Olivier Godement, Head of Product for OpenAI’s API system, gave a behind-the-scenes look at how business groups are adopting and deploying AI officials at scale at VentureBeat’s Transform 2025 event.
The Actions API and Agents SDK, the current OpenAI API director, was the subject of a 20-minute panel discussion hosted entirely with Godement, who unveiled the Responses API and its latest developer tools while presenting early adopters like Stripe and Box’s real-world patterns, safety considerations, and cost-return examples.
Here are the top 8 most important insights for business leaders who are unable to attend the program life.
Brokers Are Quickly Transitioning From Manufacturing to Prototype
Godement believes that the way AI is being used at size will change significantly in 2025. With over a million active developers worldwide using OpenAI’s API platform each month and a 70 % increase in token usage annually, AI is moving beyond experimentation.
” The last five years have been pretty wild,” says one user who said, “it’s been five years since we launched essentially GPT-3 .”
Godement emphasized that bots are no longer the only thing in demand today. ” Artificial use cases are moving from basic Q&, A to real usage scenarios where the application, the broker, you do things for you.”
In response to this change, OpenAI released two significant developer-facing resources in March: the Actions API and the Agents SDK.
When Should You Apply Single Agents vs. Sub-Agent Architectures?
A significant concept was the choice of architecture. According to Godement, single-agent loops, which encapsulate whole tool entry and context in a single model, are theoretically elegant but frequently impractical at scale.
It’s challenging to create accurate and trustworthy one agents. Like, it’s truly difficult.
Teams frequently transition toward flexible architectures with particular sub-agents as complexity increases—more tools, more feasible user inputs, and more logic.
” You may complete isolation of concerns like in software,” according to the practice that has come to be, is to effectively divide the agents into several sub-agents.
These sub-agents act similarly to members of a small team: tier one agents handle routine issues, tier one agents handle routine issues, and others escalate or resolve edge cases.
Why Is the Responses API a Step Change?
The Responses API was Godement’s first step in the development of developer tooling. Previously, developers manually arranged sequences of model call. That orchestration is now handled internally.
The Responses API is arguably the biggest new layer of abstraction we’ve created since essentially GPT-3.
Developers can express their intentions without having to simply create model flows. The Response API handles that loop essentially, saying” You care about returning a really good response to the customer.”
It also includes built-in tools for agent workflows in the real world, such as knowledge retrieval, web search, and function calling.
Security and Observability Are Components of Everything.
The two foremost concerns were security and compliance. Godement cited key safeguards that allow the OpenAI stack to be used in regulated industries like finance and healthcare:
- Refusal decisions based on a policy
- SOC-2 logging
- Support for data residency
Godement sees the biggest gap between demo and production in the evaluation department.
” I think model evaluation is probably the biggest bottleneck for widespread AI adoption,” I think.
To help teams define what success looks like and track how agents behave over time, OpenAI now includes tracing and eval tools with the API stack.
It’s really difficult to develop that confidence, that assurance that the model is accurate and reliable unless you invest in evaluation.
Early ROI Is Remarkable in Several Functions
Some enterprise use cases are already producing tangible advantages. Godement gave examples from:
- Stripe, which uses agents to speed up the processing of invoices, reports” 35 % faster invoice resolution.”
- Box, the maker of the knowledge assistants that enable “zero-touch ticket triage”
Customer support ( including voice ), internal governance, and knowledge assistants for navigating lengthy documentation are other high-value use cases.
What It Takes to Get Started in Production
Godement emphasized the importance of people in effective deployments.
” Whenever they see a problem and see a technology, there is a small minority of very high-end people who run at it.”
These internal champions don’t always come from engineering. They have persistence, which unites them.
” OK, how can I make it work,” is their first reaction.
This group, who advocated early ChatGPT use in the enterprise and are now experimenting with full agent systems, is responsible for many initial deployments for OpenAI.
He also made a point about a gap that many people overlook: domain knowledge. Engineers are not the source of the knowledge in an organization. The ops teams are in charge of it.
Making agent-building tools accessible to non-developers is a challenge OpenAI aims to overcome.
What’s Next for Enterprise Agents?
Godement provided a roadmap a glimpse. OpenAI is actively developing:
- multimodal agents that can communicate in structured data, voice, text, images, and voice.
- Long-term memory for retaining knowledge throughout sessions
- Cross-cloud orchestration to support complex, distributed IT environments
These are iterative layers that expand what is already possible, not radical changes. ” When we have models that can think for minutes, for hours, not just for a few seconds,” says the engineer,” some pretty mind-blowing use cases will emerge.”
Final Thought: Reasoning Models Are Overhyped
Godement ended the session by reaffirming his conviction that reasoning-capable models, which are those that can reflect before responding, will be the true catalysts of long-term change.
” I still have faith that those models are about the GPT-2 or GPT-3 level of maturity.” What reasoning models can do are just scratching the surface.
The message to enterprise decision makers is clear: the infrastructure for agentic automation is present. Building a focused use case, empowering cross-functional teams, and being ready to iterate are now what matters. Not in novel demos, but in durable systems that are shaped by real-world requirements and the operational discipline to make them trustworthy is where the next step of value creation lies.
English 
Romanian 