● Agentic AI Surge Hits Public Sector
For Public Sector AX to Be “More Than Talk”: Core Point Is Leadership, Organizational Culture, and Agentic AI Design
Today, I’ll summarize why AI (AX) adoption in public institutions doesn’t move fast—and what must change to create “real success,” focusing on on-site examples. Especially, the following is the core point of the article.
1) Before technology transfer, the ability to connect the problem to the right domain (work context) determines the outcome
2) Beyond generative AI, agentic AI must block “risky behaviors” and embed verification
3) While leaders’ attention shapes the atmosphere, lasting change ultimately has to connect to motivation (sense of mission and contribution to the public good)
4) “Saved time” doesn’t automatically translate into productivity → you need a design that converts that time into valuable work (innovation)
Centered on these four, I’ll summarize the bottlenecks and solutions for AI adoption in public institutions in a news style.
1. Why Public Institutions Respond “Slowly” to AI and Digital Transformation (AX/DX)
1) Top-down demands are always increasing, but domain connection is often weak
Public organizations face increasing pressure to drive AX according to government directives. However, in reality, AI has to know the “work domain.”
The key points from the field are as follows.
Without DX (work and data structure innovation), pushing AX (automation/AI application) first doesn’t lead to good results
That’s because for AI, understanding work logic + connecting data matters more than “coding skills.”
2) It’s not a lack of urgency—rather, the “recognition of crisis” doesn’t attach well to motivational fuel
In the private sector, change often accelerates during crisis situations. But in public institutions, the driving force of “crisis-driven reform” is relatively weak.
So public institutions need to create different motivations rather than rely on crisis.
In the on-site explanations, they viewed motivation connected to the public good (citizen happiness and policy contribution) — PSM: Public Service Motivation as important.
3) “Saved time” doesn’t turn into performance immediately
Even if automation reduces time for work, in many cases that time doesn’t immediately translate into innovation.
Especially in public institutions, “mandatory” work like reporting, audits, and requirement responses exists continuously.
So the remaining issue is not simple efficiency, but creating slack (empty space) for innovation and converting it into productive labor.
2. “Success Patterns” in Public Sector AX — Three Things to Look At Before Technology
Success Pattern A: Design top-down and bottom-up in parallel
Public institutions tend to have clear directives and evaluations (top-down), and in the field, they emphasized that success requires domain-based bottom-up movement to run alongside it.
That is, not “AI adoption prompted by instructions,” but find bottlenecks felt by the frontline, create the experience of changed work, and spread it.
Success Pattern B: Leadership creates “temperature,” and lasting change needs a “plus alpha”
It’s true that when leaders strongly drive AI, the atmosphere changes dramatically.
However, persistence requires other elements as well.
From the field perspective, it’s this.
Even if the leader changes, the organization has to keep running in the same way—and this part ultimately comes down to organizational culture
And they also felt that as team members understand “why changes are necessary,” voluntariness increases.
Success Pattern C: Design for controlling risk—not “engineering automation” alone
When you introduce generative AI, it’s good to start with easy areas like reports and visualization.
But public-sector work, if done incorrectly, directly affects personnel and budgets.
So the important thing is the mechanism that stops, verifies, and confirms with the user the moment the agentic AI tries to take risky actions.
3. On-Site Demonstration: How They Turned the “Total Personnel Cost Increase Rate” Excel Bottleneck Into a System
Key Background: Personnel cost calculations explode at year-end
Public institutions often have to compute personnel costs at year-end. In that process, 담당자 gather 20–25 files in Excel, link them, and the cross-check burden is huge because they have to match everything down to the unit of won.
In the field, they also explained this issue as a “risk that’s difficult to account for as cost” (dependence on know-how, safety issues).
Transformation Goal: Turn “Excel repetition” into “automatic calculation based on ERP row data”
Existing: Excel → manual整理 → linking → calculation → verification (human error possible)
Transition: Load ERP data → automatically compute based on logic → produce results in a verifiable form
How it was implemented: Agentic AI + Hel Guard (verification fence)
Here, technical details come in, but from a blog perspective, the most important thing is “including safety mechanisms.”
1) Guardrails like disabling templates (regulatory compliance)
Example: A certain template needs to be removed under government guidance, but if the AI receives an instruction to “enable it,” it’s designed not to execute and either stop or ask the user for re-confirmation.
2) Classify the scope of change (what can be touched)
Split the risk level into whether it’s just a UI change, a change to template values, or a change to the linking logic, so the AI moves starting with the safe areas.
