-report intake hard to classify
- twice a monthreview of the AI transformation plan with the commission's vice-head
- about 20one-minute "AI shorts" how-to clips posted on the internal board
As reported by the source.
Korean·Video·opens youtube.com
What it reports
- Trigger: a tool her unit built for another department came back for "after-sales service" when its data fields changed. Her conclusion: the unit cannot maintain everything it builds.
- 20 innovation leaders across bureaus take mandatory online and in-person training and meet monthly; costs are paid centrally, approved up to the chairperson. One standard curriculum bored the skilled and lost beginners, so it will be split next year.
- Most leaders are administrative staff; IT-track staff rarely join because entry-level courses waste their time.
- After a commissioned research project failed, an app asking staff to describe their processes produced 27 candidate tasks; outside experts picked the flagships. The AI transformation plan, adopted in September, is reviewed twice a month with the vice-head.
- No AI output becomes an automatic decision; a person confirms and is responsible for every result.
- Complaint vs whistleblower-report screener: it worked on the internet-side network but accuracy dropped sharply on the separated internal network with weaker models. It was rebuilt with a veteran intake officer's rules of thumb in the context; target end of November.
- Anti-graft Act interpretation helper: building was quick; collecting, cleaning and de-identifying past interpretations was the real work.
- A fair-hiring checker for job notices was built and published but little used.
- For a collective-complaint pilot she interviewed 4-5 veteran investigators for 1-1.5 hours each, because a survey would not surface how they judge cases.
How they got it
A narrated case from the practitioner who built the programme, with questions from the NIA host and a district-office data head. The English summary and figures come from auto-captions.
Limits
- Self-reported; no usage, accuracy or time-saving figures for any tool.
- Hosted by the programme that trains these champions.
- Auto-captions may garble names and numbers.
Why it matters for companies
The maintenance argument applies to any central AI team: if it builds every tool, it inherits every support ticket, so distributed builders with central funding and guidance scale better. Two practical points carry over directly: tools that work on open-network models may fail on the weaker models allowed inside a secure network, and the fix was domain expertise written down, not a bigger model. Training needs separate tracks for beginners and technical staff, or both groups drop out.
Key passages in translation
Korean original
그 모든 것들을 사실 저희 부서에서 AS를 다 한다. 아 저는 불가능하다라고 생각이 들었습니다.
If our unit has to do after-sales service on all of that, I thought: that is impossible.
Korean original
AI가 지원이나 도움은 주지만 최종적인 책임은 사람이지게 하는 방식으로 저희는 지금 이렇게 운영을 하고 있어서
We operate so that AI gives support and help, but final responsibility rests with a person.
Korean original
그분의 지금 노하우까지 지금 다 탈 털어서 이거를 지금 다시 약간 개발한 수준으로 해서 지금 내부망 이간을 지금 준비하고 있습니다.
We drew out all of that person's know-how, more or less redeveloped the tool with it, and are now preparing to move it to the internal network.
Summary based on the full transcript. Read the original for the full argument and data.Watch the original on youtube.com