Why your AI pilot stalled at the context gap

Everyone has high hopes for the value they can derive from their agentic AI projects. The dream is to get them into production where they can assist users and make autonomous decisions that drive the business forward.

But many projects get stuck in the pilot phase.

The numbers are stark:

• Only 16% of companies have deployed agentic AI at an enterprise scale.
• Nearly 80% of companies in one survey said their AI projects are constrained by data access issues
• 70% of those asked by Deloitte said they don’t feel they can adequately trust and govern agents

This results in an all-too-familiar situation where AI agents either guess about data or make it up completely. An agent that finds multiple conflicting definitions of “revenue” across data sources might arbitrarily pick one.

The problem occurs when AI agents are deprived of the governed context they need to make informed decisions. Let’s look at this context problem, why ungoverned agents fail, and how the dbt platform enables you to build scalable AI agents that everyone in your company can trust.

The four ways an ungoverned agent fails

Without trusted, governed data, an agentic AI solution that works under test conditions often fails when faced with real user questions. There are four common reasons why:

It writes unreliable SQL. This isn’t often an outright syntactic failure; usually, the SQL parses and runs. The problem is that it’s selecting the wrong values.

It invents or misreads metric definitions. Without sufficient context, an agent might use stale, missing, or fabricated metrics. This problem is exacerbated by a lack of reviews.

It has no guardrails and no audit trail. No auxiliary processes check the agent’s work, and there’s no record you can check to verify how it reached its conclusions.

It creates rising compute costs due to inefficient work. Left to their own devices, AI agents may use more compute than necessary, causing your data processing costs to spike. They’re getting the work done, but processing is eating your profits.

None of these are “sometimes agents make a mistake” issues. These are predictable and, fortunately, fixable problems. You just need to take the correct approach to data. For example:

• SQL generation can be validated through rigorous testing. You can also train your models and agents to learn how your data works.
• Revenue can be centrally defined and shared across teams using a semantic layer, instead of spread across dozens of data stores.

Machine-readable governance

In the past, we relied on humans to manually review data and ensure its accuracy. We’re producing too much data for that to be a scalable approach in the AI age.

To make agentic AI truly scalable, you need governance. But not the type where everything is written down in a large document no one reads. You need computational and machine-readable governance.

In a computational model, governance is enforced using several capabilities:

• Contracts. A contract is a machine-readable description of how the data is shaped, how it functions, and how it differs between releases, along with the endpoints used to access it. Data that doesn’t meet a contract fails to ship, keeping a class of errors out of production. The contract also enables agents to discover and use the data, particularly as its shape evolves over time.
• Tests. Data needs to be tested the same way we test software. Tests that ensure correct data can be run when shipping new data transformation changes, and run periodically in production to ensure ongoing data health.
• Semantic layer. A semantic layer provides one central location for all metric definitions. This eliminates the agent from guessing what “revenue” means. Each metric provides additional metadata that’s invaluable to AI agents: where it came from, how it was derived, and its business purpose.

Without this machine-readable approach to governance, you can’t guarantee that your AI agents will return accurate answers at scale.

The shift to agent consumption of data

Historically, the primary consumers of data have been humans. It’s quickly becoming AI agents, which we humans now rely on to help distill the vast amounts of information we keep generating.

Fivetran and dbt Labs realized that, together, we could do more to advance a new era of trusted, Open Data Infrastructure for AI at scale.

Your AI agents don’t need to languish in the prototype phase. With the dbt platform, you have the tools you need to create high-quality, governed, and trusted data that enables your agents to make accurate decisions.

To learn how, watch the full demo of the dbt platform in action.