Arctic Wolf Says Its AI Agents Now Resolve 60% of SOC Investigations Alone, and Buyers Should Ask Every Agentic SOC Vendor How Often Analysts Overrule the Machine

LONDON - Arctic Wolf put a number on the table on Oct. 27 that every agentic SOC vendor would like to own. Its 2026 Agentic Security Operations Report says AI agents in its Aurora Agentic SOC now resolve 60% of security investigations with no human intervention, that another 38% run as a partnership between agents and analysts, and that only 2% are handled by a person alone. When the first agentic elements went live in August 2025, the company says, humans handled roughly 81% of its caseload without AI assistance.

Those are large claims from the operator of a large service. The report says its figures come from a customer base of more than 11,000 organizations in more than 100 countries, and that more than 10 trillion telemetry events flow through its platform each week. It is also, by its own account, internal data. "Unless noted, figures in this report reflect our internal data for August 2025 through September 2026," the fine print says, and customer examples, calculator outputs and dollar figures "are illustrative."

None of that makes the report useless. It makes it a vendor's account of its own service, and SOC leaders should read it the way they would read any such account: for what it measures, what it leaves out and what it reveals almost by accident.

What the numbers measure

The headline figures describe the division of labor inside Arctic Wolf's operation. Monthly investigation volume grew 21% while alert volume fell about 19%. The average time from an alert-generating event to an actionable ticket fell 36%, and the median fell 24%. The platform opens 13 investigations for every billion observations and surfaces one alert for every 175.6 million. The share of investigations that escalated into a customer alert dropped from 40% to 27%, while the share of high- and critical-severity tickets moved only from 4.63% in fiscal 2026 to 4.67% so far in fiscal 2027.

Read together, those numbers say that more cases are being opened and more are being closed before a person sees them. They do not say how many of the cases an agent closed were closed correctly.

The report does describe a quality loop it calls the AI Trust Engine. Golden datasets of expert-validated investigations set the benchmark. Every change is tested offline and scored on "verdict accuracy, speed, and cost." A separate judge model grades production responses. Agents earn autonomy in stages, starting in silent mode, then recommending actions to analysts, and acting on their own only once validated and under ongoing human audit. Analyst corrections flow back into the datasets. Changes that fall short of "the quality threshold" do not move forward, the report says. It does not publish that threshold, the verdict accuracy it measures, or how often analysts corrected an agent.

The report also describes its measurement window in three ways. The key takeaways page says investigation volume grew 21% "in the eight months after" the Aurora Agentic SOC launched. The narrative says human-only work fell to 2% "one year later." The closing page says the figures "describe 13 months of operating this model at scale." Each may be accurate for a different series. A buyer comparing vendors should still ask which months every number covers.

The most useful page is about an AI that was wrong

The strongest material in the report is not a percentage. It is a case study headed "Why Humans Will Always Be Essential." A customer had outsourced its IT environment to a consultancy that used an AI tool to assess it. The tool concluded the customer had been compromised, reaching that verdict "from a thin set of indicators and publicly available reporting, while presenting the conclusion without qualification." The consultancy accepted it without independent validation.

Arctic Wolf's agentic SOC found no evidence of compromise. The consultancy disagreed, so the case went to Arctic Wolf's incident response team, whose forensic analysis, configuration reviews and artifact validation reached the same conclusion. An independent third party reviewed it as well, and the consultancy eventually aligned with the finding. The report says the costs to the organization were real even though nothing had been breached, and that the customer took the matter to its board as a governance issue.

"An AI system produced a specific, serious, and incorrect conclusion, then presented it with complete confidence," the report says. Its lesson is the right one: "AI output is evidence for a human to validate rather than a verdict to accept."

That sentence is the standard every agentic SOC, Arctic Wolf's included, should be held to. A confident false positive cost one organization an investigation, an incident response engagement and a third-party review. A confident false negative, an agent closing a real intrusion as benign, would cost far more and would be much harder to find, because nobody reopens a case that was closed.

The report's other spotlight makes a related point. In July 2026, a phishing operation known as Payroll Pirates used an adversary-in-the-middle proxy to steal Microsoft session tokens, then used Microsoft Graph inside compromised mailboxes to find payroll, HR and finance staff, aiming to redirect employee salaries. According to the report, the pattern started with "a single flagged account": an Arctic Wolf analyst noticed an account pulling up a long list of HR and payroll employees, which did not fit its normal behavior. The company then hunted across thousands of customer environments, coordinated containment for hundreds of compromised accounts and wrote six new detections. In the report's own telling, the first catch was a person's.

What buyers should ask for

Arctic Wolf deserves credit for writing escalation into its design. "An agentic SOC that cannot recognize the cases it should not handle alone is not trustworthy, regardless of how well it performs on the other 98%," the report says. It also cites the hesitation it is selling against. In the company's own survey of 1,350 security and IT decision-makers, only 14% had made AI central to their security operations strategy, and only 53% said they trusted AI to autonomously perform even a narrowly defined action such as blocking a malicious IP address at a firewall. The report adds that in the 2026 SANS AI in Cybersecurity Survey, nearly two-thirds of security professionals said they had received AI-generated guidance they later determined was wrong.

The share of investigations an agent resolves alone is a measure of throughput. Trust needs different numbers. Before signing with any agentic SOC provider, SOC and IR leaders should ask for:

• The rate at which analysts overturned or corrected agent verdicts, by case type, over the same period as the autonomy figures.
• How autonomously closed cases are sampled and rechecked after closure, and how many were later reopened as real threats.
• The verdict accuracy an agent or model change must reach in offline testing before it touches production.
• Which response actions agents may take without a person, such as isolating a device, disabling an account or forcing a credential reset, and which ones the customer must preapprove.
• Exact date ranges and denominators for every headline figure.
• The evidence behind any autonomous closure in the customer's own environment, available on request.

Dan Schiappa, Arctic Wolf's president of technology and services, said in the announcement that organizations "can no longer rely on AI-only or human-only security operations to keep pace." On the evidence of the company's own case study, he is right. The agentic SOC providers that earn lasting trust will be the ones that publish how often their agents were wrong, not only how often they worked alone.