One meaningful signal in the enterprise AI market is not a new model, but where models and agents are deployed. In May 2026, OpenAI and Dell Technologies announced a partnership to bring Codex into hybrid and on-premises environments, with the aim of letting enterprises use AI where their critical data, systems and workflows operate. This is not only an infrastructure story. It reflects a management shift: when AI starts touching real work, the deployment location and control mechanisms become part of value design.
During experimentation, organizations often begin with an accessible tool, a small group of users and low-risk situations. But when an agent needs to read customer data, query internal documents, assist engineers or update business systems, the question is no longer “which model gives the best answer?” It becomes: which data may the agent see, which tools may it call, what may it do and who is accountable when the result is wrong?
This helps explain why platforms such as OpenAI Frontier emphasize business-context connection, agent execution, evaluation-and-improvement loops and permission controls. AI value does not come only from a model’s reasoning ability. It comes from placing that model appropriately within the enterprise work architecture.
From a demo to an operating environment
A demo can be persuasive because it shows an agent completing a task under favourable conditions. An operating environment has different requirements: data change, policies contain exceptions, access rights vary, systems may slow down or fail and customers do not follow a script. Enterprises therefore need to design how an agent works within explicit limits, rather than only assess the quality of its answers.
Moving AI closer to existing systems can reduce integration friction and help address data, security or compliance requirements. However, on-premises should not be treated as automatically safe. An agent may still take a wrong action, access data too broadly or produce a decision that cannot be explained if permissions, logs and approval processes are not clearly designed.
Where data live does not replace data governance
A deployment decision needs to separate three issues that are often merged. The first is where data are stored or processed. The second is which data may enter an agent’s context. The third is which actions the agent may take using that output. An organization may keep data in its own infrastructure and still create risk if an agent is granted overly broad query rights or can change records without a verification step.
Data should therefore be classified by sensitivity and work purpose, not only by the name of a system. For every use case, enterprises should identify the minimum necessary dataset, the data owner, the access duration, the way sensitive information is masked and the signals that trigger review. This both reduces the risk surface and gives an agent enough context to create value instead of an ambiguous information repository.
Applied example
Example: an agent supporting customer requests
An insurance company wants to use an agent to help employees retrieve case information and prepare customer responses. At first, the agent should only read necessary data fields, cite the relevant policy and prepare a draft for an employee to approve. If the agent is allowed to amend a case record or make a compensation commitment, the company must design authority thresholds, approval steps, action logs and a way to handle exceptions.
The difference is that the agent is not evaluated as an answer box, but as a component of a service process. The metrics also change: not only response speed, but also the rate of cases requiring rework, policy compliance, human handoffs and the effect on customer experience.
Hybrid environments make management trade-offs visible
A hybrid environment may combine internal systems, cloud services, partner data and specialised tools. The benefit is that an enterprise does not need to replace its full existing architecture to test a new use case. In return, connection points require closer governance: which identity is used across systems, which permissions are inherited, which data may be retained and which activity must be audited.
This is why AI leaders should not direct technology questions only to IT. An architectural decision is also a decision about process, accountability, risk and operating capability. Business owners need to define outcome standards and important exceptions; security and legal teams need to define boundaries; technology teams need to translate those boundaries into access rights, technical controls and evidence that can be checked.
Five questions before scaling a use case
Before moving an agent to a larger scale, leaders should answer five questions clearly. First, which specific decision or work will the agent support? Second, what are the minimum data, tools and permissions it needs? Third, which results require human approval before they leave the system? Fourth, which metrics show that the agent is creating value without increasing error or risk? Fifth, when policy, data or systems change, who may update the agent and who is accountable for rechecking it?
These five questions do not slow AI transformation. They help enterprises avoid scaling a demo that lacks an operational owner. When answers are explicit, organizations can deploy step by step, learn from exceptions and scale on foundations that have been tested.
Implications for Vietnamese enterprises
For many Vietnamese enterprises, data are distributed across core software, spreadsheets, outsourced platforms and manual processes. An AI project should therefore start with a valuable but sufficiently narrow workflow, such as helping employees find policies, consolidate case files or prepare report drafts. Before scaling, the organization needs to clarify data quality, access rights and where people will review outputs.
The management message is not that every enterprise must deploy AI in on-premises infrastructure. The appropriate choice depends on the data type, compliance requirements, existing systems, operating capability and the value of the use case. The more important market signal is that AI is increasingly assessed as an operating layer connected to data and processes, rather than as a tool that sits outside work.
References
OpenAI. (2026, May 18). OpenAI and Dell Technologies partner to bring Codex to hybrid and on-premises enterprise environments. https://openai.com/index/dell-codex-enterprise-partnership/
OpenAI. (2026, February 5). Introducing OpenAI Frontier. https://openai.com/index/introducing-openai-frontier/
