How to Govern AI Spend at Scale

Databricks governs AI spend at scale through a data-driven operating model. Unity Gateway, Databricks' enterprise AI control plane, provides centralized visibility, budget controls, and contextual policy enforcement across models and agents. Omnigent, Databricks' open-source meta-harness, routes tasks to the appropriate coding-agent harness and lets teams apply spend policies to individual agents and sub-agents

Unity Gateway helps lower AI costs while preserving developer productivity through three principles: see it, show it, act on it.

• Centralized visibility: Connect any model, agent, or MCP server to the gateway, regardless of provider. Track usage, performance, and cost in a unified dashboard.
• Real-time budgets and spend caps: Set per-user or per-team budgets that update as tokens are consumed. Engineers can view their own spend through the Unity Gateway CLI and build a clearer understanding of their consumption.
• Smart model routing: Match task complexity to the right model and harness automatically. Many requests can be handled without the most expensive model.
• Unified AI observability and tracing: Log every request with full context in governed Unity Catalog tables, including who made the request, which model handled it, how many tokens it used, and what it cost.
• Unity Catalog governance: Govern AI models, providers, agents, and skills as objects with fine-grained permissions. This extends the same framework your data team already uses.
• Contextual policies: Apply dynamic, stateful guardrails that adapt to events during an agent session. Policies can respond to session context alongside static allow and deny rules.

Watch the webinar to see these capabilities working together in a live demo. The demo includes how Databricks cut its own coding-agent costs by approximately 50% without compromising quality.

What You'll Learn in This Webinar

• How Databricks CTO Matei Zaharia and Director of Product Kasey Uhlenhuth govern AI spend across more than 10,000 employees, and how Unity Gateway brings the same approach to your organization.
• Why blunt spending caps and blanket policies backfire, and how to design more effective controls.
• The “see it, show it, act on it” framework for data-driven AI governance.
• How Databricks benchmarked real coding tasks to identify where agents waste the most money, and what the results revealed.
• What tokenmaxxing is, why it is widespread, and how to shift toward valuemaxxing.
• How contextual policies adapt enforcement to session data, helping prevent data exfiltration while allowing legitimate work.
• How Omnigent orchestrates multiple coding agents and automatically routes tasks to the right model.

Who This Is For

• CIOs and IT leaders managing growing AI portfolios who need a unified view of spend, usage, and risk across models and providers.
• FinOps teams extending cloud cost discipline to AI workloads, where token-based pricing and autonomous agents challenge traditional tools.
• Platform and engineering teams operating model gateways who need to provide governed AI access without slowing developers down.
• Security and compliance leaders who need audit trails, data-aware access policies, and guardrails that adapt to what an agent has seen during a session.

What is Unity Gateway?

Unity Gateway is Databricks’ enterprise AI control plane. It brings FinOps, observability, permissions, routing, policy enforcement and identity management into a single layer that connects to any model, agent, MCP server or skill.

Unity Gateway is generally available and already used by thousands of customers, with more than a quadrillion tokens processed through the gateway in the past year.

What is Omnigent?

Omnigent is Databricks’ open-source meta-harness for coding agents. It provides a common layer across existing agent harnesses, so teams can combine and switch between them for different tasks instead of being locked into a single option.

Omnigent helps teams govern agentic work in two ways:

• Smart routing: Direct simple requests to more cost-efficient models and complex tasks to more capable models or harnesses.
• Spend policies: Set limits for individual agents and sub-agents, with notifications when they reach a defined spending threshold.

5 Lessons for Managing AI Costs

• Start with visibility: You cannot manage what you cannot see. Centralize usage and cost instrumentation before introducing restrictions.
• Put spend visibility in users’ hands: Showing spend to the people generating it can change behavior faster than imposing top-down caps. At Databricks, real-time visibility in the tools engineers already use proved more effective than relying on a monthly limit alone.
• Benchmark real work: Before assuming that a larger model or higher-effort setting will deliver more value, test it against your own tasks. Databricks’ internal benchmarks found that a simpler, lower-cost harness matched or exceeded the default setup’s accuracy.
• Match controls to risk: Blunt controls can create friction without addressing the underlying risk. Contextual, session-aware policies and per-task budgets can distinguish legitimate needs from problematic behavior more effectively than flat spending caps or blanket action bans.
• Route based on task complexity: Smart routing between cost-efficient and more capable models or harnesses can reduce spend without slowing teams down. The key is to match routing to task complexity instead of applying a one-size-fits-all default.

Recommended Assets

• Unity Gateway: Lower AI costs without losing productivity | Databricks 
• Introducing Omnigent: A Meta-Harness to Combine, Control and Share Your Agents | Databricks Blog) 
• Managing AI Coding Costs at Scale | Databricks Blog