“AI Isn’t the Answer to Every Problem” — Projects Without Business Value Get Dropped — BigGo Finance

Dell Technologies’ Asad Munshi, Chief Digital Officer (CDO) for Asia Pacific and Japan (APJ), emphasized on the 30th at the “Samsung AI Forum 2026” held at Samsung Electronics’ Seocho office in Seoul that enterprise AI adoption must be centered on business value, not technological showmanship.

“AI is not the answer to every problem,” he said during the session. “You have to first determine what needs to be done, what the constraints are, and what outcomes you want.” He explained that Dell adopted simplification, standardization, and automation as company-wide transformation principles before introducing AI. The logic is that automation can only be pursued after streamlining complex existing work processes and technology environments, and establishing common processes and platforms.

In fact, while business units within Dell proposed as many as 800 AI use case ideas, the company selectively applied only those with both high business value and high feasibility. This means ideas with low business value or difficult execution were decisively excluded.

Munshi identified data as the biggest bottleneck for enterprise AI. No matter how much data exists, if it is scattered across clouds, files, databases, and applications, AI cannot properly leverage it. “The future of AI is not in models but in data,” he stressed. “Real power comes from how well you manage and connect data.”

Dell is applying these principles to core areas including sales, software development, customer support, and supply chain. Currently, approximately 90% of Dell’s coding work is done with AI assistance under human supervision.

For the sales organization, Dell introduced an internal AI tool called “Sales Chat.” Previously, sales personnel spent 70% of their time on office work and 30% on customer-facing activities; now they can spend 60-70% of their time meeting customers. The internal conversational AI “Ask Dell” has 55,000 registered users processing 3.3 million requests per month.

Munshi distinguished between generative AI, which is closer to an “advisor” that answers questions, and agentic AI, which performs actual work across multiple steps using tools and data. However, he noted that not all tasks need to be converted to agents. Repetitive and standardized tasks may be better suited to traditional automation.

He also emphasized the need for control mechanisms as agentic AI expands. “AI is not a deterministic system but a probabilistic system,” he explained, noting that unlike existing enterprise resource planning (ERP) or billing and inventory management systems that always produce the same results according to defined rules, AI can produce different outcomes even with the same input.

Accordingly, he advised applying approved platforms and security frameworks, and clearly defining the scope of agent actions. He added that human oversight must be maintained while also equipping systems with a “kill switch” that can immediately halt operations when necessary. A kill switch is an emergency stop mechanism that allows humans to immediately halt system operations when an AI agent exhibits unexpected behavior or when problems arise.

He also made clear that token usage itself cannot serve as a performance metric. “It’s not about how many tokens you use,” he said. “You have to look at whether those tokens are doing productive work and creating real business value.”

In a keynote at the same forum, Richard Ho, Vice President of Hardware at OpenAI, presented real-world results from AI agent workflows. Referencing Samsung Electronics’ enterprise-wide adoption of ChatGPT Enterprise, he said, “Now that we have secured system access, the strategic challenge is to convert this into actually useful work outcomes and workflows based on each department’s domain knowledge.”

Ho was a key figure in the early development of Google’s AI accelerator, the TPU. He currently leads the “Habanero” project, OpenAI’s first AI inference chip co-developed with Broadcom in June. In his speech, he revealed how AI agents were applied to the Habanero development process.

OpenAI applied AI agent workflows to Habanero’s development, completing the process from circuit design creation to tape-out — the stage where the design is handed over to the manufacturer — in just nine months. By using AI for design optimization, the company reduced matrix operation unit area by 10% and SIMD operation unit area by 8% compared to conventional designs. “Even in complex engineering environments like semiconductor chip design, AI agent workflows actually work and produce remarkable results,” he emphasized.

The Samsung AI Forum, now in its 10th year, was held under the theme “Agentic Shift: From Intelligence to Impact.” Jun Young-hyun, Vice Chairman and CEO of South Korea’s Samsung Electronics, said, “Agentic AI will bring fundamental changes not just in technological innovation but in how we work and in organizational culture,” adding, “We will pursue substantive AI technology innovation.”

The forum also featured suggestions that physical AI requires a specialized approach. Ranjay Krishna, a professor at the University of Washington and a leading physical AI scholar, said, “The language and reasoning intelligence of large language models (LLMs) alone has limitations for physical AI operating in the physical world,” adding, “An architecture that combines small control models that respond to real-time changes with higher-level LLMs will be the key solution.”

Afternoon sessions were divided into the Device eXperience (DX) division and the Semiconductor (DS) division, sharing specific AI research outcomes and field application cases. The DS division presented on AX (AI transformation) innovation, while the DX division presented on AI technology.

The message repeatedly emphasized throughout this forum was not AI adoption itself, but conversion into actual work outcomes and business value. Dell’s case demonstrated a selective approach of applying only valuable projects out of 800 ideas, while OpenAI’s Habanero case proved that AI agents can deliver substantive design optimization results even in complex engineering domains.

As companies expand AI investments, this forum is notable for highlighting performance-driven application strategies and the importance of control mechanisms over the speed of technology adoption. Along with the outlook that agentic AI will bring fundamental changes to how work is done, the common conclusion presented was that human oversight and emergency stop mechanisms are essential given AI’s nature as a probabilistic system.