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

Cornerstone OnDemand has turned one of its messiest back-office jobs into an AI workflow. The company’s new system, Orion AI, uses Amazon Bedrock and AWS’s open source Strands Agents framework to coordinate a set of specialist agents that handle database operations, alerts, and support tasks.

The payoff is pretty concrete. What used to take the Enterprise DataOps team up to 45 minutes per database incident now takes 10. That’s a 78% drop in diagnosis time. A three-person team built the system in six months, which is not bad for something that had to sit between live operations and multiple internal teams without making a bigger mess.

Before Orion AI, Cornerstone’s engineers were doing a lot of expensive human glue work. They were hopping between tools, checking system views, comparing logs, chasing down long-running queries, and handing problems off as they pieced together a root cause. Database lifecycle work could take 10 or more manual steps, and reporting between SRE and data teams had a 15-minute delay. Redundant alerts just added static.

Orion AI’s trick is to split the work up. A coordinating agent routes requests to specialized agents in a hub-and-spoke setup, while engineers use a web app to ask questions and approve actions. The system covers infrastructure monitoring, database diagnostics, lifecycle operations, customer analytics, and knowledge-based support. On the routing side, it goes keyword-first and falls back to semantic search when needed. About 80% of queries hit the fast path in under a millisecond.

The improvements go beyond diagnosis speed. Cornerstone says manual lifecycle steps dropped from more than 10 to one interaction, the SRE-to-data-team lag became instantaneous, and redundant alerts fell by a median of 65%. Orion AI can also create a populated Jira ticket for the right on-call engineer, which is a nice way of saying it does the follow-through humans usually get stuck with.

The design choices are the interesting part. Cornerstone split agents by domain rather than by task complexity, scoped memory so live metrics always come from current state, and built in confirmation gates for destructive actions. It sounds less flashy than “AI transformation,” but that’s usually how useful systems show up: boring architecture, fewer alerts, fewer handoffs, less drama.