AI agents are tackling the 'boring' work of data engineering
MotherDuck CEO Jordan Tigani argues that while agents can automate mechanical data maintenance, they cannot replace human business logic.
AI agents are finally beginning to penetrate the data engineering and analytics sector, though their arrival has been slower than in other software domains. This shift marks a transition from manual data curation to a model where agents handle the mechanical overhead of data maintenance.
According to Jordan Tigani, CEO of MotherDuck, the delay in adoption is largely technical: large language models (LLMs) have only been reliably proficient at writing SQL for the last six to nine months. Now that this threshold has been crossed, agents are proving particularly effective at "boring" but essential tasks. Tigani notes that agents excel at mechanical work, such as patching renamed tables, inferring column values or table joins, and converting code assumptions into functional tests.
The Limits of Automation
Despite these gains, there is a hard ceiling on what agents can achieve. While they can navigate the technical structure of a warehouse, they cannot infer human-made business decisions. For example, an agent cannot independently determine the specific organizational definition of "revenue" or "customer," as these are strategic choices rather than objective warehouse facts.
Furthermore, the dream of "hands-free analytics"—where agents proactively provide insights—remains a research prototype. Tigani suggests these proactive systems are unlikely to succeed in the near term because the bar for relevance is too high and the cost of false positives is too steep for most enterprises.
The Infrastructure Gap
This transition is forcing a fundamental rethink of database architecture. Traditional engines were designed for "human patience," where a few seconds of latency is acceptable. However, agentic workflows operate on entirely different scales, executing chains of queries in milliseconds.
Tigani warns that latency acts as a compounding bottleneck; a 10x difference in engine response time can result in 10x less work completed per minute when queries are sequential. To support this, data teams must move toward infrastructure that supports rapid scaling and strict tenant isolation to manage sudden bursts of parallel queries. As Tigani puts it, "Engines tuned for human patience and engines tuned for agent throughput are not the same engines."
What to Watch
As companies move toward agentic data usage, the competitive advantage will shift to those who implement rigorous evaluation frameworks (evals) and optimize for throughput. The immediate focus for data teams will be preparing their infrastructure for high-concurrency demands while accepting that the core business logic will remain a human prerogative.