Healthcare professionals monitor a large AI-powered dashboard displaying hospital data, alerts, and cybersecurity protections.
St. Luke's University Health Network says its Security Alert Triage Agent now gives security analysts back about 260 hours a month. That is the headline of a new Microsoft customer story. It is a vendor-published account, and Microsoft hasn't had the numbers independently audited. The story is still useful for SOC teams because it also describes how the network governs AI agents before they go live.

Healthcare professionals monitor a large AI-powered dashboard displaying hospital data, alerts, and cybersecurity protections. What Microsoft reports​

Microsoft documents the Triage Agent as an autonomous agent embedded in Microsoft Defender that helps security teams triage alerts at scale. Microsoft's story says the agent, formerly the phishing triage agent, now returns roughly 260 analyst hours per month at St. Luke's.

The network has served Pennsylvania's Lehigh Valley since 1872. Microsoft's story gives its current footprint as 16 campuses, 300 outpatient sites and more than 24,000 employees.

AI use has grown well beyond the SOC. Microsoft says St. Luke's runs about 27 AI solutions and 22 predictive models, producing about 3 million predictions a day. The story doesn't explain how it counts a "solution" or a "model". It doesn't list them or say how far each has been rolled out.

How the 260-hour figure compares with last year​

An earlier Microsoft story, from September 2025, said the same agent saved "nearly 200 hours monthly". At that point the agent was still called the Phishing Triage Agent. That earlier account said the agent was autonomously closing thousands of false-positive alerts.

The gap is about 60 hours, but I wouldn't read it as a clean year-over-year productivity gain. Neither story gives a baseline period, a staffing denominator or a calculation method. We can't tell whether the workload, adoption or the way hours were counted changed.

Microsoft's own documentation also shows the product has moved on. The Security Alert Triage Agent is the same agent as the Phishing Triage Agent, with extended capabilities to triage a broader set of alert types. Microsoft's Learn page describes it as available only to customers who are part of the preview. That page may be out of date, so check current availability and licensing before you plan around it.

Microsoft has also used the St. Luke's result in other material. A Security Copilot blog post says the agent saves analysts at St. Luke's more than 200 hours every month. The Microsoft sources therefore give figures of "nearly 200", "more than 200" and "approximately 260" over time. That is a reason to read all of them as estimates.

The governance model: agents as privileged identities​

The most reusable part of the story is how St. Luke's governs AI before it acts. CISO David Finkelstein says every AI solution needs three things:

  • A clear purpose.
  • An accountable owner.
  • The minimum access necessary.

Before an agent goes live, security teams ask what data it will use, where that data goes, what it can access and how its activity will be governed. Finkelstein's summary is that agents should be treated "like privileged identities, not apps."

Microsoft says St. Luke's uses Microsoft Agent 365 to see which agents exist, who owns them, what they can access and whether that access is still appropriate. It also uses Microsoft Purview to help find and protect sensitive patient information.

Microsoft's technical material shows how those pieces fit together. In a Microsoft Mechanics walkthrough, the Microsoft 365 admin center provides inventory and governance of agents. Entra provides identity with least-privilege access and clear ownership. Defender provides runtime protection and observability. Purview monitors how agents interact with sensitive information. That maps closely onto the purpose, owner and least-access principles St. Luke's describes.

Microsoft's story doesn't give St. Luke's actual configuration, policies, licensing or setup steps. The mapping above is from Microsoft's general product description, not from the hospital's deployment.

Project Perception: six agents under evaluation​

St. Luke's is also evaluating six agents in the Microsoft Defender portal under what it calls Project Perception. They cover:

  1. Incident triage.
  2. Attack-story reconstruction.
  3. Detection authoring.
  4. Posture prioritization.
  5. Threat intelligence.
  6. Attacker-perspective reconnaissance.

These are under evaluation, not confirmed production deployments. Don't treat the 260-hour figure as evidence for them. It belongs only to the triage agent. The story gives no dates, versions or outcome measurements for the six.

Microsoft's documentation does show where these agents surface. The Triage Agent's setup steps start in the Defender portal under Perception > Agents. The Learn page also lists the prerequisites. One is Microsoft Defender for Office 365 Plan 2. Another is turning on Defender for Office 365 in the Defender XDR unified RBAC settings.

Humans stay in charge in the model St. Luke's describes. Agents assemble evidence and recommend next steps, while analysts validate the reasoning and keep decision authority. Associate CISO Krista Arndt says AI is "not replacing our security team." The story doesn't specify approval gates or when human review is mandatory.

The non-security results​

Microsoft also lists outcomes from other parts of the network. Treat them as context, not as proof of what the security agents achieved:

  • Surgical capacity: recovering unused surgical time generated $862,000 in revenue over six months.
  • Patient deterioration program: 38% survival to discharge, against a 22.5% benchmark.

The story doesn't say where the benchmark comes from, how large the cohort was, over what period it was measured, or how it was adjusted. It also doesn't explain the AI mechanism behind the surgical figure. Neither number should be read as a proven causal effect.

St. Luke's is next looking at agents that can move work across Epic, Microsoft, Databricks and ServiceNow. Associate Chief Medical Information Officer Charlie Sonday says the aim is "not another chatbot" but finishing work end to end. That is a stated direction, not something already running.

What IT and security teams can take from it​

  • Give agents an owner before they launch. An agent with no named human accountable for it is an orphan with credentials.
  • Apply least privilege to agents. Scope access to the task, then review it regularly.
  • Keep an inventory. St. Luke's cites visibility into what agents exist and what they can reach as the basis of its approach.
  • Measure and show your baseline. If a vendor claims hours saved, ask how they were counted. The St. Luke's figures show why.
  • Check licensing and preview status. Microsoft's documentation describes parts of this stack as preview. Confirm what your tenant has before you plan around it.

Bottom line​

This is a Microsoft marketing case study, and it should be read as one. The 260-hour number is a vendor-attributed estimate with no published method. The governance principles are the more durable takeaway, because they apply whichever vendor's agents you run. St. Luke's own summary is that security can be what lets an organization scale AI, rather than the brake on it. Whether that holds will depend on what its six Defender agents deliver once the evaluation finishes.

 

References

  1. St. Luke’s scales trusted AI, returning approximately 260 hours to security analysts each month - Microsoft Microsoft Tue, 17 Nov 2026 18:06:56 GMT
  2. St. Luke’s saves nearly 200 hours monthly with AI-powered Security Copilot agents | Microsoft Customer Stories microsoft.com
  3. Deploy AI agents in Microsoft Defender - Microsoft Defender XDR learn.microsoft.com