Enterprises adopting AI to investigate outages are running into an uncomfortable problem: a system can gather every relevant log, dashboard and incident report and still point to the wrong culprit.
That gap between finding evidence and proving cause was the focus of an interview with Geunwoong "Van" Ryu, founder and CEO of South Korean company Inspirio, according to a report by KoreaTechDesk. His view is echoed by several studies published in 2026 that test how well AI performs root-cause analysis in real and simulated systems.
A Plausible Suspect Is Not a Proven Cause
Inspirio builds Rootr, an enterprise knowledge platform, and Ryu's work centers on how AI links organizational records with technical systems. He described a hypothetical e-commerce case in which payment failures rise soon after a service update, making the release the obvious suspect.
A retrieval-augmented generation system could pull release notes, past incident reports, procedures and dashboards that all seem to support that theory. In his example, however, the real trigger is a promotional rule configured two weeks earlier and scheduled to switch on just after the update.
The release draws attention because of when it happened, while the older rule escapes scrutiny. Relevant documents, in other words, are not the same as causal proof.
Research Shows AI Can Lose Track of Good Evidence
A May 2026 study from Peking University and Huawei reviewed 100 failed investigations by a single-agent AI system. Researchers found 43 cases of what they called evidence dilution, where key causal signals faded as the investigation went on.
In eight of those cases the system saw evidence for the correct cause but dropped it before concluding. In the other 35, the right answer stayed on the list but was ranked below an incorrect one. In one microservice example, traces pointed to packet loss in a recommendation service, yet the AI favored another service it had simply observed more often.
Timing and Dependencies Shape a Defensible Diagnosis
Ryu argues that incident knowledge systems should preserve five kinds of information: precise event timing, system dependencies, historical ownership, links to past incidents and the reasoning behind human decisions.
In the payment scenario, investigators would need to separate when the rule was created, when it became active and when failures began, then show how the rule could reach payment processing. Graph-based approaches such as GraphRAG can help organize these links, though a knowledge graph alone does not prove one event caused another.
Cloud and Kubernetes Studies Reveal Real Limits
A June 2026 study by Amazon Web Services researchers tested a probabilistic causal-graph method on 35 labeled production incidents. It named the exact root cause in 74.3% of cases and placed the correct cause among its top candidates in 85.7%. The framework had been used in more than 800 operational incidents.
Supabase's February 13, 2026 incident report offers a real-world example. A monitoring-service deployment unintentionally enabled an AWS networking restriction affecting its Ohio-region services, and the relevant CloudTrail event appeared in logs but was overshadowed by more prominent networking changes.
In Korea, Honam University researchers introduced KubeIn, which combines Kubernetes dependency graphs, information retrieval and large language models. Tested on 20 simulated failures, it correctly identified the leading cause in all ten it could detect, while the other ten fell outside its coverage. Separately, Samsung SDS presented QAFD-RAG at ICLR 2026, an approach that adjusts graph relationships to each query so retrieval does not simply favor nearby or highly connected information.
Outdated Graphs Can Mislead Investigators
A July 2026 systematic review in Computers, Materials & Continua examined 31 studies on graph-based retrieval in IT operations. Of 22 focused on service-dependency graphs, 16 did not explicitly explain how they handled topology changes or stale graph data.
That matters because a dependency map must reflect the system as it was at the moment of failure, not as it looks today. Services gain new dependencies, configurations get replaced and scheduled rules activate long after they are written.
What Buyers and Builders Should Take Away
Ryu draws a clear line: AI should narrow hypotheses and organize evidence, while humans declare the cause and approve action. He favors answers backed by citations, graph relationships and follow-up questions, and warns against polished output that hides doubt. "The worst pattern I see is uncertainty being covered by confident prose," he said.
For enterprises evaluating AI observability tools, the reporting suggests three tests: whether retrieved records are genuinely relevant, whether timing and dependencies support the proposed cause, and whether engineers can see what evidence would separate the leading explanation from alternatives.
For Korean AI startups and their investors, the bar is rising. Demonstrations should be tested against misleading recent changes, competing explanations and incomplete or outdated system data before a fluent answer is treated as a finding.
Source: KoreaTechDesk





