Cornerstone Cuts Database Diagnosis Time by 78% With 13 AI Agents on AWS

On October 7, 2026, AWS published a case study describing Orion AI, a database operations system built by Cornerstone OnDemand with Amazon Bedrock and Strands Agents. The company reports that average diagnosis time fell from about 45 minutes to 10 minutes, a 78 percent reduction. The architecture uses multiple specialized agents rather than assigning every operational task to one general-purpose model.

THE OPERATIONAL PROBLEM: Cornerstone's Enterprise DataOps team previously investigated SQL Server incidents by checking system views and logs across tools. Database lifecycle processes involved more than ten manual steps, reports between SRE and data teams lagged by about 15 minutes, and redundant alerts made urgent signals harder to spot.

REPORTED OUTCOMES: AWS describes a 78 percent reduction in diagnosis time, a 70 percent reduction in manual lifecycle steps, a median 65 percent reduction in redundant alerts, and the removal of a 15-minute reporting delay. These are results reported for Cornerstone's environment, not guaranteed outcomes for other organizations.

THIRTEEN DOMAIN-SPECIFIC AGENTS: Orion AI combines a coordinating meta-orchestrator with 13 specialist agents covering database diagnostics, session blocking, live SQL state, infrastructure monitoring, operational analytics, knowledge retrieval and related tasks. Narrow tool access keeps each agent focused on its domain.

THE ROLE OF THE ORCHESTRATOR: The central agent interprets a request, selects specialists and assembles their findings. According to AWS, it holds routing and control-flow tools rather than the operational tools belonging to the specialists. This makes responsibilities and tool access easier to separate.

KEYWORD-FIRST ROUTING: Roughly 80 percent of requests take an in-memory keyword route that resolves in under a millisecond, according to the case study. Ambiguous requests fall back to semantic search using Amazon Titan Text Embeddings V2. This hybrid design reserves more flexible retrieval for questions that need it.

THREE STAGES OF SQL DIAGNOSIS: One agent inspects wait types, long-running queries and blocking chains. Another investigates root blockers and potential remediation. A third queries live SQL Server state directly. Separating observation, root-cause reasoning and current-state checks helps avoid treating old context as live evidence.

CONNECTING EXISTING TOOLS WITH MCP: Shared capabilities such as SQL diagnostics use Model Context Protocol connections to external tool servers. Other sources are accessed through SDKs or REST APIs. Orion AI does not force every integration through the same abstraction; it chooses shared interfaces where reuse matters.

MEMORY IS NOT LIVE TELEMETRY: DynamoDB holds same-session context, AgentCore Memory supports cross-session recall, and an ephemeral scratchpad shares findings between agents. Requests about current operational metrics bypass memory and query live systems instead. This is an important safeguard against stale diagnostic conclusions.

LIMITING MEMORY RETRIEVAL: AWS describes parallel retrieval across memory tiers with a 4,000-token budget and a 500-millisecond timeout. More retained context is not automatically better. Relevance, freshness and latency must be controlled.

HUMAN APPROVAL FOR DISRUPTIVE ACTIONS: Destructive operations require explicit confirmation within a five-minute window and default to denial on timeout. Orion AI also uses input validation, SQL-injection protections, output redaction, role-based permissions and per-request cost tracking.

OBSERVABILITY IN PRODUCTION: Orion AI runs as containerized services on Amazon ECS. CloudWatch collects operational logs and metrics, while AWS X-Ray traces calls across agent executions. Operators need visibility into routing, tool usage, latency and failures, not just the final AI response.

THREE PEOPLE AND SIX MONTHS: AWS says a three-person team built Orion AI in six months. That is a useful case-study detail, but implementation effort will depend on existing infrastructure, APIs and organizational support. Ongoing maintenance should be included in any business assessment.

DESIGN LESSONS FOR OTHER TEAMS: The reusable principles are domain-scoped agents, keyword routing before semantic fallback, and live-system checks for real-time questions. A cautious rollout can begin with read-only diagnostics, measure accuracy and investigation time, then expand permissions only after controls are validated.

WHAT THIS MEANS FOR AUTONOMOUS OPERATIONS: Orion AI demonstrates how agents can assist with investigation, evidence gathering, remediation recommendations and handoffs. Sustainable results still depend on data freshness, permissions, auditability and human oversight. Operational control matters as much as agent autonomy.