The Data Foundation Behind Agentic AI
The conversation about artificial intelligence often centers on models. But in an interview with Constellation Research’s Hoger Muller, Couchbase’s Tim Finley makes a case for looking beneath the model: a company’s data is what gives AI useful context—and what can set one company’s AI apart from another’s.
Better context can mean lower AI costs
Organizations may use the same AI models, Finley notes, but they bring different data: the knowledge of their customers, operations, and business. Getting that data and context right can help AI systems produce more relevant results. It may also reduce costs by avoiding unnecessary model calls.
That cost pressure grows as AI systems move from occasional prompts to always-on agents. During the interview, Finley cited a Cloudflare analysis reporting that autonomous agents and machines account for more than 60% of web traffic—18 months ahead of schedule—and a Cisco study that found agents generated 450% more network traffic than humans performing the same tasks. These figures were cited in the conversation; they underscore the infrastructure demands organizations should consider as agentic workloads scale.
A converged platform for AI workloads
Finley described Couchbase’s approach as a converged data platform, combining a scalable JSON database with SQL capabilities. He said the company’s recent AI investments include vector search, model services, LLM functions, semantic caching, and agent memory.
One example is Couchbase’s integration with Amazon Bedrock. According to Finley, semantic caching reduced LLM calls by approximately 40% to 60% in Couchbase’s testing. Fewer calls can help lower costs and latency while preserving access to capable models.
AI data planes and MCP tools
Couchbase launched its AI Data Plane in June, Finley said, bringing together capabilities such as model services, agent memory, and an agent catalog. He also described an early customer that put multiple agents into production within weeks of general availability.
Finley said customers have shown strong interest in Couchbase’s Model Context Protocol (MCP) offering. The company is evaluating roughly 110 tools and skills developed by field engineers for potential future releases. The appeal, he suggested, is not limited to agents: tools that make data capabilities easier for agents to use can also be useful to the people working with them.
Governance that scales with agents
More agents mean more activity to monitor: tool calls, handoffs between agents, model access, and token usage. Finley said Couchbase can capture this “digital exhaust” in JSON, where teams can query it with SQL to investigate system behavior, including what is succeeding and failing.
He also described a longer-term goal: full provenance at scale, allowing organizations to determine what an agent knew, what it did, and what resulted from each action. Finley said Couchbase aims to deliver this within the next calendar year.
The interview’s broader message is that governance and innovation need not be opposing priorities. Governance is essential, but treating it as a reason to delay building agents may leave organizations behind. A practical path is to build with governance in mind and choose infrastructure designed to support visibility and control as workloads grow.
The takeaway: Agentic AI makes the data foundation more—not less—important. Relevant context, efficient infrastructure, and traceable activity can help organizations scale AI while keeping its costs and behavior in view.