AI governance has become a board-level concern as enterprises scale up AI experimentation. The question for executives is no longer whether governance matters, but how to design it so that it enables speed, innovation, and trust at the same time. In an interview published on the Databricks blog, David Meyer, senior vice president of product at Databricks, offered a notably practical view of where organizations make real progress, where they get stuck, and how today’s governance decisions shape tomorrow’s options. Rather than treating AI governance as new or abstract, his argument returns to first principles: engineering discipline, visibility, and accountability.
AI governance as a way to move faster
One of the clearest patterns Meyer describes is that governance challenges are both organizational and technical, and tightly linked. On the organizational side, leaders want teams to move faster without creating chaos. Organizations that struggle tend to respond by avoiding risk: they centralize every decision, add burdensome approval processes, and inadvertently slow everything down — often ending up in a worse position, not a safer one.
Strong technical governance, by contrast, can unlock organizational resilience. When leaders have real visibility into what data, models, and agents are in use, they do not need to control every decision manually. Teams get more freedom because leadership understands what is happening across the system. In practice, that means teams do not ask permission for every model or use case — access, auditing, and updates are controlled centrally, and governance operates by design rather than by exception.
Between two extremes
Meyer describes two failure modes. Some companies declare themselves “AI-first” and encourage everyone to build freely. It works for a while — people move fast, there is excitement — until the organization suddenly has thousands of agents, no real inventory, no idea of costs, and no clear picture of what is running in production. Other organizations try to control everything up front, keep approval choke points in place, and find that almost nothing worthwhile ever ships, while teams feel constant pressure of falling behind.
The companies getting it right sit in the middle. Within each business function they identify AI-literate people who can guide experimentation locally. Those people compare notes across the organization, share what works, and narrow the recommended toolset — and going from dozens of tools to two or three makes a bigger difference than most expect. The organizational reasons agent initiatives lose momentum are explored further in this analysis of why AI agent projects stall.
Agents are not as new as they seem
A striking theme of the conversation is that agents are less novel than they appear. Their key traits are familiar: they spend money continuously, expand the security surface area, and connect to other systems — all problems enterprises have handled before. The same principles used to govern data assets and APIs apply. If no one knows where an agent exists, no one can turn it off; if an agent touches sensitive data, someone must be accountable for it. Organizations that assume agent systems need an entirely new rulebook are, in Meyer’s framing, mostly wrong — borrowing proven lifecycle and governance practices from data management covers most of the ground.
Start with observability
Asked where an executive should begin, Meyer’s answer is observability. Meaningful AI almost always relies on proprietary data, so organizations need to know what data is being used, which models are involved, and how those pieces combine to form an agent. Many companies use multiple model providers across different clouds; when those models are managed separately, cost, quality, and performance become hard to understand. Managing data and models together lets teams test, compare, and improve more effectively — and matters more as the ecosystem changes, because leaders need to evaluate new models without rebuilding the stack each time.
On where progress is fastest: knowledge-based agents stand out — pointed at a document set, they immediately let people ask questions and get answers. The catch is that many such systems degrade over time as content changes, indices go stale, and quality declines. Most teams do not plan for this. Sustaining value requires a system that continuously refreshes data, evaluates outputs, and improves accuracy; without it, organizations often see strong activity for a few months followed by declining usage and impact.
Treating agentic AI as an engineering discipline
Organizations balancing speed with trust treat agentic AI as an engineering problem, applying the discipline used for software: continuous testing, monitoring, and deployment. Failures are expected; the goal is not to prevent every problem but to limit the blast radius and recover quickly. Teams that operate this way move faster and with more confidence — and, in Meyer’s telling, a system where nothing ever goes wrong is probably being run too conservatively.
Trust follows the same logic. It does not come from assuming the system will be perfect; it comes from knowing what happened when something went wrong. That requires traceability — what data was used, which models were involved, who interacted with the system. With that level of auditability, an organization can afford to experiment more. Large-scale distributed systems have always been run this way: optimized for recovery, not the absence of failure. As AI systems grow more autonomous, that mindset becomes more important.
Building an AI governance strategy
Rather than treating agentic AI as a break from the past, the interview frames it as an extension of what enterprises already know how to operate. Three themes stand out:
- Use governance to enable momentum, not impede it. Foundational controls let teams move quickly without losing visibility or accountability.
- Apply familiar engineering and data practices to agents. Inventory, lifecycle management, and traceability matter as much for agents as for data and APIs.
- Treat AI as a production system, not a one-time launch. Continued value depends on ongoing evaluation, fresh data, and fast detection and correction of problems.
Limitations and what to watch
- The perspective comes from a senior Databricks executive, and the recommendations naturally align with the capabilities of unified data-and-AI platforms; the underlying principles (observability, inventory, lifecycle discipline) are vendor-neutral, but tooling choices deserve independent evaluation.
- Remarks are paraphrased from the published interview; exact wording is in the linked original.
- Governance standards for autonomous agents are still immature industry-wide, and practices described as sufficient today may need strengthening as agent autonomy increases. Broader vendor-landscape context is available in this overview of integrated AI governance platforms.
The overall conclusion is clear: sustainable AI value comes neither from chasing the newest tools nor from shutting everything down, but from building a foundation that lets an organization learn, adapt, and move forward with confidence.
