SOC Copilot: Agentic AI for Security Operations
Agentic copilot that auto-investigates incoming security alerts, assesses risk, and tells analysts what to do next, grounded in approved SOC playbooks.
Problem
The enterprise Security Operations Center (SOC) handles approximately 300 alerts each week through this workflow. Routine investigations such as risky logins required 30 to 60 minutes each, while complex, hard investigations (roughly 10% of total volume) routinely took up to a full working day per alert.
Analysts faced alert fatigue, repetitive manual pivoting between disparate consoles, and varying levels of playbook adherence across rotating shifts. As alert volume scaled, security leadership needed to absorb volume without adding headcount.
Discovery & User Research
In October 2025, I initiated analyst shadowing sessions across multiple security tiers to understand where triage time was actually lost. The findings revealed that analysts did not struggle with taking final mitigation actions; they lost hours gathering context across systems, cross-referencing internal SOPs, and synthesizing initial findings.
I structured a proof of concept and presented the discovery insights at the Scaled Agile PI planning session (PI 25.4), aligning engineering, architecture, and security leadership on an agentic solution.
Key Architectural Decision
Should we adopt the general enterprise Microsoft Copilot with prompt grounding, or build a purpose-built SOC Copilot?
We chose to build a purpose-built SOC Copilot in Microsoft Copilot Studio. General enterprise copilots pulled broad ungrounded context and lacked operational integration with ticketing infrastructure. A purpose-built product allowed us to:
- Enforce strict grounding exclusively in approved, version-controlled SOC incident playbooks.
- Eliminate generic hallucinated mitigation steps by locking responses to defined security taxonomy.
- Implement a staged shipping strategy: deliver a grounded assistant first, followed by connected pipeline automation.
Shipping Strategy: Two-Stage Evolution
Release 1 (January 2026): Grounded Copy-Paste Assistant
Instead of delaying launch while awaiting an enterprise cloud tenant migration, we shipped a v1 RAG assistant directly in Microsoft Teams. Analysts pasted alert payloads and received instant playbook-grounded guidance. The usage telemetry and positive feedback proved clear operational value, securing executive buy-in and funding for the complete tenant migration and integration.
Release 2 (March 2026): Connected Agentic Pipeline & Web UI
Following the cloud tenant migration, we integrated TheHive alerts directly into the agent via Power Automate. When an alert arrives, the agent automatically reasons over severity, observables, indicators of compromise, and analyst notes, cross-referencing verified playbooks.
User feedback on the Teams interface showed it felt slow and opaque. I led the move from Teams to a dedicated, responsive web application that exposes the agent's step-by-step reasoning, tool invocations, and playbook citations in real time. Transparency resolved skepticism, and adoption held steady.
Product Management Role
I owned the product end-to-end, partnering with two engineers:
- Defined product specifications, guardrails, failure-mode behaviors, and engineering acceptance criteria.
- Designed the end-to-end agentic workflow, prompt grounding architecture, and human-in-the-loop governance.
- Established weekly sampling quality control where senior analysts review agent suggestions and flag discrepancies.
- Personally demoed the system at the ART's PI 26.2 system demo to Business Owners and portfolio leadership, authored the complete analyst onboarding guide, and trained team leads globally.
Results & Measurable Impact
- 89% Active Analyst Adoption: Achieved across global operations. The 5 non-adopting analysts are senior staff who exclusively handle rare, highly critical incident escalations outside standard playbooks.
- 10-15 Minute Investigation Time: Research and triage time on hard alerts fell from about a full day to 10-15 minutes, validated on a 50 real-alert test suite and direct analyst observation.
- ~₹50L / yr Value Created: Estimated analyst time saved annually, enabling the security operations center to absorb growing alert volume without adding headcount.
- Quality Control: Continuous weekly review loop maintains high fidelity and prevents drift.
What's Next
Our upcoming roadmap focuses on deepening connected security data sources:
- Pulling live identity and sign-in signals directly from Microsoft Entra ID.
- Extracting device compliance telemetry via Microsoft Intune.
- Querying incident correlations in Microsoft Sentinel and live endpoint detection signals from CrowdStrike.