Join the event trusted by corporate leaders for almost two decades. VB Transforma connects people building AI Real Enterprise. Learn more
On VentureBeat’s Transform Conference 2025Olivier GODEMENT, the head of the API OPENAI platform product, provided a behind -the -scene look at how corporate teams accept and implement AI agents on a vast scale.
In a 20-minute panel discussion, I hosted only with Godem, former researcher Stripe and the current head of OPENAI API unpacked the latest tools of OPENAI-UP programmers and SDK agents and at the same time emphasizing real patterns, security considerations and examples of reimbursement of costs from early adopters.
For corporate leaders, they cannot participate in a live session, here are the 8 most significant take -out:
Agents quickly pass from the prototype for production
According to Godem 2025, it means a real change in a vast -scale AI way. With over a million energetic programmers using the OPENAI API platform around the world, and the consumption of tokens increased by 700% year on year, and goes beyond experiments.
“Five years have passed since we triggered the GPT-3 … and man, the last five years were quite wild.”
GODEMENT emphasized that the current demand no longer applies to chatbots. “Cases of use of artificial intelligence go from simple questions and answers to the actual use of cases in which the application, agent, can do things for you.”
This change prompted OpenAI to launch the two main tools addressed to programmers in March: The API answers and SDK agents.
When to exploit individual agents compared to subordinate architecture
The main topic was the architectural choice. GODEMENT noticed that disposable loops, which contain full access tools and context in one model, are conceptually elegant, but often impractical on a vast scale.
“Building accurate and reliable individual agents is difficult. It’s really difficult.”
With the escalate in the complexity of tools, more possible input data of users, more logic-female often head towards modular architecture with specialized sub-agents.
“Practice has appeared, there is basically to break up agents into many sub-agents … You would make a distribution of fears as in software.”
These sub-agents act like roles in a tiny team: Sicing agent classifies intentions, agents with one level support routine problems, while others escalate or solve the edge.
Why API Answers is a change of steps
GODEMENT has established API of answers as a fundamental evolution in the instrumentation of programmers. Earlier, developers manually organized sequences of model connections. Now this orchestration is supported internally.
“API of answers is probably the largest new layer of abstraction that we have introduced because almost GPT-3.”
It allows programmers to express intentions, not just configuring the flows of models. “You care about returning a really good answer to the customer … API basically serves this loop.”
It also includes built-in possibilities of searching for knowledge, searching websites and causing functions-exhorts whose enterprises need an agent’s work in the world.
Observation and safety are built -in
Safety and compatibility were the most significant. GODEMENT quoted key handrails that make the pile profitable OpenAI for regulated sectors, such as finance and healthcare:
- Policy -based refusals
- SOC-2 login
- Data residence support
The assessment consists in the fact that Godem sees the largest gap between demo and production.
“My hot shot is that the assessment of the model is probably the biggest bottleneck for mass adoption AI.”
Opeli now includes tracking and tools to evaluate with a pile of API to aid teams determine what success looks like and track how agents work in time.
“Unless you invest in a rating … It is really difficult to build this trust, the certainty that the model is accurate, reliable.”
Early roi is noticeable in specific functions
Some cases of using an enterprise already provide measurable profits. GODEMENT shared examples from:
- Stripewhich uses agents to accelerate invoice, reporting “35% faster invoice resolution”
- Boxwho introduced knowledge assistants who enable “Zero-Dotyk ticket triage”
Other cases of high value include customer service (including voice), internal management and knowledge assistants to navigate dense documentation.
What you need to start in production
GODEMENT emphasized the human factor in successful distributions.
“There is a small fraction of a very high class of people who, when they see the problem and see technology, run on it.”
These internal masters do not always come from engineering. What unites them is perseverance.
“Their first reaction is: ok, how can I work?”
Opeli sees a lot of initial implementations driven by this group – people who pushed the early exploit of chatgpt in the enterprise and are now experimenting with full agents’ systems.
He also drew attention to the gap referred to: specialist knowledge in the field. “Knowledge in the enterprise … does not lie with engineers. It lies in OPS teams.”
Sharing tools for building agents for non -inorganizers is a challenge that OpenAI aims to solve.
What next with corporate agents
GODEMENT took a look at the road map. Opeli is actively working on:
- Multimodal agents which may affect the text, voice, images and structured data
- Long -term memory to maintain knowledge in various sessions
- Cross orchestration To support intricate, distributed IT environments
These are not radical changes, but the iterative layers that expand what is already possible. “When we have models that can think not only for a few seconds, but for a few hours … It will allow some stunning cases of use.”
Final word: reasoning models are under the under
GODEMENT closed the session, confirming his belief that models serving reasoning-these can reflect before the answer-the real ones will be real enabling long-term transformation.
“I will continue to find out that we are almost at the level of GPT-2 or GPT-3 maturity of these models … We are still scratching the surface, what models of reasoning can do.”
In the case of company decision makers, the message is clear: the infrastructure for agency automation is here. What counts is to build a targeted exploit case, strengthening interfunctional teams and readiness for iteration. The next phase of creating values is not about novel demonstration versions-but in lasting systems, shaped by needs in the real world and operating discipline to make them reliable.