Back to Devoured - November 20, 2026
DEVOURED
How security leaders see AI being used in the SOC

How security leaders see AI being used in the SOC

DevOps Elastic
While 96% of security teams use AI, only 18% have enforced governance, leading to a disconnect between executive confidence and practitioner reality.
What: An Elastic survey of 400 security pros found that despite widespread AI adoption, most teams lack governance and visibility. A significant gap exists between executives—who see no negative impacts—and practitioners, who report hidden costs and tool sprawl.
Why it matters: Security teams are repeating the same mistakes seen in early cloud adoption: letting unmanaged tools proliferate before establishing a unified data foundation or cost-control framework.
Deep dive
  • Governance Deficit: Only 18% of teams have formal AI usage frameworks.
  • Human-in-the-Loop: 52% of teams require human authorization for all AI decisions; 89% prioritize future workflows that keep humans in the approval chain.
  • Hidden Costs: 57% of teams report cost increases due to tokens and licenses; 62% expect budgets to increase further in five years.
  • Tool Sprawl: Teams juggle four disparate categories of AI tools, complicating threat intelligence narratives.
  • Executive Disconnect: Leadership is 4x more likely to report 'no negative impacts' than frontline practitioners.
Decoder
  • SOC (Security Operations Center): A centralized unit that deals with security issues on an organizational level.
  • Token: A basic unit of text processing used by LLMs; the fundamental unit of cost in most AI API pricing models.
Original article

How security leaders see AI being used in the SOC

96% of security teams are now using AI. Whether it's through chatbots, native platform features, in-house agents, or custom-built tools, AI has become standard equipment in the SOC. Yet, only 18% of teams have structured and enforced frameworks for how AI should be used in the SOC.

Security leaders aren't struggling to adopt AI. They're struggling to control it, explain it, and defend it when a board or a regulator asks What happened? Why did the system do that? Without an answer to these questions, the technology isn't the risk… the lack of insight into its process is.

Different levels of AI autonomy

Data from the recent AI Adoption in the SOC report — based on a survey of more than 400 cybersecurity professionals using AI in their day-to-day — reveals what teams are actually allowing their AI agents to do:

  • 52% require a human to authorize all AI decisions before it takes any action.
  • 43% trust low-risk tasks but require verification for higher-consequence outputs.
  • 3% permit AI to take actions it deems appropriate with no human in the loop.

Looking ahead, 89% of respondents say the future of AI in security operations is agentic workflows paired with detailed human oversight. This means security professionals aren't rejecting agentic AI; they're asking for it with a human as a central part of the approval chain. This is a deliberate design choice, and it's the right one.

We insist agents have all the guardrails — show your work, ask for sources, then verify, verify, verify.

This is a necessary, calculated approach toward a practical tool that’s only going to grow as a share of analyst usage.

Security is held to different standards with AI

Every function in the business benefits from faster AI: Sales wants faster account research, finance wants faster forecasts, and IT wants faster ticket resolution. But security operations as a function carries the burden of higher consequence when AI gets ahead of itself.

If, for example, an AI agent flags an incident, takes action, and something breaks, then a legitimate business process gets blocked, a false positive triggers a costly response, or worse, a real threat slips through because the agent made a probabilistic guess that looked confident but wasn't correct. And now, you're in front of the board or a regulator. They want to know what happened, why the system acted the way it did, and who approved it.

"We trusted AI" isn't a viable answer.

Human oversight in AI-driven security is business risk reduction. A transparent, explainable AI workflow gives you something to point to: here's what the system saw, here's what it recommended, here's who approved the action, and here's why. That paper trail is what separates a defensible incident response from a credibility crisis.

And it works both ways, too. While AI is incredible at flagging anomalies that may otherwise have been missed, only human analysts understand the broader business implication of halting operations for every low-incident flag. Analysts need this visibility into exactly how the AI reached its conclusion in order to tune it appropriately to scale with the business.

AI will find a vulnerability in legacy code and just decide the application can't be used. A human is the one who has to do that risk analysis. Sometimes you override security because you actually have to run a business. That's the human piece: knowing what risks are necessary to enable business operations.

The hidden costs nobody budgeted for

There's a second thread running through the report’s research that CISOs need on their radar: AI is not the free efficiency win it was pitched to be. Yes, the benefits are real:

  • 61% of teams report faster investigations.
  • 53% report an improved overall security posture.
  • Nearly all teams (99.8%) report some benefit from AI adoption.

