Unlike humans who debate flawed data in meetings (a "data brawl"), AI agents confidently present a single, incorrect number from bad data. This creates a "silent failure" where the error is persuasive and unnoticed, a phenomenon Salesforce's Gaurav Pathak calls "garbage in, gospel out."
When enterprise AI agents like support chatbots fail due to bad data, they create new problems such as increased escalations. The human employees who must clean up these messes and deal with the consequences are termed "sin eaters," absorbing the cost of the AI's failures.
For enterprise applications, the choice of AI model is a minor factor. Salesforce's Gaurav Pathak argues that 95% of the battle is getting the right business data—the context—to the agent. This reframes AI investment from a focus on cutting-edge models to a focus on data infrastructure and management.
Beyond model building, AI engineers need three critical skills: 1) implementing evaluations and traces to create self-improving systems, 2) ensuring the right data context reaches the model, and 3) managing token economics to optimize for both cost and performance.
An enterprise's data landscape is like a supermarket of unlabeled cans. Without metadata, an AI agent must "open and smell" every data asset to find what it needs, burning massive amounts of tokens. Well-structured metadata acts as the can's label, allowing agents to find the right data efficiently and affordably.
Previously, pitching investments in data infrastructure was difficult. Now, as companies deploy AI agents like support bots, the negative consequences of bad data—such as increased escalations to human agents—are painfully clear. This direct business impact makes the return on investment for data management much easier to justify.