Customer evidenceKorean healthcare provider · anonymized

Developers wanted AI. The hospital needed a boundary it could control.

  • Patty Code
  • PCCP

A controlled Patty Code and PCCP rollout helped 18 developers understand and maintain aging systems without sending patient information or operational systems beyond the hospital’s boundary.

Initial participants
18 developers
Scope
6 repositories · non-production
Observation period
Approximately 8 weeks
Primary environment
Java/Spring · Oracle · Windows · legacy integrations
01

A small team was responsible for several generations of systems

The provider operates a patient portal, scheduling, billing and insurance integrations, staff applications, and interfaces around its EMR environment. Its internal development team is small relative to the number and age of those systems, and knowledge of older integration code was concentrated among a few senior engineers.

Developers wanted AI assistance for code exploration, test generation, and repetitive maintenance. Security agreed that the productivity potential was real. The unresolved question was whether the hospital could explain and constrain the information sent to models and the actions an agent could perform.

02

The previous choice was effectively allow or prohibit

The organization could not rule out patient identifiers, source code, configuration, or sensitive operational context leaving its controlled environment. It also lacked dependable answers about which model received context, which files the AI read, whether it could execute commands or install packages, and who had authorized an action.

The official policy therefore prohibited or heavily restricted external AI coding tools. There was no operating model connecting developer demand with the hospital’s obligations.

03

Patty Code supplied the experience; PCCP supplied the basis for approval

Patty Code initially attracted the engineering team, but PCCP made the deployment approvable. The hospital could place explicit boundaries between models, repositories, tools, and execution without redesigning its entire development environment around AI.

04

The hospital started small and explicitly excluded production

The initial scope covered 18 developers, six repositories, and approximately eight weeks. It included scheduling, billing integrations, an internal portal, notification services, shared libraries, and one older integration application. Live EMR data and production deployment were out of scope.

05

Implementation

The control path came first

PCCP was deployed inside the controlled environment as the authorized path for AI-assisted development. Patty Code was provided to participating developers, while model access and execution passed through the same policy layer.

  • Production databases and patient identifiers: denied

  • Normal repository reading and modification in approved workspaces: allowed

  • Package installation and sensitive directories: additional approval

  • Production deployment: prohibited

Policies were tuned from actual usage

The initial configuration required approval too often and interrupted developers. Usage evidence was then used to automate low-risk actions while retaining approval for sensitive files, tools, and environment transitions.

06

Observed change

4일 → 2.5–3일
Routine maintenance turnaroundFor applications included in the pilot
2–4시간 → 45–90분
Initial unfamiliar legacy-code investigationObserved time to a useful first analysis
50%+
Reduction in preparation time for many code reviewsCustomer team’s pilot observation
07

Correct code interpretation could still differ from operational reality

Even when the agent interpreted source code correctly, the architecture represented in code was not always operationally relevant. Senior engineers validated the initial system map, and blocked context or actions were treated as evidence that controls were working—not as product failures.

08

It began as a productivity project; governance made adoption possible

Developers valued what Patty Code could do. Security and compliance valued the boundaries and records supplied by PCCP. Because the use case remained people directly operating an agent, there was no need yet to add Crew or Pilot.

09

Evidence and disclosure

Customer-reported and pilot-observed. Figures are approximate, limited to the participating applications and period, and are not an organization-wide productivity measurement.

The organization name and identifying system details are withheld. No direct quotation is presented; interview notes have been reconstructed as third-person editorial narrative.

RELATED

Related products and technology

View all customer stories