Agents in the Wild

This directory is no longer updated. Last update: November 16, 2026.

Salesforce's SATA agent matched human security analysts about 95% of the time on past cases

Salesforce built SATA, an agent for first-line triage of several hundred security alerts a day. Before production, the team compared it with analyst decisions on historical cases. Low-confidence cases stay with people.

Organization
Salesforce (Detection and Response)
Industry
Enterprise Software
When
May 2026
Evidence
Company reported
Agents and tools
SATA Agent

Reported outcome

Roughly 95% agreement with human analysts on historical security cases, measured before production use. Salesforce gives no production accuracy or time-saved figure.

  • roughly 95%
    Agreement with analysts (historical cases)
  • several hundred per day
    Alert volume
  • 80,000
    Employees covered
  • 20% reduction
    Containment time goal (not yet met)

Caveats

Stated by the company that uses the agent (blog, earnings call, interview). Not independently audited.

  • 95% is agreement with analysts on past cases, not verified correctness in production.
  • Salesforce gives no sample size for the test.
  • Speed results come from initial testing and early estimates. The post gives no time-saved number.
  • The 20% cut in containment time is a target, not a result.
  • Salesforce sells AI agents, so it has a commercial interest in this story.

What the agent does

SATA does first-line triage for Salesforce's detection and response team. It reviews signals, gathers context from logs and other sources, and prioritizes the cases that need an expert.

Every decision gets a confidence score. Lower-confidence cases stay with people. Multiple agents review decisions, and spot checks validate automated outcomes.

What happened

Before production use, the team ran SATA against historical security cases and compared its answers with analyst decisions. That test produced roughly 95% agreement.

In initial testing, the agents triaged hundreds of cases in what Salesforce calls a fraction of the manual time. The next target is a 20% cut in incident containment time.

What to copy

  • Replay past cases and compare with past human decisions before going live.
  • Score confidence on every decision and keep the low scores with people.
  • Keep spot-checking after launch. Agreement before launch does not prove accuracy after it.

Starter prompt

Written by us, not by Salesforce. Use it to replay closed cases, not live alerts.

Closed security case: {case_without_final_decision}

Give your triage decision: benign, suspicious, or confirmed threat.
Give a confidence from 0 to 100 and the evidence you used.
List what data you would still need.
Do not look at or guess the analyst's final decision.

Sources

  1. How Salesforce Built an AI Security Agent for Autonomous Threat Triage — Salesforce Engineering, May 19, 2026

Added Nov 9, 2026 · Sources last checked Nov 16, 2026.

Get new use cases weekly

One short email. Sources included.

The newsletter is paused. Past issues stay in the archive.

Related use cases