As companies weigh artificial intelligence deployments based on business results rather than technical performance, building trust in integration with existing systems and in security and control frameworks has emerged as the central challenge. Cases of AI sharply cutting task processing times are growing, but companies are stepping up technical and procedural vetting before applying the technology to actual work.
Amazon Web Services held a press briefing for AWS AI WAVE 2026 at the Grand InterContinental Seoul Parnas in Seoul's Gangnam District on the 23rd, saying corporate customers' interest is shifting from AI technology itself to concrete improvements in work processes and measurable results. Choi Young-jun, head of AI technology at AWS Korea, and Kwon Soon-il, chief strategy officer at Upstage, discussed changes in corporate AI use, how agentic AI can be applied to real work, and the direction of cooperation between AWS and Upstage.
"In the past, many companies simply said they wanted to adopt AI. This year, they start by asking exactly which tasks they want to improve," Kwon said. "Recently, many companies set key performance indicators in advance before adopting AI."
Companies reviewing AI deployments are demanding concrete evidence of how much working time can be cut, rather than model performance. Citing a U.S. insurance company, Kwon said: "In the past, the focus was on explaining how good the AI technology was. Recently, how much task processing time can be reduced has become the yardstick for proof-of-concept decisions." After adopting AI, the company cut review time from 50 to 60 minutes per case to six minutes. Companies are paying more attention to efficiency gains than to model performance.
Choi, the AWS Korea AI technology chief, said GS Retail (007070.KS) used an AI agent built on Amazon Bedrock to respond to operational failures, cutting the average time for root-cause analysis to about two minutes from 30 minutes.
Higher efficiency alone does not immediately broaden AI adoption. "Established companies already have work procedures and systems in place, so AI that ignores them and simply produces results cannot actually be used," Kwon said. "Demand is growing to verify integration with legacy systems and compliance with governance." Upstage offers a platform called Studio that allows companies to define and manage existing work procedures.
Kwon identified trust as the biggest obstacle to spreading agentic AI in the workplace. "Technology and cost issues will largely be resolved over time, but social consensus on whether AI can be trusted with work is still lacking," he said.
Citing a comparison of performance on document interpretation tasks, he said: "People score 91 to 92 points, including errors, while AI comes close to 98 points, yet many companies are still reluctant to apply it without human involvement." He attributed this to the technology's relative newness and to a shortage of evidence showing whether AI adheres to procedures and governance. That is improving in the agent era, he said, and trust will build further as controls such as security measures become visible.
Verifying that AI complies with existing work rules is especially important in regulated industries such as finance, he stressed. "Even if AI can assess creditworthiness accurately, whether a financial firm will trust and use it is a separate question," Kwon said, describing an approach that automates follow-up work while keeping the rules of existing systems intact.
Choi said AWS supports companies in limiting the scope of AI data access and execution rights through guardrails and policy controls in AgentCore. "It is important to put control frameworks in place so that companies can comply with security requirements and internal rules as they apply AI to existing work," he said.
The approach to adoption should also vary with each company's circumstances. "One option is for management to lead a shift in core tasks where the benefits outweigh the cost of oversight and control; another is to apply four or five small automation projects first to build trust within the organization," Kwon said. "Rather than one being better, the choice should match the company's decision-making structure."
Upstage, meanwhile, is continuing cooperation on model training optimization and joint sales under a strategic partnership agreement signed with AWS in 2024. "Based on the technical experience and business track record we have built at home, we hope our cooperation with AWS expands into global markets as well," Kwon said.