3) Embed a verification step
To reduce the possibility that links break after the work or that data consistency gets compromised, they included verification flows before and after execution.
Why this matters: “B y vibe coding” reduces maintenance risk
The point emphasized in the field was maintenance.
If you build it once via outsourcing, it can be difficult to respond to changes in logic or directives, and costs can rise.
So even if the data structure or the method for calculating management evaluation changes, designing the system logic in a form that can be updated was seen as an advantage.
Impact (time savings): 3–4 weeks → possible to reduce by more than half
In the demo case, they mentioned that when the Excel-based work that took 3–4 weeks is systematized, it can cut the time by more than half.
The key point is that, along with “faster results,” there were also psychological and operational benefits: “the outcome can be confirmed right away, and uncertainty about errors decreases.”
4. To Expand Public-Sector AX More Widely: Standardization and Sharing (Platformization)
Public-sector logic is similar, but since ERP/data differs, a complete copy-paste isn’t feasible
As the number of public institutions grows to around 350, if there are “shareable standard modules” like an agent market, the spread speed could increase.
However, in reality:
Because each institution’s ERP structure differs, the required data and linking method also differ, making customization essential
So even if there are more than 50% similar areas, the remaining differences still have to be reflected at the design stage.
Still, a “system I use alone” has no real meaning
The message from the field was strong here.
The system you build shouldn’t end with a single owner using it alone; you need to expand value through organizational sharing and learning structures
They mentioned approaches like operating a learning organization and cross-check collaboration.
5. Priority for Studying “Latest Trends”: Stop chasing technology and focus on the problem
Technology changes too fast, making “following trends” inefficient
In the field, they also advised on recent AI learning approaches.
Instead of following the technology itself, you should bring technology in as a means to solve problems in your own work
It was a direction that cautioned against putting the cart before the horse—when technology becomes the purpose.
And the conclusion: Organizational culture, leadership, and problem awareness create onboarding
AI is a tool, and organizations move through people.
When new technology arrives, to make people actually change “how they work,” onboarding (learning) + leadership signals + alignment with organizational culture must come first.
“The single most important line” summarized separately in this article
The success or failure of public-sector AX doesn’t depend on how much AI you’ve adopted; it depends on organizational design that resolves domain bottlenecks while controlling risky behavior (including verification) and converting saved time into innovative labor
And the axis of that organizational design is leadership (temperature) + public service motivation (voluntariness) + standard sharing/learning (persistence).
Checklist: Items you can use immediately when pushing AX in public institutions
– Whether the system is designed together with someone who has knowledge of business logic (domain)
– Whether it’s more than just generative AI, and whether the structure extends all the way to actual estimation and verification
– When agentic AI moves, whether there are guardrails like “stop,” “verification,” and “approval”
– Whether the leader’s drive hardens into organizational culture/process rather than being “one-time”
– Whether design is in place so that time created through automation truly converts into innovation, including KPI/work allocation/learning
– Finally, whether all of this is turned into a template/module that enables sharing and reuse, and whether it’s being expanded
Primary content you want to convey (conclusion in one paragraph)
The biggest reason AI adoption doesn’t go smoothly in public institutions is not simply a lack of technology; it’s because, without understanding the work domain, AX is pushed only from the top down, or the design is weak in converting time created by automation into valuable innovative labor.
On the other hand, in successful cases, beyond generative AI, they systematized estimation logic with agentic AI and blocked risky actions while embedding verification, and, together with leadership-driven momentum, they created voluntariness through public-good motivation (public service motivation) and secured persistence through shared-learning structures for team members.
< Summary >
– Public-sector AX has limitations if it relies only on top-down directives, and the key is domain connection for DX.
– Innovation is possible only when there is slack, but since saved time doesn’t automatically convert, it requires design.
– Motivational drivers like public-good motivation (PSM) become an important force instead of urgency about crisis.
– When applying agentic AI, you must stop risky behavior (guardrails), classify the scope of change, and embed verification steps.
– By replacing the Excel bottleneck with automatic estimation based on ERP row data, it was possible to reduce the 3–4 week workload by more than half.
– Only with standard sharing/learning does the built system become an organizational asset rather than a personal tool for a dedicated owner.
[Related Posts…]
- The Changes Agentic AI Brings to Public Institution Operations
- Automating Total Personnel Cost Estimation: A Data and Logic Perspective That Removes the Excel Bottleneck
*Source: [ 티타임즈TV ]
– AI로 바뀌는 공공기관, 성공하려면 무엇이 달라야 할까? (김창일 한국학중앙연구원 박사)
