Connecting enterprise data to the age of AI agents

For the past several years, the enterprise AI conversation has focused on models: which one reasons best, which one handles the largest context window, and which new release moves closest to human-level performance. That focus made sense while model capability was the clearest constraint. But as reasoning improves and models become commoditized, the decisive question is changing. It is no longer simply, “What can the model do?” It is, “What can the model understand about my business — and what can I trust it to do with that understanding?”

In a recent conversation, George Fraser, CEO of Fivetran and dbt Labs, and Andi Gutmans, Vice President and General Manager of Google Agentic Data Cloud, explored what this shift means for enterprise leaders. Their conclusion was clear: practical commercial usage of AI will be defined less by access to increasingly capable models and more by access to reliable data, relevant context, scalable infrastructure, and appropriate controls. As Gutmans put it, “the data is really the lifeline of AI.”

[CTA_MODULE]

The AI bottleneck has moved to data

Most enterprises do not lack data. They lack a dependable way to make the right data available, understandable, and usable at the moment an AI system needs it. Critical information remains fragmented across applications, databases, clouds, documents, and operational systems. Definitions vary across teams. Multiple tables may claim to represent the same customer, product, or transaction. Even when the data is technically accessible, an agent may not know which source is authoritative or how a metric relates to the business.

That makes enterprise AI readiness a different challenge from model readiness. A highly capable model connected to incomplete, stale, or poorly defined data will still produce unreliable results. Better reasoning cannot compensate for missing business context. As Gutmans observed, “the bottleneck has actually moved to the data.”

Solving that bottleneck requires more than moving larger volumes of information into a model’s context window. In fact, more context can make an AI system less accurate. Fraser described Fivetran’s experience using a semantic model to help an internal agent understand its data. The model provided a map of the business — including important entities, relationships, and definitions — that the agent could use to construct queries. Adding more information did not necessarily improve the result. Sometimes it sent the agent in the wrong direction.

This is an important lesson for leaders racing to connect AI to everything. Context quality matters more than context quantity. The goal is not to expose every available table, document, and definition at once. It is to give an agent the smallest relevant set of trusted information needed to complete a task accurately. That approach can improve results while reducing latency and token costs.

Bring AI to the data — not data to another silo

Preparing data for AI does not necessarily require building a separate AI data stack. Early approaches often copied enterprise data into specialized vector databases so models could search unstructured content. But that separation introduces another system to synchronize, secure, govern, and maintain. As vector search becomes a native capability of databases and data platforms, structured data, unstructured data, and SQL can increasingly work together in the same environment.

That matters because business questions rarely fit neatly into one technical category. A user might ask which customers mentioned a particular issue in their call transcripts and how much revenue those customers generated last quarter. Answering that question requires both semantic search across unstructured conversations and SQL analysis of structured financial data. When both types of data are available in one platform, agents can answer the full question without extracting, loading, and reconciling results from disconnected systems.

Keeping AI close to the data can also improve accuracy and efficiency. A database can narrow the relevant dataset before running vector search, helping the agent retrieve higher-quality context while reducing unnecessary token usage. Instead of sending everything to the model, the data platform helps identify the information most likely to produce the right answer.

Open formats extend that principle across clouds and processing engines. Enterprise data rarely lives in one environment, and the AI market is changing too quickly for organizations to lock their data into a single model or platform. Open table formats such as Apache Iceberg allow multiple engines to work with the same underlying data, giving enterprises the flexibility to adopt new AI capabilities without rebuilding their existing foundation for analytics and creating another copy of their data.

An open, integrated foundation gives agents the data and context they need to reason effectively. But access to trusted data does not automatically grant an agent the authority to act on it. Once an agent can understand the business and recommend a decision, leaders must determine what it should be allowed to do next.

Autonomy is a spectrum, not a destination

The shift from copilots to autonomous agents is often described as a simple progression. First, AI answers questions. Then it recommends actions. Eventually, it acts on its own. But enterprises will not cross that boundary everywhere at once. Autonomy will develop unevenly, based on the risk, reversibility, and business impact of each decision.

An organization may be comfortable allowing an agent to draft a customer response, adjust a marketing campaign, or resolve a routine support issue. It may allow an agent to replenish $10,000 of inventory but require human approval for a $10 million purchase. The relevant question is not whether an enterprise is “autonomous.” It is where each use case belongs on the spectrum of autonomy.

That makes AI adoption as much a leadership and governance challenge as a technical one. Leaders must define which actions an agent can take, what evidence it needs, when it must escalate, and how its decisions will be reviewed. They also need a way to expand autonomy gradually. An agent might begin by observing a process, then recommending an action, then acting within a narrow threshold, and finally handling a broader class of decisions once its performance is proven.

Trust grows through controlled experience. Reliable data and semantic context enable better reasoning. Security, lineage, and auditability make that reasoning inspectable. Clear action boundaries limit the consequences when an agent is wrong. Together, those capabilities allow an organization to increase autonomy without treating trust as a leap of faith.

Connecting AI to enterprise data is the future

The model race will continue, and model capability will keep improving. But the enterprises that create lasting advantage will be those that solve the harder operational questions beneath the model: how to connect fragmented data, how to provide the right context without overwhelming the agent, how to support machine-scale workloads, and how to decide which actions should remain human.

Fivetran helps create that foundation by continuously moving data from hundreds of operational sources into platforms such as Google BigQuery and open data lakes. Fivetran preserves the fidelity of source data, dbt turns that raw data into tested and governed business models, and Google Cloud brings structured data, unstructured data, vector search, and AI together. The result is less time spent stitching systems together and more time focused on how AI can create value.

[CTA_MODULE]