Samsung Electronics deployed ChatGPT Enterprise and Codex to its complete South Korean workforce and to all workers in its System eXperience division worldwide on June 21, 2026, marking what OpenAI described as certainly one of its largest enterprise rollouts to this point. The DX division — which builds Galaxy smartphones, house home equipment, and client electronics — spans tens of 1000’s of workers throughout a number of continents. The transfer completes a reversal three years within the making: in March 2023, Samsung banned generative AI instruments company-wide after engineers unintentionally uploaded proprietary supply code and confidential assembly data to the general public model of ChatGPT, exposing knowledge that left Samsung’s management earlier than the corporate realized what had occurred.
What makes the June 21 announcement distinct from Samsung’s earlier multi-vendor AI rollout, introduced June 9, is specificity: that is the formal settlement overlaying ChatGPT Enterprise’s enterprise-grade safety structure — zero mannequin coaching on buyer knowledge, identity-based entry administration, data-loss-prevention controls — alongside Codex, the agentic software-building tool OpenAI launched as a cloud-based coding agent in 2025. Samsung’s System Options division, which handles semiconductor manufacturing, operates beneath tighter restrictions and isn’t a part of the worldwide Codex deployment. The complete Korea workforce and world DX division are in scope.
Codex Was Constructed for Builders — Samsung Is Giving It to Everybody Else
The headline element within the Samsung deployment will not be the size. It’s the viewers. Codex started as a device for writing, reviewing, and debugging code. At Samsung, it’ll advertising groups, manufacturing workers, product designers, and company features — roles with no formal programming background.
Roh Tae-moon, President and Head of Samsung’s DX Division, framed the shift in phrases that went past productiveness: “This isn’t merely about introducing AI as a office device. It marks the place to begin for essentially remodeling the best way we work and execute.”
Harrison Kim, Basic Supervisor of OpenAI Korea, made the identical distinction specific: “Samsung Electronics is adopting AI not as a device particular for sure groups or duties however as a core platform to boost the best way workers work and innovate globally.” The implication is direct: a advertising analyst who has by no means written a line of code can now describe an inside device they want, and Codex will try to construct it.
Greater than 5 million folks now use Codex each week throughout technical and non-technical workflows, according to OpenAI. Data employees — folks exterior engineering — now symbolize roughly 20 % of that consumer base and are rising at thrice the speed of builders. The fastest-growing duties amongst non-technical Codex customers are knowledge evaluation, analysis, and the creation of stories and work paperwork. In South Korea particularly, weekly energetic Codex customers grew by almost 800 % between February 1, 2026, and the June 21 announcement date, reflecting the momentum OpenAI was using when it locked in Samsung.
How Codex Truly Works: Agent Loop, Sandboxes, and the Prefix Cache
What permits a device initially constructed for knowledgeable engineers to be deployed meaningfully to a company-wide viewers is an structure that doesn’t require the consumer to grasp programming — solely to explain an issue.
On the heart of Codex is an agent loop. When a consumer submits a request, Codex constructs a multi-layer immediate — stacking setting context, project-specific instruction information, sandbox permission guidelines, and the consumer’s message — and sends it to the underlying mannequin for inference. If the mannequin’s response will not be a closing reply however a device name (“run this shell command,” “learn this file,” “execute these exams”), the agent carries out the device name, appends the end result to the immediate, and sends the entire thing again to the mannequin once more. This cycle can repeat dozens of instances earlier than the consumer sees a response. A single request like “construct a workflow that extracts final week’s gross sales knowledge and codecs it as a abstract report” would possibly set off file reads, knowledge queries, code execution, and end result validation earlier than returning a completed artifact.
Every job runs in its personal remoted cloud sandbox, preloaded with no matter repository or knowledge context is related. That isolation is what makes company-wide deployment manageable for an organization with Samsung’s safety necessities: a producing worker’s workflow can not work together with a semiconductor engineer’s codebase.
