AI Agents Need a New Foundation

The hype around AI agents is palpable, but a stark reality is emerging: many organizations are building advanced AI capabilities on a foundation not designed for action. Databricks co-founder Arsalan Tavakoli-Shiraji highlights that the gap between AI activity and tangible value often stems from architectural shortcomings. Enterprises are struggling to move beyond experimentation and task automation because their existing data infrastructure and governance models are ill-equipped for agentic execution.

Tavakoli-Shiraji points out that the most common pitfall is underestimating the complexity beneath the AI models themselves. The true challenge lies in unifying siloed data, implementing robust governance that understands agent behavior, and imparting deep semantic understanding of the business context to these virtual workers. The anti-pattern involves data locked in disparate systems, governance treated as an afterthought, and a subsequent scramble to understand why agents fail in production.

Why Old Architectures Fail Agents

Traditional analytics architectures, built for dashboards and batch pipelines, are fundamentally misaligned with the demands of AI agents. Dashboards, often static and difficult to interrogate, can't provide the real-time, drill-down capabilities needed when agents require immediate context. Batch processing, designed for slower decision cycles, simply cannot keep pace with the shrinking window between observation and action required by agentic systems.

This architectural mismatch is a critical bottleneck. Agents need to interact with data at speeds and scales that legacy systems cannot support. The infrastructure must shift from serving human analysts to powering automated, low-latency applications. This is where a new approach, like Databricks' Lakebase, becomes essential. It offers a transactional database designed specifically for the agentic world, working alongside existing analytics layers without replacing them.

Governance Beyond Outputs

As AI agents transition from generating reports to taking actions—sending emails, updating records, executing decisions—governance failures become significantly more impactful. The assumption that agents can inherit human permissions is flawed; humans possess contextual awareness and gut instincts that agents lack. Agents operate on goals and constraints, and without appropriate oversight, this can lead to unintended consequences.

Organizations must treat governance not as a post-deployment fix but as a prerequisite. This involves understanding agent permissions, tracking their actions, and ensuring they operate within defined boundaries. Robust governance is the lynchpin for unlocking real value from AI agents at scale.

The Path to Value Without Sprawl

The fastest route to successful AI agent deployment, avoiding technological sprawl, involves two key steps. First, clearly define what success looks like before beginning. Pinpointing specific outcomes—productivity gains, new revenue streams, cost reductions—is crucial for backward-engineering the right technical approach. Technology itself is not the objective.

Second, isolation is key. Creating a dedicated, focused pilot team, shielded from legacy technical debt and existing policy constraints, allows for rapid iteration and learning. This agile approach helps define what "good" looks like in a real-world context before scaling to the broader organization. This allows teams to Databricks AI Agents effectively.

The uncomfortable truth for leaders is that the agentic era demands a holistic re-evaluation of underlying processes, not just bolting AI onto existing workflows. Just as replacing steam engines with electric ones without redesigning factory floors yielded minimal gains, organizations must ask: if we were building this from scratch with AI in mind, what would it look like? This requires change management, enablement, and a clear vision of success, tackling all aspects simultaneously.