Peter Yang - AI Agents for Facility Operations

Hello!

Fault Detection and Diagnostics Agent

Location/asset Over-investment Agent

High risk work order management agent

Work Order Assignment and Cost Allocation Agent

Fault Detection and Diagnostics Agent (loop)

Location/asset Over-investment Agent (loop)

High risk work order management agent (loop)

Work Order Assignment and Cost Allocation Agent (loop)

Key decisions

Turn fragmented operations into a coordinated decision workflow

Unify priorities across specialized agents

Faults, safety risks, work allocation, and investment decisions lived in separate workflows. I organized four specialized agents within one weekly action view, giving each a clear responsibility while sharing a consistent path from operational signal to recommended next step. This helped facility teams compare what needed attention without reconstructing priorities across tools.

Connect recommendations to evidence and human control

A recommendation needed enough context to support an operational decision. I paired prioritized actions with work-order records, compliance gaps, ownership, lease timing, and planned spend. The framework distinguished completed auto-assignments from unresolved exceptions and kept costly, critical, or ambiguous decisions open for investigation and human review.

Representative interaction

A shared workflow connects weekly priorities to supporting evidence, focused investigation, and a decision about what to do next.

01
Identify what needs attention
Specialized agents surface urgent risks, recurring faults, assignment exceptions, and potential over-investment in a shared action view.
02
Investigate the recommendation
Review the relevant records and rationale, compare affected work orders or locations, and ask follow-up questions to clarify scope and consequences.
03
Decide the next action
Use the evidence to determine whether to escalate, schedule maintenance, review ownership, or reassess planned spending, with human review for high-impact decisions.

The production-code prototype was used to explore workflows and align Product and Engineering. The figures shown are scenario data, not measured operational outcomes.

AI Product DesignEnterprise SoftwareFacility OperationsCode Prototyping

Overview

Where I Started

Problem Statement

Facility teams had to monitor faults, high-risk work orders, assignment exceptions, cost allocation issues, and potential over-investment across disconnected workflows.
Important signals competed for attention, leaving operators to manually determine what required immediate action, what could be handled automatically, and what needed human review.

Context

The opportunity was not simply to add a chatbot. It was to create a coordinated set of agents that could turn operational data into weekly recommendations grounded in urgency, risk, financial exposure, and accountability.
I used four agent threads to define the model: FDD work order triage, high-risk work order prioritization, over-investment review, and assignment and cost allocation review.
The design challenge was balancing automation with transparency, ensuring that users could understand, verify, and approve recommendations before acting on high-impact operational decisions.

What I Did

Responsibilities & Contributions

Defined the agent framework: Structured diagnostics, risk management, investment review, and assignment workflows into specialized agents with clear triggers, decision types, and user outcomes.
Designed the first-response pattern: Made each agent open with a concise recommendation, the evidence behind it, and suggested next steps so users could understand what mattered before continuing.
Separated automation from approval: Let confident recommendations support bulk handling or auto-assignment, while routing critical, costly, or ambiguous cases back to human review.
Established trust and control patterns: Surfaced urgency, rationale, supporting data, and recommended actions without implying that the system should take high-impact action without user confirmation.
Prototyped in the production codebase: Used Claude Code to build the interactive experience directly in the product environment, testing design decisions against real components and technical constraints while creating a shared artifact for Product and Engineering alignment.

Impact

4 specialized agents

Covered diagnostics, risk management, work allocation, and investment oversight.

1 prioritized action view

Brought cross-workflow recommendations into a single weekly experience.

4 operational workflows

Connected complex signals to clear next steps through a shared interaction model.

1 reusable AI pattern

Established a foundation for adding future agents without creating separate experiences.

What I Learned

Key Insights

AI becomes valuable when it reduces decision effort, not when it produces more information. The strongest responses translated noisy tables into a clear first decision.
A useful agent needs a crisp opening move. Prioritized lists, grouped exceptions, financial exposure, and recommended next steps helped users understand the decision before exploring details.
Trust came from visibility and control. Users were more willing to act when they could understand why an agent made a recommendation, inspect the evidence, and confirm the scope.

Critical Constraints

Building data quality and confidence varied by source, requiring the interface to communicate uncertainty, assumptions, and data lineage without overwhelming users.
High-impact operational decisions required human review, clear accountability, and a path to investigate underlying evidence before taking action.
A shared interaction framework needed to support different evidence types and operational responsibilities while preserving a consistent path from recommendation to investigation and review.