The engineering element that retains this economically viable at scale is prompt caching via the prefix property. Each new Codex job appends recent content material to the top of an present immediate, which suggests the outdated immediate is at all times a precise prefix of the brand new one. That prefix property lets OpenAI reuse computation from prior inference calls — so though the uncooked knowledge despatched to the API grows quadratically as a dialog extends, the precise mannequin computation stays nearer to linear. This isn’t an incidental optimization. It’s what makes working thousands and thousands of iterative agent interactions each day financially sustainable. The cache is fragile — altering the mannequin, adjusting device order, or altering sandbox configuration breaks it — however when it holds, it’s the mechanism that makes the economics of company-wide agentic AI work.
When conversations develop lengthy sufficient to hit the mannequin’s context window, Codex compacts: it replaces the complete dialog historical past with a compressed illustration that preserves the mannequin’s understanding of the duty by way of an encrypted payload, then continues from there. The mannequin powering Codex as of 2026 is GPT-5.5, OpenAI’s agentic-first base mannequin, which changed the sooner codex-1 variant fine-tuned from the o3 reasoning mannequin.
Citizen Growth at Industrial Scale
Low-code and no-code platforms have promised for years to allow non-technical employees to construct software program. Microsoft Energy Apps, Salesforce Lightning, and related instruments put drag-and-drop utility builders within the palms of enterprise analysts and operations workers. They achieved partial adoption, however by no means crossed from area of interest to default as a result of they required employees to study a brand new paradigm — visible programming — that remained unfamiliar to most.
The Codex mannequin differs in a single structural means: the interface is pure language. There is no such thing as a paradigm to study. An worker describes what they need; the agent loop handles the interpretation from description to working code. This doesn’t remove the necessity for human judgment — the output nonetheless requires overview, and complicated duties require iteration — nevertheless it removes the step the place an thought requires somebody with engineering credentials to hold it additional.
Samsung’s deployment is the primary large-scale check of whether or not this mannequin can displace the IT request queue as an organizational establishment. The normal enterprise workflow: a non-technical worker submits a request, waits for prioritization, works with a developer to specify necessities, waits once more for supply, and receives a device that will or might not match what they envisioned. The Codex workflow: that very same worker describes the concept, iterates in pure language with the agent, and should have a working prototype in the identical session.
NVIDIA reached a comparable milestone earlier in 2026, deploying Codex to greater than 10,000 workers throughout each engineering and non-engineering features. Samsung’s deployment is bigger in declared scope and, as a result of it contains manufacturing and advertising employees at an organization of Samsung’s breadth, represents probably the most expansive company wager but that enterprise AI deployment has moved past being a developer story.
The Safety Structure That Made the Reversal Potential
Samsung’s 2023 ban was not a reflexive response — it was a exact analysis. Engineers importing code to a client AI device didn’t represent a safety breach within the conventional sense; no attacker was concerned. What occurred was that knowledge left Samsung’s perimeter completely as a result of the buyer model of ChatGPT might retain consumer inputs and doubtlessly use them for mannequin coaching. That publicity was the issue.
ChatGPT Enterprise resolves this on the contract degree. Enterprise agreements with OpenAI decide to not coaching fashions on buyer knowledge by default, and to data-retention controls and processing agreements that the buyer product doesn’t carry. Samsung bolstered this with its personal gate: entry is granted solely to workers who full inside AI safety coaching, a requirement the corporate constructed after working a two-month proof-of-concept from April to Might 2026 during which 2,500 DX division workers examined ChatGPT, Gemini, and Claude concurrently.
The groundwork was laid by way of Samsung SDS, Samsung’s IT companies arm. On December 23, 2025, Samsung SDS signed a reseller partnership with OpenAI, turning into the primary Korean firm licensed to ship and technically help ChatGPT Enterprise for native shoppers. That deal got here six weeks after Samsung Chairman Lee Jae-yong and OpenAI CEO Sam Altman signed a broader letter of intent on October 31, 2025, overlaying cooperation throughout semiconductors, knowledge facilities, and cloud infrastructure beneath OpenAI’s Stargate program — the context that finally made a workforce-level AI deployment the following logical step.
