Innovation & Tech

Security AI needs context as teams face 4,400 daily alerts

Unite.AI says security agents need context, bounded autonomy and pre-scaled governance as teams face over 4,400 daily alerts and enterprises manage 2 million endpoints.

By Olivia Hart

4 min read

AI Can Scale Security. But Scale What?
AI Can Scale Security. But Scale What?Conny Schneider / unsplash

What's News

  • Some research cited by Unite.AI puts average daily alerts per security operations team above 4,400.
  • Some enterprise environments contain 2 million endpoints.
  • Fine-tuning for individual enterprise customers can support large-scale investigations at a more predictable cost.
  • The essay warns that using frontier models for millions of investigations will bankrupt companies.
  • Recommended metrics include accuracy, false-positive and false-negative rates, remediation time, cost, escalation quality and policy adherence.

By Olivia Hart

Can AI keep up with the alert volume?

Some research cited in a Unite.AI Thought Leaders essay puts the average number of daily alerts received by a security operations team above 4,400, a volume that exceeds realistic manual review capacity.

The essay notes that estimates vary, but describes teams as swamped by more activity, telemetry and alerts than they can investigate manually. It expects the burden to grow as organizations expand, add tools and AI enables more attacks.

The essay points to enterprises with 2 million endpoints, where manual investigation puts defenders at a disadvantage against increasingly agentic attackers. AI agents can absorb investigations and repetitive work, but more data does not automatically improve threat detection or response.

“Without this context, automation risks scaling noise and generating irrelevant or misleading conclusions,” the essay says. An agent examining millions of endpoints must know what it seeks, what behavior is unusual in that environment and which findings merit further investigation.

Takeaway: More alerts do not call for more automation; they call for automation that understands the environment.

What does organizational context actually mean?

Security environments differ in infrastructure, applications, operational requirements, regulatory obligations and risk tolerance, according to the essay. Defenders must keep applications running while meeting compliance requirements and avoiding operational disruption, so identical signals cannot carry identical meaning across environments.

Agents need organizational knowledge: historical investigations, established procedures and experienced analysts’ judgment. The essay says much of it cannot simply be fed in a spreadsheet; it must be observed in action and learned from defenders at work.

Enterprises must translate that institutional knowledge into systems capable of making consistent, context-appropriate decisions at machine speed. They also need mechanisms to capture prior case decisions, procedures, evidence interpretation, custom processes, custom tools and institutional knowledge held informally by staff.

Takeaway: Useful AI must learn how the organization works, not merely consume more data.

How can enterprises put context into AI systems?

The essay presents model specialization as one option. 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.

Fine-tuning is not the only route to context. Retrieval, memory, workflow orchestration and explicit policy controls are alternatives, but the central test is whether the selected method improves performance in the environment where the agent will operate.

Performance must also be evaluated as models, workflows and enterprise environments change, the essay says. Synthetic environments, or digital twins, can test different agent versions against representative settings when production data is unavailable or outdated.

Takeaway: Context matters only when the chosen design improves results in the live environment.

Where should agents act without human approval?

The essay says correctly identifying a threat solves only part of the challenge. Deployment must define authorized actions, when human approval is necessary and how to avoid disrupting legitimate business operations; trusted use depends on clear boundaries between machine and human decisions.

The essay says not every organization can rely on frontier-model investigations for every alert because the cost will be far too high. “Companies that use frontier models for millions of investigations are going to bankrupt themselves,” it warns. It therefore calls for the most efficient investigation method, whether a small language model (SLM) or a proprietary model.

The essay’s rule is proportionate autonomy. Routine, well-understood tasks may run autonomously; consequential or disruptive actions may require human approval or additional authorization. Coordination across teams also matters: a threat requiring a firewall configuration change must pass to the responsible team with authorization and accountability for remediation.

Governance and workflow design are therefore core deployment requirements, not controls added after agents scale. The essay treats authorization, handoffs and accountability as design elements from the outset.

Takeaway: Autonomy expands only when organizations can define its limits, costs and responsibility.

How should leaders judge security AI?

The essay says organizations should judge agents by outcomes rather than activity alone. It recommends tracking:

  • Investigation accuracy
  • False-positive and false-negative rates
  • Time to investigate and remediate
  • Cost per investigation
  • Escalation quality
  • Adherence to authorization policies

Continuous evaluation remains essential as models and workflows change. Testing versions against representative cases and using synthetic environments can expose regressions before updates reach production, according to the essay.

Security leaders must balance the benefits of automation against operational deployment risks. Application availability, compliance obligations, organizational risk tolerance and implementation timelines all shape how quickly capabilities can be adopted, the essay says.

The essay says AI can address a longstanding security capacity problem, but raw scale is not the goal. The opportunity comes from 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.

Takeaway: The next proof point is operational: teams must show that agents can reduce cost and remediation time without increasing false positives, unauthorized actions or service disruption.

Source: unite.ai

Share this article:

More from Olivia Hart

Olivia Hart

Show full bio

Staff writer covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.

2,535 articles

Related articles

« Previous article