Enterprise AI Is Blocked by Context, Not Model Intelligence

Ali Ghodsi’s central claim is that enterprise AI is being held back less by model intelligence than by the absence of usable organizational context. The models are already “plenty smart,” he said. The reason companies do not yet have “hundreds of agents working together” while employees manage “thousands of agents” is that agents lack the documents, data, processes, relationships, and institutional knowledge required to work across a business.

Ed Ludlow framed that as Ghodsi’s argument that “AGI is already here” and that intelligence is not the constraint. Ghodsi’s answer was direct: “It’s the context.” If companies could put “all the data, all the processes, everything that people are thinking” into AI systems, he said, the systems could help far more than they do today.

That is the rationale behind Databricks’ Genie Ontology. Ghodsi described an enterprise ontology as a “big web” of knowledge and relationships: who opened a document, who read it, what the document is about, how it connects to other information, and what the important elements inside it are. Much of that information already exists inside companies, he said. The problem is that it has not been made available to AI systems in a form they can use.

His analogy was Google search. Current agents, in his telling, often behave as if there were no search index: they visit one site, read it, follow another link, read that, and keep moving step by step. That is slow, expensive, and not very good. An ontology is meant to be the enterprise equivalent of the search index: computed in the background, then available when a user asks a question or an agent needs to act.

Ghodsi drew Databricks’ product distinction around a simple contrast.

By “recite,” he meant an assistant that searches for a document containing an answer and summarizes what it finds. That works only when the answer has already been written down somewhere. It breaks down when the question requires a live calculation, such as “Who are the new customers on the platform last 24 hours?” or “Which of our biggest customers churned in the last 10 days?” Those answers may not exist in any document. They have to be computed from company data.

Ghodsi said Genie 1’s difference is numeracy and an ability to understand enterprise data, aided by the ontology. The system is supposed to calculate from live data rather than merely retrieve a prior written answer.

Novo Nordisk was his main example. Scientists running experiments, he said, previously had to spend a long time determining how an obesity study was progressing, how a drug was affecting control groups, and what appeared in the results. With Genie 1, according to Ghodsi, those results can be computed live “in a fraction of the time.” He connected the company to Ozempic, but the example’s function was narrower: enterprise AI has to answer questions whose answers do not already exist as text.

The infrastructure argument extends beyond Genie. Ludlow pressed Ghodsi on a broader claim: systems built for human users do not necessarily work well when agents become the users of software, databases, and infrastructure.

Ghodsi began with software creation. The cost of creating software, he said, is “almost going to zero,” but every piece of software still needs a database behind it. Existing databases, in his account, were designed around human patterns of use. Agents behave differently: they move quickly, experiment, sometimes break things, and may create thousands of databases. Companies do not want to pay for thousands of databases, and agents lose much of their value if they have to wait 20 minutes for database operations.

Databricks’ Lakehouse is positioned as the response to that mismatch. Ghodsi said ordinary software needs databases behind the scenes, while more complicated data-science questions often require a separate database, known in the industry as a data warehouse. Lakehouse brings those systems together into one database that he described as “fit for agents.” In that setup, agents can operate on stored information while also asking more analytical or AI-like questions, such as what a curve appears to look like.

Prada was his example. Ghodsi said the company stores its KPIs inside Lakehouse, allowing agents to make those KPIs available quickly to company leadership. He described the result as faster work, higher accuracy, and lower cost.

The tension in Ghodsi’s argument is that he also acknowledged agents are not yet everywhere in the real world. Asked what is constraining Databricks’ growth and deployment, he did not point first to model capability or capital. He pointed to organizational approval.

Building the graph of enterprise context means handling company data, and that requires security and legal review. Databricks is trusted, he said, and has been doing this for 15 years, but security teams still need to examine the systems and legal teams still need to approve them. AI adoption for critical use cases is “working its way through” organizations as they become more comfortable with the technology.

That creates a split in current enterprise AI use. Ghodsi said most usage remains “as a chatbot,” because chatbot use has been approved in most organizations. The harder transition is using AI for critical infrastructure: financial reporting, KPI generation, or drug-development workflows where accuracy matters and the output is tied to consequential decisions.

In those areas, adoption takes longer because companies are still assessing the systems and their risks. At the same time, Ghodsi said the AI shift has accelerated Databricks’ growth and that the company is “growing faster than ever before” because of it.

Ludlow cited reports that Databricks was raising money in a new round valuing the company at up to $175 billion. Ghodsi did not confirm a fundraise. He said Databricks had not announced one, described the reports as rumors, and added that the company is “always talking to investors.”

The uses of capital, as Ghodsi described them, are tied to Databricks’ broader enterprise-AI pitch. The company is expanding into new product categories and building them “agentic-first” rather than adding AI after the fact. He named Customer Lake, a new marketing product aimed at marketing teams, and LakeWatch, a security product launched two months earlier. Entering those categories requires resources, he said, as do AI research, international expansion, and availability in more regions. AI researchers “don’t come cheap,” he added.

Ghodsi also reaffirmed that Databricks is not rushing to go public. He said he still believes Databricks will ultimately be a public company and wants to be public, but he views the current year as a bad time for an IPO. His stated reason was market congestion and uncertainty around several “mega IPOs,” one of which had already happened. He said he would prefer to go public when conditions are calmer and more predictable, and that other CEOs considering IPOs have told him they also do not want to list between major offerings.