OpenAI’s Pre-IPO Enterprise Push
The Samsung deal lands at a particular second in OpenAI’s company historical past. The corporate filed its confidential S-1 registration assertion with the U.S. Securities and Change Fee on Might 22, 2026, with the submitting publicly confirmed on June 8. Goldman Sachs and Morgan Stanley are main the providing, with a goal public debut as early because the fourth quarter of 2026 at a valuation analysts estimate between $852 billion and $1 trillion. Enterprise income now accounts for greater than 40 % of OpenAI’s roughly $2 billion in month-to-month income and is on observe to achieve parity with client income by yr’s finish.
A deployment of Samsung’s scale — introduced 13 days after the S-1 affirmation — features each as a business milestone and as a showcase for enterprise buyers evaluating OpenAI’s capacity to win and retain the world’s largest firms as prospects.
Samsung itself is already a part of OpenAI’s {hardware} provide chain: Samsung Electronics is supplying superior reminiscence semiconductors for OpenAI’s next-generation AI infrastructure. The June 21 deployment expands that relationship in a qualitatively completely different course — from supplying the chips that run OpenAI’s fashions to embedding these fashions into the each day workflows of Samsung’s world workforce.
The Korean enterprise AI market displays this consolidation. OpenAI’s roster of Korean shoppers now contains LG Electronics, LG Uplus, LG CNS, GS E&C, Samsung SDS, TVING, Krafton, Toss, MUSINSA, Korea Zinc, Nexen Tire, HanaTour, Day1Company, and Worksphere. IT-services arms of all three of Korea’s largest conglomerates — Samsung SDS, SK AX, and LG CNS — now carry OpenAI distribution agreements, creating an institutional layer that routes enterprise AI choices by way of channels already embedded in Korean company operations.
What Samsung Plans to Measure
Samsung has set a goal to finish coaching for its full world workforce by the top of 2026. What it has not publicly disclosed is the way it will measure success — adoption charges, device high quality scores, productiveness metrics, or output quantity — although the safety structure requires workers to finish inside coaching earlier than gaining entry, which can produce no less than a pure enrollment funnel.
The outcomes shall be carefully watched by the various enterprises now evaluating their very own company-wide AI methods. If Samsung can reveal that non-technical employees utilizing Codex construct instruments that cut back the load on engineering groups, the IT request queue might show much less sturdy than most firms assume.
Ceaselessly Requested Questions
What’s ChatGPT Enterprise and the way does it differ from common ChatGPT?
ChatGPT Enterprise is a contract-tier product designed for company deployments. Not like the buyer model, it commits by default to not prepare OpenAI’s fashions on buyer knowledge, supplies identity-based entry administration and data-loss-prevention controls, helps zero-data-retention agreements, and contains safety governance options that permit firms to deploy AI inside their present compliance frameworks. Samsung’s 2023 ban was triggered partly by the truth that client ChatGPT lacked these protections.
How does OpenAI Codex work for non-technical employees?
Codex makes use of an agent loop that interprets pure language descriptions into working software program. A non-technical worker describes a job — a workflow, an inside device, a knowledge report — and the agent iteratively calls instruments, runs code, reads information, and executes exams inside an remoted cloud sandbox till it produces the end result. The consumer sees the output, not the underlying code, until they select to overview it. The mannequin working Codex as of 2026 is GPT-5.5, OpenAI’s agentic-first base mannequin.
Why did Samsung ban generative AI in 2023 and reverse course in 2026?
In March 2023, Samsung engineers uploaded proprietary supply code and confidential assembly data into the general public model of ChatGPT, which might retain consumer inputs and doubtlessly use them for mannequin coaching, sending Samsung’s knowledge exterior its perimeter. Samsung banned generative AI instruments company-wide in response. The reversal turned potential when enterprise contract tiers launched data-use commitments that resolved the unique concern, and when Samsung constructed its personal access-control gate requiring inside AI safety coaching earlier than any worker can use the instruments.
What does the Samsung deployment imply for enterprise software program improvement?
Samsung’s rollout is among the many first company-wide deployments of an agentic coding device to non-technical employees at industrial scale. If the deployment succeeds in enabling advertising, manufacturing, and company workers to construct useful software program instruments with out engineering help, it’s going to check whether or not the IT request queue — the standard bottleneck between an worker’s thought and dealing software program — might be meaningfully shortened by AI brokers. The end result will affect how different massive enterprises construction their very own AI transformation methods.