Nov 19 2026
Security

Building an Agentic SOC: When AI Agents Join Your Security Team

AI agents can accelerate investigations and reduce repetitive work, but security leaders need reliable data, enforceable permissions and clear limits on autonomous response.

Security operations teams spend a lot of time assembling information for investigating alerts, from identifying the affected user to checking device activity and searching other systems for related events.

That workload is growing, leaving analysts with less time to challenge an explanation, investigate an unfamiliar threat or determine how a response could affect the business. Enter the agentic security operations center, which delegates parts of that investigative work to AI agents.

In building an agentic SOC, CISOs must make decisions about which systems agents can access, what evidence supports their conclusions and when a person must authorize action. Faster investigations are valuable only if the organization can trust — and control — what happens next.

Click the banner below to learn how to start building security confidence.

 

What Is an Agentic SOC?

An agentic SOC uses AI agents to select tools, gather evidence and determine investigative steps within boundaries established by the organization. Unlike a chatbot that summarizes an alert, an agent can pursue an investigation across connected systems.

Jeff Pollard, vice president and principal analyst at Forrester, says the change involves granting agents operational authority.

“An agentic SOC delegates investigation and explicitly authorized response tasks to AI agents that can select tools, gather evidence and determine next steps,” he explains.

An agent can investigate suspicious activity and recommend containment, with an analyst reviewing the evidence and approving any response before it takes effect.

WATCH: Why your organization should prioritize cyber resilience for the agentic era.

How AI Agents Reduce Investigation Work

Disconnected tools can turn one incident into several alerts, each requiring an analyst to retrieve context, while fragmented security and infrastructure telemetry contribute to that burden, notes Kamal Hathi, senior vice president and general manager of Splunk at Cisco. 

“Today’s SOC analysts are overwhelmed by alert fatigue and fragmented data, spending more time on research than response,” he says.

For a suspicious login, an agent could retrieve identity information, examine the device’s security posture and compare the activity with historical behavior and threat intelligence. The analyst receives a timeline and supporting evidence instead of starting with isolated records.

The workflow can also extend beyond initial triage, with agents searching for other affected systems, assembling incident reports and checking whether containment stopped malicious activity.

“Buyers should ask for these steps to be demonstrated live, because triage quality depends on the data the agent can see,” says Christopher Kissel, research vice president in IDC’s Security and Trust group.

Kissel recommends testing those capabilities against real evidence.

Kamal Hathi
Today’s SOC analysts are overwhelmed by alert fatigue and fragmented data, spending more time on research than response.”

Kamal Hathi Senior Vice President and General Manager of Splunk at Cisco

Build the Data and Integration Foundation

Agents need access to endpoint, identity, network, cloud and application telemetry, supplemented by asset ownership, vulnerability information and business context. Without that context, a technically plausible response could disrupt an essential service.

Kissel suggests normalizing data into a common model and evaluating integration depth alongside autonomy claims. Connections to security information and event management, ticketing, identity and firewall systems must support the investigative and response tasks the organization intends to delegate.

Security leaders should establish spending limits and investigate repeated tool calls or loops that consume resources without resolving a case. Those compute requirements should include model hosting, inference latency and token consumption.

Hathi advocates federating data across sources to support correlation without requiring every record to be centrally ingested, noting whatever architecture an organization chooses, agents must retrieve relevant evidence reliably and preserve its provenance.

DISCOVER: How data quality and governance can lay a foundation for trusted AI solutions.

Keep Human Review and Oversight Meaningful

Approval requirements should reflect the consequences of an action. Isolating a production server, disabling a critical account or changing a business-sensitive system warrants scrutiny beyond a routine evidence lookup.

Jim Richberg, head of cyber policy and global field CISO at Fortinet, cautions that excessive approval requests can undermine oversight.

“Balanced against this is the risk of consent fatigue: Presenting human operators with a flood of oversight and approval requests risks reducing their ability to give due attention to each one,” he says.

Reviewers need enough time and context to reject a recommendation, with an interface exposing supporting records, conflicting evidence and potential business disruption.

As agents take on routine work, analysts will spend more time auditing results and resolving exceptions. Organizations must also capture operational knowledge that analysts previously carried in their heads. 

“The agent needs standard operating procedures, asset ownership, identity data, change requests and the reasons behind exceptions,” says Michelle Abraham, research vice president in IDC’s Security and Trust group. “Someone must build and maintain that.”

Engineers’ responsibilities expand to workflow testing, output validation, version control and rollback, she adds.

Click the banner below to learn how organizations are using AI for cybersecurity.

 

Prevent Errors From Spreading Between Agents

A mistaken conclusion can become more dangerous when another agent treats it as established evidence. Teams should preserve the difference between source records and interpretations at every handoff.

“Grounding constrains what a model reads, not what it concludes,” Abraham says.

Independent evaluation should challenge verdicts and verify that cited records support them. Missing information should trigger escalation rather than an unsupported conclusion.

Each agent also needs a distinct identity, an accountable owner and permissions limited to its assigned tools and tasks. Operators must be able to revoke access and stop execution.

Pollard recommends enforcing action limits and stop conditions outside the model. Instructions alone cannot constrain an agent that has excessive authority. Response actions should follow approved playbooks, with additional approval for irreversible or widespread changes.

READ MORE: Why you need to rethink identity, access and security for agentic AI.

Establish AI Governance Before Expanding Authority

Audit records should show what an agent accessed, recommended and changed, which policies applied and who authorized consequential actions. Security and legal teams must assess applicable privacy, sector and AI requirements against the actual deployment.

The EU Artificial Intelligence Act applies in stages; organizations should not assume every agentic SOC automatically qualifies as high-risk. Current European Commission guidance places Annex III high-risk requirements in December 2027 and product-related high-risk requirements in August 2028.

“Regulations reinforce what quality governance already requires,” says Dan Schiappa, president of technology and services at Arctic Wolf.

From his perspective, transparency, accountability and meaningful human oversight should remain the top priorities for organizations building or buying an agentic SOC.

“An organization should be able to explain how AI-driven decisions were made, display who is accountable, and keep records of agent actions and reasoning,” Schiappa says. 

DIVE DEEPER: Find out how organizations are leveraging automated incident response and recovery.

Expand Autonomy Through Measured Stages

Schiappa says he sees building an agentic SOC as a “coordinated transformation,” not a single deployment.

“The sequence matters because each phase depends on the integrity of the one before it,” he explains. 

Organizations should begin with reliable telemetry, documented procedures and baseline measures of investigation quality. Introduce read-only assistance for enrichment, summaries and queries before allowing agents to propose responses for human approval.

They should also test historical cases and measure missed threats, incorrect conclusions, inconclusive results, disruption and cost. Sampling automatically closed cases helps expose errors that faster processing might otherwise conceal.

Pollard adds that accountability remains with the security team as delegation expands.

“Operationally, the biggest change is delegating authority to agents,” he says. “SOC analysts and security leaders still own responsibility for consequences.”

J Studios/Getty Images
Close

New Research from CDW Explores AI and Cybersecurity

Learn how AI is helping IT teams manage risk and improve resilience.