The Reality of the 'Agent Era' Shown by OpenAI and Databricks: How GPT-OSS, Governance, and Data Infrastructure Are Changing Management
AI adoption for businesses is finally moving from 'experimentation' to 'core infrastructure'.
The symbol of this is the partnership between OpenAI and Databricks. In this article, based on the conversation between Sam Altman (OpenAI CEO) and Ali Ghodsi (Databricks CEO), we will organize the 'next phase of enterprise AI' and the 'future of GPT-OSS'.
1. The Essence of the OpenAI × Databricks Partnership: The 'Last Mile' of Models and Data
Databricks provides the enterprise data infrastructure, and OpenAI provides the cutting-edge models. By natively integrating the two, this partnership offers a complete package of 'high-performance models × proprietary enterprise data × governance'.
Ali puts it this way.
“Every enterprise customer wants to use OpenAI. However, data is confidential, and requirements such as privacy, auditing, and GDPR must also be met.”
Meanwhile, Sam also emphasizes that OpenAI's focus is shifting significantly from consumers to enterprises.
“Enterprise is now one of our biggest focuses. Enterprise usage has grown 7x this year.”
The point is the reality that enterprise adoption will not progress just because 'there is a good model'.
An infrastructure that can safely handle confidential data and Databricks as the glue that embeds models into business workflowsare the keys to OpenAI's growth strategy.
2. The Next Phase: AI Agents and the Extension of the 'Task Horizon'
2-1. The 'Task Horizon' Expanding from 5 Seconds → 5 Minutes → 5 Hours
An interesting metric Sam mentioned is the 'length of a task that a model can complete with a 50% probability = task horizon'.
Early GPT-3.5: Tasks at the 'few-second level', such as coding
GPT-4 generation: Tasks at roughly the '5-minute level'
GPT-5 generation: Already tasks at the '5-hour level'
“Many enterprise tasks are composed of tasks that take months to years. Extending the task horizon to that point will be an important research theme from now on.”
In other words, the view is that 'agents' that work over long periods while maintaining context, rather than one-off prompt responses, are becoming the main battlefield.
2-2. Context is the Greatest Lever
Ali, while touching on the 'context optimization technology inspired by genetic algorithms' developed by Databricks, explains:
“Automatically picking only the 'context needed right now' from vast amounts of internal corporate documents and providing it to the model. There is no need for manual prompt tuning.”
he explains.
Rather than relying solely on the intelligence of the model, 'what context to provide, when, and how' is what determines agent performance — this is the shared understanding between OpenAI and Databricks.
3. The field that AI agents have begun to change: Turning tasks that were not done into 'tasks that can be done'
The specific examples Ali cited are quite vivid.
Pharmaceuticals (AstraZeneca): An agent reads 400,000 documents at once to extract information necessary for research and regulatory compliance
Financial Institutions: An agent reads SEC filings and related materials to provide analysts with the 'seeds' of investment ideas
Insurance and Healthcare: Feeding vast amounts of hospital documents into an LLM to automate risk assessment and the extraction of critical information
Sam points out here that 'not only are tasks being done faster and cheaper, but tasks that no one was doing at all are being created'.
'Features and analyses that we thought were great ideas but that we could never get around to are now testable just by adding one agent'
'Tasks that we had given up on because we couldn't do them' are, for the first time, becoming economically rational thanks to AI — this perspective shows an impact that goes beyond mere productivity improvement.
4. The bottleneck for adoption will be 'governance,' not 'intelligence'
As model performance increases, corporate concerns shift from 'intelligence' to 'control'.
Ali lists the following as 'essential features' in the Databricks × OpenAI integration.
Audit logs that can track every operation
Access control by user and department
Checks to ensure model output aligns with brand guidelines and compliance
Business guardrails, such as 'are you recommending a competitor's product?'
Sam agrees, asserting that
'The 'constraint' for enterprise AI adoption is not intelligence or price, but governance'.
This realization indicates that the phase has shifted from 'trying out LLMs within the company' to 'building an infrastructure that can be entrusted with core business operations'.
This realization indicates that the phase has shifted from 'trying out LLMs within the company' to 'building an infrastructure that can be entrusted with core business operations'.
5. The Future of GPT-OSS: Will the Day Come When We Run GPT-5 Class Models 'Locally'?
OpenAI's 'GPT-OSS (Open Weights Model)' was also a key topic of the conversation.
Sam summarizes the current situation as follows.
The demand for the 'most powerful models' running in the cloud is overwhelmingly high
However, there is definitely a need to control them within one's own environment or on devices
OpenAI is in a position where it should do both
And, as a long-term goal,
"someday we want to provide a GPT-5 class model as open-source weights that can run on a single device"
He also mentioned this quite ambitious vision.
Behind this is 'privacy and freedom,' which Sam repeatedly emphasized.privacy and freedom
"If AI is to become a core infrastructure of life, it will be important for it to work offline and be able to function locally"
How cloud AI and local AI will divide their roles is likely to be a major theme moving forward.
6. What Executives Should Do 'Right Now': Preparing Data Infrastructure and Definitions
Finally, in response to the question, 'So what should executives do starting tomorrow?', Ali provides some extremely practical advice.
-
Preparation of data infrastructure
Integrate data scattered across the company into a state that agents can access
Make data sleeping on-premises and siloed databases reusable as a 'source of context'
-
Clarification of definitions (metrics)
Clearly state definitions such as 'What is churn?', 'How do we define sales?', and 'Fiscal year/quarter cutoffs' in a form that the model can reference
"Those definitions are definitely written somewhere in some document in some department. Whether or not you can connect them in a way that AI can read will determine success or failure."
Sam also adds,
"The models are already smart enough. What is needed is to teach them 'where to look'."
he supplements.
Conclusion: From a Competition of Models to a Competition of 'Integration' and 'Freedom'
The partnership between OpenAI and Databricks is not just about saying, 'You can now use powerful models.'
The trinity of enterprise data, AI models, and governance
Transitioning to agents that handle long-term tasks
The vision of GPT-OSS spanning cloud and local environments
AI utilization based on governance, privacy, and freedom
In the coming years, the challenge companies will face is less about 'which model to choose' and more about'how to integrate them, how much to delegate, and with what degree of freedom to operate them.' This is the real competition.
It can be said that this discussion clearly illustrates that future vision.