OpenAI is no longer content to simply offer ChatGPT to businesses or to sell its models via API.
Since the arrival of GPT-4o, its technologies have gradually woven into enterprise software, from document management to CRM and cybersecurity.
This evolution can be read in three phases. First, the model becomes a feature of the software. Then it becomes the engine for agents capable of performing tasks. Finally, it begins to insert itself directly into business workflows.
GPT-4o: AI Becomes a Software Feature
The May 2024 launch of GPT-4o accelerates the diffusion of OpenAI’s models into professional applications. Its ability to process text, images, and voice enables publishers to add AI features to their products without requiring customers to switch environments.
Box thus integrates GPT-4o into Box AI to analyze content stored on its platform. Zendesk uses it in its customer service tools. Salesforce integrates it into its AI features. Intercom also relies on OpenAI’s models for its Fin support agent.
The model becomes then an invisible building block of the software. The user does not necessarily choose OpenAI; they use a function of their application, whose engine is partly supplied by OpenAI.
The shift is significant for OpenAI. Access to the enterprise market no longer goes solely through direct sales of ChatGPT or by a company deciding to connect its own apps to an API. Software publishers themselves become a distribution channel.
GPT-5: The Model as the Engine for Agents
The second phase emerges with reasoning models and agentic capabilities.
With GPT-5, OpenAI no longer highlights only content generation, but the model’s capacity to reason, use tools, and accomplish more complex tasks. Software publishers begin to incorporate these capabilities into their own agents.
Software development is the most visible field. GitHub Copilot, Cursor, Windsurf, Vercel, JetBrains, GitLab, Factory, Lovable, Augment Code, and Cognition are among the players that have integrated or tested GPT-5 in their products.
At JetBrains, GPT-5 is notably integrated into Junie, its development agent, as well as AI Assistant. At Vercel, the model powers notably v0 and the platform’s agent functions.
The difference from the first phase is substantial.
With GPT-4o, OpenAI powered features. With the reasoning models, it begins powering agents.
These agents can analyze context, use multiple tools, chain operations, and produce a result without the user having to lay out each step.
The same evolution appears outside of development. Harvey uses OpenAI models in its legal tools and workflows. Notion integrates them into its agents. Salesforce is developing Agentforce around this logic of agents capable of interacting with the company’s data and applications.
The integration of a model into software thus gradually changes in nature: it is no longer just about asking it to generate a response, but about entrusting it with part of the work.
2026: Models Insert into Business Workflows
The third phase is now especially visible in cybersecurity.
With Daybreak, OpenAI is coordinating a network of partners to integrate its specialized models into security products and services. OpenAI’s published list includes Akamai, Cato Networks, Check Point, Cisco, Cloudflare, CrowdStrike, Darktrace, Elastic, Fortinet, IBM, Okta, Palo Alto Networks, Proofpoint, SentinelOne, Sophos, Tenable, Trend Micro and Zscaler.
This time, the model does not simply add generation or search capacity to the software. It can intervene directly in an operational process.
At Cloudflare, OpenAI’s cyber models are used to help discover and remediate vulnerabilities. CrowdStrike integrates them into its security workflows. Elastic uses them to accelerate certain investigation and detection operations. Tenable pairs them with its exposure data to analyze risks related to agents, skills, MCP servers, and playbooks.
The principle is now different. The software provides the model with its data, context, and tools; the model brings its reasoning capabilities while the publisher retains control of the workflow and its governance mechanisms.
Cybersecurity is particularly revealing of this evolution because the consequences of a wrong decision can be immediate. Integrations must therefore combine the model’s capabilities with rules, action limits, supervision, and, in many cases, human validation.
But the logic extends beyond cybersecurity.
In data tools, CRMs, legal software, or collaborative platforms, the model no longer simply enriches the software; it participates in executing the tasks for which the software is designed.
OpenAI Shifts Position in the Software Chain
This evolution has significant economic implications.
The market is no longer limited to which model companies will choose. A new battleground is now playing out: in which software will this model be embedded?
A chief information officer can use a function powered by OpenAI without having chosen OpenAI as the model provider at the level of their AI infrastructure. They can work in GitHub Copilot, use Harvey for legal teams, leverage Notion, or rely on a cybersecurity platform that embeds GPT Cyber.
