What disappears from human review when AI clears the queue?
How security leaders can evaluate the cases that automation removes from an analyst’s view
Karina Carlos
A shorter security-alert queue can represent useful work. Analysts may have more time for cases that require investigation. But the queue shows only what remains available for review. If an automated system closes part of the incoming work, evaluating the analyst’s screen alone leaves part of the process unexamined.
The question for a security leader concerns the cases that no longer reach that screen. What evidence would show that their exclusion remains justified? The answer requires a view of the selection process as well as the decisions made after selection.
The benefit and the unresolved question
James Bono’s 2025 Microsoft-authored preprint randomized a phishing-triage task among 167 analysts. A modeled workflow automatically resolving emails classified as benign yielded an estimated 6.53 times the manual baseline’s true positives per analyst-minute. In a separate full-queue comparison, the AI-aware group missed 46% of injected false negatives, versus 17% in control. These are different analyses, not a demonstrated production-wide efficiency gain.
The results make queue design worth examining. An organization could gain useful capacity by removing low-value work while still needing evidence about the errors in that removal. Requiring people to recheck every item would create a different workflow, with its own costs and possible errors.
In a separate 2026 mixed-methods observational study, partly based on self-reports, Jack Tilbury and Stephen V. Flowerday report a relationship between attention shifting toward manual tasks and reduced monitoring of automation. I rely here on the publisher’s abstract and structured sections; the full results and instruments were not available for audit. The evidence does not establish causation.
Measure what entered before celebrating what remains
Suppose a team reports that it handled its visible queue faster after introducing automatic closure. That observation leaves several possibilities open. The same work may have been completed more efficiently, fewer cases may have required investigation, or some cases may simply have stopped reaching analysts.
I would ask the team to describe the population entering the process and account for where those items went. How many were retained for review, closed automatically, combined with another case or excluded before a case was created? The categories and counting unit need to be consistent enough to compare across periods.
This matters particularly when a dashboard reports a success rate only for the surviving queue. Performance on those selected cases cannot establish the quality of the excluded cases. A decision about the system’s scope needs evidence from both sides of that selection.
Preserve a case that can be reconstructed
An automatically closed item should leave enough of a record for an authorized reviewer to reconstruct why it left the queue. The organization should define that record before deciding that the item no longer needs attention.
My proposed review would examine whether the relevant input, disposition, timing and applicable system configuration can be connected. It would also distinguish an item judged benign from an item never processed because required information was missing. Combining those states under one closure count would obscure different problems.
Retention should be proportionate to the risk and applicable privacy and security requirements. The aim is to preserve usable evidence under controlled access, with defined limits. Keeping every piece of sensitive content indefinitely would create additional exposure without necessarily making the decision easier to reconstruct.
Inspect excluded work without recreating the whole queue
Sampling automatically closed cases is one proposal worth testing. Its design should follow the question being asked. A sample chosen to investigate a suspected weakness serves a different purpose from a sample intended to estimate error across the entire excluded population.
A team should explain how cases were selected, which periods or case types are absent, and how uncertain the resulting estimate remains. Finding no error in a small review cannot establish that the closed population is safe. A changed sampling method can also change the reported rate without any change in system quality.
The review needs a credible basis for deciding whether an exclusion was justified. Repeating the original automated label provides no independent check. Reviewers need appropriate evidence and expertise, and disagreements should remain visible rather than being resolved solely by deference to the model. These are design proposals, not a validated sampling protocol.
Let later evidence reach earlier closures
The classification of a case may deserve reconsideration when new information arrives. I would test whether later evidence can be connected to an earlier closure and whether the original record can be recovered without reconstructing it from memory.
The same question applies when the operating context changes. A new data source, altered configuration or different mix of incoming cases may change what an earlier evaluation can support. The team should define which changes require renewed scrutiny and who evaluates the resulting evidence.
This keeps the history of the decision available. A correction should show what was previously concluded and why the judgment changed, so that the organization can learn from the exclusion rather than simply replacing its label.
Include the work of verification in the performance account
NIST SP 800-61 Revision 3 recommends filtering events for human analysis and tuning monitoring to acceptable false-positive and false-negative levels. It also addresses roles and performance measures. This is official guidance, not proof of an AI-review design or a universal legal requirement.
A local evaluation should count the work required to maintain those conditions. Sampling, resolving review disagreements, investigating later evidence and maintaining records consume capacity. If that work is omitted, the apparent time saved can conceal a new obligation assigned to nobody.
Applying Human Authority to this context, I would ask whether the organization can examine the judgments embedded in removing cases from attention. The relevant capacity belongs partly to the review process around the queue, rather than only to the analyst working through it.
A smaller queue still needs an account
An organization need not expose every automated action to manual review to take responsibility for its operation. It does need a defensible account of what has been removed, how that choice is evaluated and what could cause it to be reconsidered.
The strongest evidence for expanding automation would therefore include both the capacity gained and the limits of visibility that remain. A queue can become smaller while the organization’s understanding becomes better. Establishing that improvement requires examining the work that the queue no longer shows.
Comments