Samsung Electronics has completed a dramatic reversal of its 2023 ChatGPT ban, deploying OpenAI's ChatGPT Enterprise and Codex to its entire South Korean workforce and all employees in its Device eXperience division globally as of June 21, 2026. This marks one of OpenAI's largest enterprise rollouts to date, affecting tens of thousands of workers across multiple continents. The move is significant not for its scale, but for who is using the technology: marketing teams, manufacturing staff, product designers, and corporate functions with no formal programming background.

What Changed Since Samsung Banned ChatGPT Three Years Ago?

In March 2023, Samsung implemented a company-wide ban on generative AI tools after engineers accidentally uploaded proprietary source code and confidential meeting records to the public version of ChatGPT. The data exposure left Samsung scrambling to contain the damage. The June 21 announcement represents a complete policy reversal, but with critical safeguards in place. This time, Samsung is using ChatGPT Enterprise, which features enterprise-grade security architecture including zero model training on customer data, identity-based access management, and data-loss-prevention controls.

The deployment also excludes Samsung's Device Solutions division, which handles semiconductor manufacturing and operates under tighter restrictions. This selective rollout reflects a calculated approach to risk management, allowing the company to expand AI adoption while protecting its most sensitive operations.

How Is Codex Being Used Beyond Software Development?

Codex began as a tool for writing, reviewing, and debugging code. At Samsung, it is being deployed to roles that have never written a line of code in their lives. The distinction matters because it challenges a long-standing assumption about AI automation: that non-technical workers need to learn new skills to benefit from AI tools. Instead, Codex operates on a different principle entirely.

According to OpenAI, more than 5 million people now use Codex every week across technical and non-technical workflows. Knowledge workers outside engineering now represent roughly 20 percent of that user base and are growing at three times the rate of developers. The fastest-growing tasks among non-technical Codex users include data analysis, research, and the creation of reports and work documents.

In South Korea specifically, weekly active Codex users grew by nearly 800 percent between February 1, 2026, and the June 21 announcement date. This explosive growth reflects the momentum OpenAI was riding when it locked in Samsung as a major enterprise customer.

"This is not simply about introducing AI as a workplace tool. It marks the starting point for fundamentally transforming the way we work and execute," said Roh Tae-moon, President and Head of Samsung's DX Division.

Roh Tae-moon, President and Head of Samsung's DX Division

How Does Codex Actually Work for Non-Programmers?

The architecture that makes Codex accessible to non-technical workers relies on what OpenAI calls an agent loop. When a user submits a request, Codex constructs a multi-layer prompt that stacks environment context, project-specific instruction files, sandbox permission rules, and the user's message. This prompt is sent to the underlying model for inference. If the model's response is not a final answer but a tool call (such as "run this shell command" or "read this file"), the agent carries out the tool call, appends the result to the prompt, and sends the whole thing back to the model again. This cycle can repeat dozens of times before the user sees a response.

A single request like "build a workflow that extracts last week's sales data and formats it as a summary report" might trigger file reads, data queries, code execution, and result validation before returning a finished artifact. The user never needs to understand what is happening behind the scenes.

Steps to Understanding Codex's Technical Foundation

• Agent Loop Architecture: Codex uses an iterative process where the model makes decisions, executes tool calls, and refines its approach based on results, repeating until a task is complete without requiring user intervention between steps.
• Isolated Sandboxes: Each task runs in its own isolated cloud sandbox preloaded with relevant repository or data context, preventing a manufacturing employee's workflow from interacting with a semiconductor engineer's codebase and maintaining security at scale.
• Prompt Caching via Prefix Property: Every new Codex task appends fresh content to the end of an existing prompt, allowing OpenAI to reuse computation from prior inference calls and making company-wide agentic AI economically viable at scale.
• Conversation Compaction: When conversations grow long enough to hit the model's context window, Codex replaces the full conversation history with a compressed representation that preserves the model's understanding of the task through an encrypted payload.

The model powering Codex as of 2026 is GPT-5.5, OpenAI's agentic-first base model, which replaced the earlier codex-1 variant fine-tuned from the o3 reasoning model. The prefix caching mechanism is not an incidental optimization; it is what makes running millions of iterative agent interactions daily financially sustainable.

Could Codex Replace the Traditional IT Request Queue?

Low-code and no-code platforms have promised for years to enable non-technical workers to build software. Microsoft Power Apps, Salesforce Lightning, and similar tools put drag-and-drop application builders in the hands of business analysts and operations staff. They achieved partial adoption, but never crossed from niche to default because they required workers to learn a new paradigm: visual programming. That paradigm remained unfamiliar to most employees.

Codex differs in one structural way: the interface is natural language. There is no paradigm to learn. An employee describes what they want; the agent loop handles the translation from description to working code. This does not eliminate the need for human judgment; the output still requires review, and complex tasks require iteration. But it removes the step where an idea requires someone with engineering credentials to carry it further.

Samsung's deployment is the first large-scale test of whether this model can displace the IT request queue as an organizational institution. The traditional enterprise workflow involves a non-technical employee submitting a request, waiting for prioritization, working with a developer to specify requirements, waiting again for delivery, and receiving a tool that may or may not match what they envisioned. The Codex workflow is different: that same employee describes the idea, iterates in natural language with the agent, and may have a working prototype in the same session.

NVIDIA reached a comparable milestone earlier in 2026, deploying Codex to more than 10,000 employees across both engineering and non-engineering functions. Samsung's deployment is larger and represents a more ambitious test of whether agentic AI can fundamentally reshape how enterprises handle software development and automation requests.

"Samsung Electronics is adopting AI not as a tool specific for certain teams or tasks but as a core platform to enhance the way employees work and innovate globally," explained Harrison Kim, General Manager of OpenAI Korea.

Harrison Kim, General Manager of OpenAI Korea

The stakes are high. If Samsung's deployment succeeds, it could signal a shift in how enterprises think about IT infrastructure, workforce productivity, and the role of non-technical workers in software development. If it falters, it may reveal fundamental limitations in how well AI agents can handle the complexity and context-sensitivity of real-world business problems.