Skip to main content
Back to the villageDONGGU’S FIELD NOTESAI-PLATFORM
The workshop · selected work
PLACE
(주)GS
ROLE
Full-Stack
STARTED
2024-04
★ Case 04 / 5Forward Deployed Engineer Donghyun Kang

GS Inc. AI Agent Platform

On a non-developer self-run environment a 6-person team built on top of open-source Dify, I owned two axes — the org-mapped permission model and the non-developer training.

(주)GSFull-Stack2024.04 ~ 2024.10
01.Discover

Everyone wanted to use AI, but they had to wait for IT's schedule.

Marketing, HR, customer service — all wanted to use AI. The entry barrier, permission management, and IT-team dependency stopped every attempt.

"I want to use AI but I have to wait for IT's calendar." That line repeated across teams. Even when a tool got built, change requests never ended, and IT could not carry every team's asks.

The domain lived with the team that owned it. A centralized "build a tool for you" model could not reach that distance.

Outcome ▸AI adoption = IT bottleneck. Self-run by the team itself: zero.

Field notes

Note 01

Tool-building never ended

Even built tools attracted endless change requests — the tool model had no closing point.

Note 02

IT was the ceiling

One IT team could not absorb every team's AI ask.

Note 03

The domain lived with the team

The team using the workflow understood it better than IT ever could.

02.Build with AI

Designed organizational dynamics, not another tool — supported self-running teams through permissions and training.

Decision 01

RBAC mapped to the org chart

Dify's default admin/user split did not match the way the company is actually structured by department, title, and role. Mapped the org chart itself into the permission model as a matrix.

BeforeDify default — global admin/user, two tiers
AfterA multi-tier matrix across department, title, and role

tradeoffComplexity vs. accuracy — accuracy first

Decision 02

Non-developer AX training

Knowing that installing the tool alone would not produce self-running teams, ran a training track that taught "how to build" alongside adoption. Permission tier and learning tier moved together on one track.

BeforeTool adoption and learning ran on separate tracks
AfterAdoption and training on one flow — permission ↔ learning, 1:1

tradeoffLaunch speed vs. how long self-running actually lasts

Decision 03

Organizational dynamics — closing the fear of sprawl

AX adoption lands best when the internal heat is right — the will to solve with AI plus environmental support. At that moment, what stops fear-of-sprawl from producing meaningless MVPs is not technology; it is the permission model (boundary) plus training (learning path) — how you compose the organizational dynamics.

BeforeAdoption = MVP sprawl; fear of sprawl pushes things back into central control
AfterPermission boundaries × learning paths close the sprawl — self-running becomes possible

tradeoffAdoption speed vs. how long self-running lasts

03.Prove with metrics

On top of permissions and training, employees build and teams run it themselves.

Internal AI tools

~140

Employee-built (platform-wide, 2025)

Build speed

1 month → under 1 week

Professional dev → employee self-build (platform-wide)

Group-company adoption

20+

MISO platform (as of 2025)

DX leads on one workflow

230 people

Platform unified workflow (2025)

What this case is really about
“An engineer who spotted the gaps in the company's org, language, and training context, then designed the permission model and learning path that let non-developers use AI safely on their own.”
Discover

Read the gaps in the org, language, and training context as a pattern.

Build

Designed the RBAC mapped to the org chart and wrote the non-developer training and ops playbook — permission tier and learning tier kept on a single track.

Prove

Platform-level change that landed on top of permissions and training — about 140 employee-built tools, cycle time from one month to one week, 20+ group companies adopting (as of 2025).

Linked principlesEvidence-DrivenOwnership