How Cornerstone OnDemand cut database diagnosis by 78% with Amazon Bedrock

Cornerstone OnDemand, Inc., which provides workforce readiness solutions and serves 140 million users across 186 countries, built a multi-agent AI system called Orion AI to move database operations from reactive work to proactive, self-orchestrating workflows. The system uses Amazon Bedrock for managed access to foundation models and Strands Agents, an open source agent orchestration framework from AWS, for the orchestration layer. A three-person team delivered it in six months.

Before Orion AI, Cornerstone's Enterprise DataOps team spent approximately 45 minutes per database incident querying system views and cross-referencing logs before handing work across teams. With Orion AI, database diagnosis dropped from 45 minutes to 10, a 78% reduction. Manual database lifecycle steps fell from 10 or more to a single interaction, a 70% reduction. Reporting lag between the site reliability engineering team and the data team went from 15 minutes to instantaneous, and redundant alerts fell by a median of 65%, with only 3 to 4 of every 10 previous alerts reaching engineers.

Orion AI is a hub-and-spoke system in which a meta-orchestrator delegates to specialized child agents. Engineers interact through a web application. It uses 13 domain-specific agents covering areas including infrastructure monitoring, database diagnostics, database lifecycle operations, customer analytics, and knowledge retrieval from runbooks. Three SQL Server agents handle distinct investigation stages: a database diagnostics agent, a session blocking analysis agent, and a real-time SQL diagnostics agent.

Routing is keyword-first with a semantic search fallback; approximately 80% of queries resolve through an in-memory keyword fast-path in under a millisecond, with fallback to semantic search powered by Amazon Titan Text Embeddings V2. Orion AI is deployed as containerized services on Amazon Elastic Container Service, reaching data sources through Model Context Protocol servers and direct SDK or REST API calls. Amazon Bedrock AgentCore provides a cross-session memory layer with short-term, long-term, and inter-agent tiers, while live-metric requests bypass memory. Guardrails include prompt-level constraints, a human-in-the-loop confirmation gate with a five-minute window that defaults to deny on timeout, custom input validation and output sanitization, and routing-level protection. Amazon CloudWatch and AWS X-Ray provide observability.