AI becomes more useful not when it answers better, but when it can integrate into the workflow.
At DevDay, OpenAI showed a combination that seems more practical to me than another "smarter and faster": Agents API for task execution and Decisions API for quick decisions and routing. An agent can receive a task, determine the next step, call an external tool, and return the result. Decisions API handles narrow forks: classify the request, choose the route, understand whether escalation to a human is needed.
This is an important shift. We stop building business logic around a single chatbot and start designing workflows where the model is one component of the process.
But there's a false temptation to call such a system autonomous and release it without constraints. In a real product, you still need access permissions, limits, action logs, and confirmation points. Otherwise, "an agent that does things itself" quickly turns into "an agent that everyone double-checks."
For developers and architects, the main question now is not "which model to choose?" but "which decisions can be safely delegated, and which should be left to humans?" The cost of integration with accounting systems and the real benefit of the new APIs will depend on this.
Source: Latent Space breakdown.