Tool Sprawl Meets Agentic AI: Why SOCs Are Rethinking the Triage Stack
Forty tools, forty-three day dwell times. Vendors are pitching agentic AI as the fix. Analysts have questions.
The average enterprise security team runs north of 40 tools. EDR, SIEM, SOAR, CSPM, NDR, identity analytics, the long tail of point products bolted on after the last incident. Each one generates telemetry. Most of it overlaps.
And yet dwell times sit around 43 days.
That number is the uncomfortable subtext to every vendor pitch landing in analyst inboxes right now. The pitch goes like this: assistive AI — copilots, summarizers, natural-language query — was the warm-up. Agentic AI, where models plan and execute multi-step investigations on their own, is what actually moves the needle on mean time to respond.
It's worth separating capability from intent here.
Capability-wise, the agentic shift is real. Models can now chain tool calls, pivot across data sources, write and run queries, and produce a triage verdict without an analyst clicking through five consoles. Several detection-and-response vendors have already shipped agent frameworks that operate against their own telemetry, and a handful are exposing MCP-style interfaces so the agents can reach into third-party tools.
Intent — meaning whether SOCs actually want autonomous action on production — is messier.
The operational case is straightforward. Tier-1 triage is repetitive, well-bounded, and a known burnout driver. If an agent can close out the 70% of alerts that are obvious false positives, analysts get their time back for hunting and adversary tracking. That's the bull case.
The bear case is familiar to anyone who has watched a SOAR playbook misfire at 3 a.m. Autonomous agents inherit the blast radius of whatever credentials they hold. An LLM-driven response action that disables the wrong account, isolates the wrong host, or pushes a bad block list is an incident in itself. Prompt injection via attacker-controlled log data is not theoretical; researchers have already demonstrated indirect injection against AI-assisted analyst workflows.
A few questions worth asking any vendor selling agentic SOC tooling:
- What's the human-in-the-loop boundary, and is it configurable per action class?
- How is the agent authenticated to downstream tools, and what scopes does it hold?
- Is there an audit trail that a forensic investigator could actually reconstruct?
- What happens when the model is wrong, and who owns the post-incident review?
The industry data on dwell time hasn't materially improved in years despite enormous tooling investment. That's a process problem as much as a technology one. Agentic AI may compress triage. It will not, on its own, fix alert fatigue rooted in noisy detections, undertuned rules, and overlapping coverage from too many overlapping products.
The shift is coming regardless. The interesting question is which SOCs treat agents as a junior analyst with a leash, and which hand over the keys.
