Elastic reports agentic AI pipeline lifted alert triage accuracy from 60% to 92%

What Elastic reported

Elastic's internal InfoSec team says it raised the accuracy of security alert triage from 60% to 92% by replacing simple AI summaries with an agentic pipeline built on its own tooling. In a post published on dev.to, the team describes how its Security Operations Center moved from having a language model annotate raw alerts to a system in which specialised agents investigate alerts with the same background material a human analyst would want.

The number stands out because agentic AI in security operations is usually discussed in aspirational terms, and this is a concrete before-and-after measurement from a live SOC. One caveat worth stating up front: it is Elastic reporting on Elastic products and Elastic's own team, so it is a self-reported figure rather than an independent evaluation.

Three agents in a synchronous pipeline

The pipeline, assembled with Elastic Workflows and Agent Builder, divides the work among three agents:

• Pattern Finder reviews historical cases and prior analyst feedback to surface recurring trends before an investigation begins.
• L1 Investigation runs a focused set of external lookups against the alert and works toward a verdict.
• Summarizer packages the outcome into a report formatted for Slack and Kibana.

The three run synchronously rather than as free-running autonomous agents. According to the post, that design choice keeps token costs down and shortens the time between an alert arriving and a finished report being available.

Context, not just model output

The team attributes most of the accuracy gain to grounding rather than to model choice. The agents receive past case records, the team's internal investigation guides, and user risk data held in Workday — the kind of institutional knowledge that lives outside the alert itself and that a generic summariser never sees.

Without that grounding, the team describes AI-written summaries as untrusted material — what it bluntly calls "slop" — that analysts cannot verify. With it, the output becomes intelligence analysts can check and act on. That distinction is the difference between an AI feature people tolerate and one they actually rely on during an incident.

Results delivered inside existing workflows

The reports land where analysts already work. Summaries appear in Slack and Kibana, and because the pipeline's reasoning is transparent, the team says analysts can often resolve a case with one click rather than re-doing the investigation themselves.

Every analyst decision also flows back into the system. The post describes this feedback loop as central to the design, since it exposes where the agents were wrong and gives the team a way to keep improving the models and the workflow over time, rather than treating deployment as the finish line.

Why it matters

Two things make this more than a routine vendor case study. First, it is a quantified, real-world data point for agentic AI in security operations — a domain where the gap between conference demos and production reliability is notoriously wide. Second, it suggests where the gains actually come from: not from a smarter base model, but from context engineering. Historical cases, written investigation guides and HR-adjacent risk data turned the same class of model from a summariser into something closer to an investigator.

For security teams weighing agentic tooling, the practical lesson is that the hard part is plumbing — connecting agents to the organisation's own knowledge — not prompt tuning. And for readers evaluating the claim, the healthy stance is to treat 92% as Elastic's experience with its own stack: encouraging, specific, and awaiting replication by teams without a vendor's incentive to publish flattering numbers.