支援から実行へ：企業はどのようにAIを活用しているかFrom assistance to execution: How enterprises put AI to work

OpenAI has published research examining how enterprises are moving beyond experimental use of generative AI toward workflows in which software systems carry out multi-step tasks with limited human oversight. The findings matter because they suggest a shift in how organizations value these tools: from conversational assistants that draft text or answer questions to what the industry calls agentic AI, where models plan, take actions, and complete work autonomously. According to the research, ChatGPT and Codex are central to this transition, and firms at the frontier of adoption appear to be pulling significantly ahead of their peers.

The central theme is the distinction between assistance and execution. In an assistance model, an employee prompts a system, reviews the output, and integrates it manually into a task. In an execution model, the system is given a goal and is trusted to carry out a sequence of steps—retrieving information, calling tools, generating code, or updating records—before returning a finished result. OpenAI's research frames this as a maturing pattern across enterprises rather than a single feature, and it positions ChatGPT and Codex as the primary vehicles through which companies are making that move.

Codex is OpenAI's coding-focused agent, designed to handle software engineering tasks such as writing functions, fixing bugs, running tests, and working across a codebase rather than producing isolated snippets. Its inclusion in the research reflects how software development has become one of the earliest and most measurable areas for agentic deployment, since code can be executed, verified, and iterated on within defined boundaries. ChatGPT, meanwhile, serves broader knowledge work, and enterprise offerings such as ChatGPT Enterprise and connected tools have been positioned to bring these capabilities into corporate environments with administrative controls, data protections, and integration options.

A recurring point in the research is the gap between so-called frontier firms and everyone else. Frontier adopters are described as organizations that have restructured workflows, invested in internal expertise, and built the governance needed to let AI systems act rather than merely advise. The suggestion is that these firms are capturing disproportionate benefits, which is consistent with a broader pattern in technology adoption where early and committed implementers tend to compound their advantages over time. It is worth treating this as an observed trend rather than a guaranteed outcome, since adoption maturity and measured results can vary widely by industry and by task.

For context, agentic AI relies on several underlying components that have advanced in parallel. Tool use, sometimes called function calling, allows a model to invoke external systems and APIs. Retrieval methods connect models to a company's own documents and data so responses are grounded in internal knowledge. Longer context windows and improved reasoning let models sustain multi-step tasks, while orchestration frameworks coordinate how an agent breaks down and sequences work. OpenAI has released developer tooling in this direction, and competing efforts from other major AI providers point to an industry-wide bet that agents, not chat alone, represent the next phase of enterprise value.

The move toward execution also raises practical considerations that enterprises are still working through. Granting a system the authority to take actions increases the importance of guardrails, permissions, audit trails, and human review at critical checkpoints. Reliability remains a concern, because an autonomous process that makes an error can propagate it across several steps before a person notices. Questions of security, data handling, and compliance become more acute when agents touch production systems or sensitive records. These factors likely explain why the research emphasizes organizational readiness alongside the technology itself.

Taken together, the research portrays a market in transition rather than a finished shift. Many organizations appear to remain in the assistance phase, using AI to accelerate individual tasks, while a smaller group is redesigning processes around autonomous execution. For companies evaluating their own strategy, the implication is that the tools are increasingly capable of doing more than drafting and summarizing, but that realizing that capability depends on investment in workflow design, oversight, and staff skills. As with earlier waves of enterprise software, the eventual winners are likely to be defined less by access to the technology and more by how deliberately they integrate it into how work actually gets done.