Setting the hook
Why would a refinery race to build its own generative AI while margins compress, demand patterns shift, legacy systems creak, and thousands of frontline workers rely on mobile devices for every shift task?A sharper question sat behind that one: could a heavy-industry incumbent convert curiosity into enterprise-scale adoption before the technology itself shifted again, and do it without jeopardizing mission-critical operations that cannot blink?
Why this story matters
The stakes were concrete and immediate. Energy and chemical businesses managed volatile feedstock costs, stringent safety requirements, and intense global competition—an unforgiving mix that punished hesitation. Traditional digital programs often stretched across years and multiple vendors, leaving frontline teams stranded between good intentions and tools that never fit the pace of real work. Any attempt to bring generative AI into this environment needed to be safe, fast, and directly useful.
GS Caltex approached the dilemma by recasting risk. Leadership championed a practical standard—“good risk taking”—that shaped how teams tried new ideas and how quickly they shipped them. That stance mattered beyond rhetoric. In a few months, roughly 85% of employees became regular users of AiU, the company’s in-house generative AI platform, and about 95% of mobile-only workers engaged on their devices. In a field where digital pilots often stall, those numbers signaled a cultural and operational shift.
Inside the cultural and platform shift
Executives set the tone. CEO Hur Sae-hong described “good risk taking” as responsible progress—permission to act quickly with guardrails, not a license for recklessness. CIO/CDO Lee Eunjoo turned that belief into operating norms: get something small into people’s hands, learn from the results, then iterate. The message landed clearly—stop waiting for perfect; make it safe to try. A visible recognition loop reinforced the behavior, with frontline teams presenting monthly to leadership and peers, and failure cases discussed without stigma.
This mindset flowed into the design of AiU. The platform sat alongside core tools so that employees did not have to jump between systems to use it. It acted as both a knowledge interface and a development environment, letting staff query internal sources, summarize information, and prototype simple agents within the same workspace. Under the hood, AiU remained model-agnostic, able to work with providers such as OpenAI and Anthropic, which reduced lock-in and simplified upgrades as models improved. The bet was clear: assume the AI layer would change; invest in data and applications that endure.
The early build moved quickly. An AWS-based pilot launched in days and drew feedback from more than 300 participants, capturing requirements from finance, legal, HR, operations, and sales. The team then built on MISO—the GS Group’s AI transformation platform—adding GS Caltex–specific layers for data access, security, and workflow integration. From pilot to full platform took about six months, a timeline that reflected both leadership urgency and a deliberate choice to ship in thin slices rather than wait for a grand unveiling.
What changed on the ground
What employees received was not a single app but a gateway to agents that solved concrete tasks. Finance used an FAQ assistant to handle recurring questions. Legal teams accelerated crude contract reviews by drafting summaries and issue lists in minutes. HR and administration routed complex medical reimbursements through AiU to reduce manual back-and-forth and ease end-of-month backlogs. In operations, a safety briefing agent helped partner companies prepare faster with consistent, up-to-date guidance. Sales and PR analyzed customer feedback across channels with less effort and more consistency.
Adoption was broad because AiU lived where work already happened. For field teams who ran on mobile, the platform fit within daily routines rather than asking for new ones. The “AiU playground” lowered the barrier to entry: employees could try ideas, test prompts, and spin up simple agents without opening formal projects or waiting for external vendors. As small wins accumulated, teams pulled the platform into more processes, creating a self-reinforcing loop of experimentation and impact.
The platform also turned legacy into leverage. An underused enterprise search tool regained relevance once AiU added retrieval and summarization across Microsoft 365 documents, HR information, and work rules. Instead of launching a costly replacement program, the company layered AI on top to unlock value already stored in its systems. This pattern—augment first, replace later—generated quick wins, conserved funds, and built credibility for larger modernization efforts.
Lessons and next moves
Quotes from leadership framed the change in human terms. “Good risk taking,” Hur said, “is not moving fast for speed’s sake—it is moving responsibly so that learning compounds.” Lee described the shift in governance as flipping digital from reactive to proactive: “Don’t wait for a ticket; shape the roadmap with the business, measure outcomes, and retire what no longer serves.” Those statements were not slogans; they were guidelines used to prioritize, fund, and evaluate work.
Measurements followed the rhetoric. Usage data captured breadth—nearly the entire company tried AiU—and depth, with about 85% becoming regular users and mobile-only adoption around 95%. Performance indicators shifted from proxy IT metrics to outcomes: time-to-draft in legal cut significantly; administrative reimbursement cycles shortened; customer satisfaction improved where feedback classification fed faster responses; and cost avoided through system rationalization was reinvested in resilience and digital foundations. Monthly showcases surfaced best cases and documented failed attempts as institutional learning.
Industry watchers recognized the approach as aligned with a DAX mindset: advance digital and AI together, keep flexibility at the AI layer, and harden the data and application core. Externally, multi-model support hedged against a volatile market. Internally, the digital academy expanded skills through short courses and multi-month tracks in RPA, Tableau, Python, AI, and data science. Executives completed generative AI training, signaling that the expectation to learn applied to the top as much as to the front line.
A working playbook others could adapt
Several patterns translated into a practical operating system. First, codify culture: reward small, fast releases and publish lessons from failure as rigor, not theater. Second, architect for change: decouple the AI layer, support multiple LLMs, and secure data access inside daily workflows. Third, productize ownership: let line-of-business teams build and evolve agents in a low-friction playground, with the DX group as an enabler and curator rather than a gate. Fourth, measure what matters: tie efforts to productivity, cost, and customer experience; sunset tools that underperform; reinvest gains into resilience and data quality. Fifth, broaden talent: develop hundreds of practitioners, not a handful of specialists.
Governance backed the playbook. The portfolio was rationalized to remove redundancy and reduce underutilized spend, while selective outsourcing expanded beyond infrastructure into HR, procurement, and legal systems. Savings funded disaster recovery, business continuity, and shared components that supported AI at scale. Embedding AiU into existing tools reduced context switching, and monthly showcases kept momentum visible and expectations consistent.
The roadmap pointed toward a unified intelligence layer capable of answering practical questions on demand—“What’s new with this customer?”—by stitching together work-in-progress, recent interactions, and open issues across divisions. Maintaining flexibility at the AI layer while fortifying shared data models and workflows kept options open as models evolved, ensuring that the durable assets—the data and the business applications—compounded in value.
The bottom line
The story closed on execution rather than aspiration. By legitimizing “good risk taking,” building a model-agnostic platform anchored in core data, and pushing ownership to the line of business, GS Caltex had moved beyond pilots into enterprise habit. Agents improved legal throughput, trimmed administrative backlog, accelerated safety briefings, and pulled customer signals into decisions faster. System rationalization and selective outsourcing had reallocated spend toward resilience and DX foundations, while the digital academy scaled capability across the front line.
What remained instructive for other incumbents were the next steps that followed: extend embedded AI into more legacy processes before replacements, deepen single-view snapshots to shrink decision latency, continue rotating best frontline builders into center-led roles to spread practices, and keep performance reviews tied to business outcomes over proxy indicators. In short, the path forward had been about turning curiosity into capacity, capacity into outcomes, and outcomes into a durable advantage that accumulated with each release.
