Thought Leaders
AI Can Scale Security. But Scale What?

Key PointsProcessing more data does not automatically improve threat detection without…
View more
- Processing more data does not automatically improve threat detection without organizational context.
- Fine-tuned models can support large-scale investigations at more predictable cost than frontier models.
- Autonomy should be proportionate to risk with governance designed before agents are scaled.
Enterprise security teams today face more activity, telemetry and alerts than they can realistically investigate manually. Estimates vary, with some research suggesting the average security operations team receives more than 4,400 alerts a day. As organizations expand and add more tools, and AI enables more attacks, that number is bound to grow. We’re already seeing examples of enterprises that have two million endpoints in their environment, and security teams that are investigating these alerts manually are not fighting a fair fight against attackers that are increasingly agentic.
For defenders, AI agents can change that equation by taking on investigations and other repetitive tasks across the environment. However, the ability to process more data does not automatically translate into better threat detection or response. What’s needed is not just scale but context.
The Importance of Context in Evaluating Scale
An agent investigating millions of endpoints must understand what it is looking for, what constitutes unusual behavior in that particular environment and which findings warrant further investigation. Without this context, automation risks scaling noise and generating irrelevant or misleading conclusions.
And it’s important to keep in mind that not every organization will be able to rely on frontier model investigations for each alert; that will get far too costly.
Security environments differ in their infrastructure, applications, operational requirements, regulatory obligations and risk tolerance. Defenders must keep applications running while managing compliance requirements and avoiding operational disruption.
To be effective, agents must incorporate organizational knowledge, including historical investigations, established procedures and the judgment of experienced analysts. Much of this knowledge isn’t something that can just be fed to an agent in a spreadsheet or other format; it’s something that needs to be observed in action, by learning from experienced defenders at work.
For enterprises, the big question here is how to take that institutional knowledge and translate it into systems capable of making consistent, contextually appropriate decisions at machine speed.
How to Operationalize Organizational Knowledge
Organizations need mechanisms to capture and apply the knowledge that informs investigations, including previous case decisions, established procedures and how analysts interpret evidence. Enterprises may have “custom” types of processes, tools and tribal knowledge that need to be captured, as well.
Model specialization is one potential approach. Fine-tuning models for individual enterprise customers can support large-scale investigations at a more predictable cost while maintaining strong performance. Fine-tuned open-source models may be particularly useful for tasks involving very large endpoint populations.
Organizations must also evaluate agent performance as models, workflows and enterprise environments change. Synthetic environments, or digital twins, can provide representative settings for testing different agent versions when production data is no longer available or has become outdated.
Fine-tuning is not the only way to provide context. Retrieval, memory, workflow orchestration and explicit policy controls are alternative approaches. The central requirement is that the chosen method improves an agent’s performance in the environment where it will operate.
Autonomy at Scale Requires Clear Boundaries, Oversight and Coordination
Correctly identifying a threat is only part of the challenge. Agents must understand which actions they are authorized to take, when human approval is necessary and how to avoid disrupting legitimate business operations. Determining where agents can act autonomously and where human decision-making is required is essential to trusted deployment.
Autonomy at scale also requires economy of scale. Companies that use frontier models for millions of investigations are going to bankrupt themselves. That makes finding the most efficient method of investigating with an AI model, whether a Small Language Model (SLM) or a proprietary one, paramount.
Autonomy should be proportionate to risk. Routine, well-understood tasks may be suitable for autonomous execution, while consequential or disruptive actions may require human approval or additional authorization. Enterprise security also depends on coordination across teams. An investigation that identifies a threat requiring a firewall configuration change must transition seamlessly to the team responsible for that system, with appropriate authorization and accountability for remediation.
Governance and workflow design are therefore essential components of agent deployment, not controls that can be added after the technology has been scaled.
The Next Measure of Security AI Success Is the Quality of Its Outcomes
Organizations should assess agents using outcome-based measures, including investigation accuracy, false-positive and false-negative rates, time to investigate and remediate, cost per investigation, escalation quality and adherence to authorization policies.
Continuous evaluation is essential as models and workflows change. Testing different agent versions against representative cases and using synthetic environments can help organizations identify regressions before deploying updates into production.
Security leaders must balance the benefits of automation against the operational risks of deployment. Application availability, compliance obligations, organizational risk tolerance and implementation timelines all influence how quickly new capabilities can be adopted.
AI offers security teams an opportunity to address a longstanding capacity problem, but raw scale is not the end goal. The real opportunity lies in combining machine speed with organizational context, appropriate authority and human judgment so agents can distinguish what matters, take the right actions and involve the right people when necessary.












