There is a quiet failure mode in modern security operations that does not show up on any dashboard. The alerts get triaged. The mean-time-to-respond keeps dropping. The agentic AI handles tier-one volume that used to bury a team of analysts. By every metric leadership tracks, the SOC is healthier than it has ever been. And then a genuinely novel incident lands, the automation produces a confident answer that happens to be wrong, and nobody on the floor has the instinct to push back.
That is skill erosion, and Gartner has put a number on it. The firm projects that by 2030, 75 percent of SOC teams will experience erosion in foundational security analysis skills due to overdependence on automation and AI. Read that again. It is not a fringe risk affecting laggards. It is the projected default outcome for three out of four teams that adopt automation the way most teams are adopting it right now.
This is not an argument against the AI-powered SOC. The volume problem is real, the analyst burnout is real, and the gains from automation are real. The argument is narrower and more uncomfortable: the same tooling that makes a SOC faster can, if deployed carelessly, make the humans inside it worse. And degraded humans are exactly what you cannot afford during the incidents that automation was never going to catch.
What Skill Erosion Actually Looks Like
Skill erosion is not analysts forgetting how to do their jobs overnight. It is subtler and slower than that. It is the gradual replacement of investigation with approval.
When an analyst spends three years building investigations by hand, they accumulate something hard to name: a sense of when a process tree looks wrong, when an authentication pattern is suspicious before the rule fires, when a "benign" finding deserves a second look. That intuition is built by friction, by chasing down false leads, by being wrong and understanding why.
Now picture the same three years spent clicking approve on AI-generated verdicts. The analyst sees the conclusion but never builds the reasoning that produced it. They learn to trust the output. They do not learn to interrogate it. Gartner's framing is precise here: if analysts spend their time approving AI decisions rather than constructing investigations themselves, the underlying expertise atrophies. The work still gets done. The capability to do the work without the machine quietly disappears.
The danger is concentrated exactly where it hurts most. Automation excels at the well-defined, high-frequency cases that dominate the alert queue. The incidents that actually threaten an organization, the novel intrusion, the supply chain compromise, the attacker deliberately operating below detection thresholds, are precisely the ones where automated confidence is least reliable and human judgment matters most. A SOC optimized entirely for the common case is structurally unprepared for the case that matters.
The Human-on-the-Loop Trap
Much of the industry has reframed this problem as a shift from "human-in-the-loop" to "human-on-the-loop." The pitch is appealing: let the AI handle volume, let humans supervise strategy. In principle that is the right division of labor.
In practice, "on the loop" can become a polite name for not being in the loop at all. Supervision without engagement is not supervision. A human who reviews a stream of machine verdicts, almost all of which are correct, develops automation bias fast. After the two-hundredth correct verdict, the two-hundred-and-first gets a rubber stamp. The supervisory role looks like oversight on the org chart and functions like a rubber stamp on the floor.
This is the trap worth naming explicitly, because it is invisible in the metrics. A human-on-the-loop SOC and a SOC with severely eroded skills can produce identical dashboards right up until the moment they do not. The difference only reveals itself under pressure, which is the worst possible time to discover it.
The goal, then, is not to abandon supervision models. It is to make sure the humans doing the supervising still have the muscle to override the machine when the machine is wrong, and the judgment to recognize when that moment has arrived.
Countermeasures That Actually Build Muscle
The good news is that the practices that fight skill erosion are not exotic. They are mostly old disciplines applied with new intent. The shift required is treating skill maintenance as a deliberate operational activity, not something that happens automatically as a byproduct of doing the job, because in an automated SOC it no longer does.
Rotations off the automation. Periodically pull analysts off the assisted queue and have them work raw telemetry. Not as punishment, but as practice. An analyst who has to build an investigation from logs once a month keeps the underlying skill alive. One who never does loses it.
Purple teaming as routine, not theater. When red and blue actually collaborate, defenders are forced to reason about attacker behavior rather than wait for a rule to fire. The 2026 conversation around purple teaming is worth a caveat: a red team and a blue team running scripted exercises in the same room is not purple teaming, and increasingly the handoffs themselves are being automated. That is fine for coverage, but the learning value for humans comes from the messy, unscripted reasoning, so protect that part from full automation.
Tabletop exercises for novel scenarios. Tabletops cost almost nothing and exercise exactly the judgment that automation cannot. Run scenarios the playbooks do not cover, where there is no automated answer to approve, and force the team to reason out loud.
Deliberate manual practice. Treat investigation like a craft that needs sharpening. Capture-the-flag exercises, detection engineering work, threat hunts where analysts form and test their own hypotheses. The common thread is that the analyst produces the reasoning, not just the verdict.
Keep humans genuinely in the loop on the hard cases. Route the ambiguous, the novel, and the high-stakes to humans who investigate from scratch, while letting automation own the high-volume routine. Gartner's own recommendation points the same direction: aim for task and workflow optimization rather than end-to-end automation, and scale human roles like incident investigator in areas such as alert contextualization rather than trying to replace the workflow outright.
None of this is free. Every one of these practices spends time that could be spent closing tickets. That is the actual decision in front of security leaders: treat that time as overhead to be minimized, or as the premium you pay to keep a capable team. The 2030 projection is, in large part, a forecast of which choice most organizations will make by default.
Reframing the Metrics
Part of why skill erosion goes unmanaged is that nobody measures it. SOC dashboards track throughput, response time, and coverage. None of those capture whether the humans can still operate without the machine.
If you want to manage this, you have to measure it, even crudely. How do analysts perform on incidents where automation offered no answer? How often does a human override an automated verdict, and how often is that override correct? Can your tier-one team build an investigation from raw telemetry under a time limit? These are imperfect proxies, but they at least make the invisible visible. A SOC that never measures its human capability has no way of knowing it is eroding until it fails.
The cultural reframe matters too. Approving correct AI verdicts all day should not be the definition of a good analyst. The analysts worth keeping are the ones who can tell when the confident answer is wrong, and that skill is built by doing the hard work, not by supervising it.
How Safeguard Helps
Safeguard is built on the premise that reliability lives in the verification and orchestration layer above the model, not in any single model's confidence. Our Multi-Agent TAOR Deep Think AI engine runs findings through multiple verification agents rather than emitting one unchallenged verdict, which cuts false positives and, just as important, surfaces the reasoning behind a finding so analysts can engage with it instead of rubber-stamping it. Because the platform is model-agnostic and bring-your-own-model, engines like OpenAI Daybreak or Anthropic Mythos plug in as components under that verification layer, and we measure value in cost per verified finding, the metric that keeps both the automation and the humans honest. If you are working out how to scale your SOC without dulling your analysts, reach out.