But 85% also report negative impacts, and the most common one cited by 57% of respondents is cost increases from tokens, tools, and licensing.

This isn't a rounding error. More than half of security leaders say their overall security budget is higher today because of AI adoption, and that number climbs to 62% when projecting five years out, with 29% expecting costs to be "much higher." Efficiency gains are real, but so is the bill, and without cost controls built in from the start, AI spending can spiral quietly until it shows up as an uncomfortable line item at budget review.

The same pattern shows up in tool sprawl:

  • 75% of security teams use general-purpose chatbots.
  • 66% use native AI features in security tools.
  • 43% use in-house tools built by IT.
  • 32% use custom agents built by the security team.

That's four categories of AI tooling, often without a shared data foundation connecting them. Fragmentation like that makes it harder for analysts to build coherent threat intelligence narratives and nearly impossible for leaders to present a single, confident answer about how AI is being used across the organization.

This use of this “sidecar AI” also wastes spend on prompts and queries for models that don’t have the proper context to be helpful. Or, even worse, AI without visibility into the full security telemetry data set can give teams the ok, providing the team with false confidence.

The leadership-practitioner gap that needs to close

Here's something that should concern every CISO reading this report: Your frontline team may be experiencing a very different reality than the one you're describing upward.

Executives are more than four times as likely as practitioners to say AI adoption has had no negative impacts (26% versus 6%). Team managers and practitioners are more likely to be disappointed that AI didn't meet expectations than leadership is (38% versus 28%). Executives also overestimate how often their teams rely on native AI tools (75% reported versus 60% actual) and underestimate how often teams build their own workarounds (38% believed versus 49% actual).

This gap reveals a false confidence surveyed CISOs have about their teams' AI adoption. If frontline analysts are quietly patching together chatbots, in-house scripts, and sidecar AI tools because the native AI tooling in their existing platform isn't getting the job done, you won't find out from a dashboard.

Closing this gap starts with asking practitioners directly what they're actually using to get the job done, rather than assuming the AI solution one team or manager chose at a point in time is scaling across every analysts’ workload into today’s applications.

What makes AI trustworthy enough to act on

So, how does a CISO move from "AI gives us faster answers" to "AI gives us answers we can defend?” As with most anything around discussions of AI performance, it all comes down to context.

An AI recommendation is only as trustworthy as the data behind it. If your threat intelligence, alert history, and investigation context live in disconnected systems, AI doesn't have the full picture, and neither do you. Teams in this research cited exactly this problem: 29% said their data lacks the structure needed for effective automation, and 37% said they struggle to meet governance, compliance, or audit requirements as a direct barrier to adopting automated workflows.

A unified, searchable data foundation changes that equation. When an AI agent can draw on the complete, correlated context of your environment rather than a narrow slice of it, its recommendations become explainable. You can trace the "why" behind every suggested action.

The architecture question CISOs should be asking their teams right now is not "which AI tool should we buy next?" But rather "Can we actually explain, end to end, how our AI reached this conclusion?" If the answer is no, that's the gap to close before adding more automation.

Building security operations that are ready for what's next

The survey data is unambiguous: teams with mature AI adoption report significantly more value than teams still in the pilot stage (100% report moderate or significant value compared to just 60% at the limited-use stage).

The argument is for building the foundation that makes AI adoption durable instead of a source of ongoing risk. A few starting points to accomplish this:

  • Make human approval the default, not the exception. The data shows this is already where most security teams land instinctively, but this needs to be formalized from the top down.
  • Build cost visibility into every AI deployment from day one. Don't let token spend and licensing creep become a surprise at renewal time.
  • Unify your data before you scale your automation. Explainability depends on context, and context depends on having your threat intelligence, alerts, and investigation history in one searchable place.
  • Check in with your practitioners directly. The gap between what leadership believes and what analysts experience is often the first sign that governance isn't keeping pace with adoption.
  • Treat every AI-assisted decision as something you may need to explain later. Build the audit trail now, not after an incident forces the question.

The organizations that get this right won't be the ones with the most AI tools, they'll be the ones that can explain, clearly and confidently, exactly how every AI-assisted decision was made and why a human stood behind it.

Explore all 56 articles in today's digest

Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi