
On November 18, 2026, Strike48 released The State of Agentic Security H2 2026, a report based on a survey of 534 security and IT leaders. Strike48 sells an agentic orchestration platform for security and IT environments, so these are a vendor’s survey figures, self-reported by the respondents it reached. A security operations center is not a finance department or an HR desk, but it is a workflow with intake, triage, action, and sign-off. Those steps show up in every function a business owner is trying to hand to AI.
What Strike48 reported about agents in production
Strike48 reported that AI agents are in production at 74% of the organizations it surveyed, compared with 36% in its May survey. Deployment did not bring autonomy with it. Among teams with agents in production, 47% still have a human execute or pre-approve agent actions, and only 30% allow high or full autonomy.
Breadth appears to matter, though Strike48 did not claim that more use cases cause more trust. Among teams running agents across multiple production use cases, 14% cited distrust of output, compared with 42% of teams with a single use case in production. High or full autonomy was 41% for the multiple-use-case teams and 8% for the single-use-case teams. Teams that already trust their agents may simply be the ones expanding.
The barrier ranking moved. Lack of trust in AI outputs was the top barrier in the previous survey and now ranks fourth. Data infrastructure ranks first, with 31% of leaders saying they lack the infrastructure to support effective agents. On letting agents take action, regulatory and compliance risk was the top concern at 46%, followed by agents acting on incomplete or incorrect data at 41%.
What leaders say they need before an agent acts without approval
Before letting an agent act without approval, 57% of respondents want a track record of accuracy in their own environment and 53% want a full audit trail. Another 46% want technically enforced limits, 44% want a precisely defined scope, and 40% want the ability to reliably reverse actions. The CIO or CTO was the most frequently named approver for agents with authority to act in production, at 40%, compared with 33% who name the security leader.
On people, 87% of respondents are changing their L1 analyst function because of AI agents. Of the respondents, 49% are hiring for new skills, 42% are retraining existing L1 analysts to supervise or manage agents, and 37% are shifting L1 analysts into L2 or investigation roles. Meanwhile, 20% have reduced L1 headcount and 19% have slowed or frozen L1 hiring.
Strike48’s president, Tim Leehealey, framed the gap this way: “Putting AI agents into production is one thing. Trusting them to make decisions and take action is another.”
What the release does not say
The release reports no error rates, no cost figures, and no time saved, and it does not list which tasks the agents perform. It also does not say how long a track record must run before autonomy is granted. A business owner cannot copy a 74% deployment figure into a board slide as proof of results.
Role assignment comes before the tooling decision
Basil’s view is that deploying AI into a workflow is a role-assignment question before it is a tooling question. He splits AI work into seven roles, Researcher, Editor, Coder, Calculator, Liaison, Ideator, and Navigator, with a human holding final authority, a method he calls HAIA-RECCLIN. Applied to any workflow in this survey’s pattern, the first job is to write down which role each step needs and who or what holds it.
In a triage workflow, an agent can plausibly take the Researcher role, gathering and correlating context, and the Calculator role, scoring and counting. The Navigator role, which lays out the options and the trade-offs without resolving them, still needs a person when the decision carries consequence. That is Basil’s reading, not Strike48’s. Strike48 describes its platform as using deterministic and cognitive agents with configurable human oversight; it does not map its product to roles, and Basil is not claiming it does.
The 42% retraining L1 analysts to supervise agents is the part of the data Basil finds most useful for a small or midsize business. Supervising an agent is a different job from doing the task, and it is a skill. A person who could not have done the work has trouble catching the agent’s error in it. Basil’s authorship test applies: a reviewer who cannot defend the output line by line to a skeptical expert forwarded it and did not review it.
Where machine checks end and a named person begins
Several items on the leaders’ wish list are machine-side controls: an audit trail, technically enforced limits, a defined scope, and reversal capability. Basil classes platform monitoring, evaluations, and models checking models as Responsible AI, which is machine checking machine. It has real value, and Basil trusts automation in contained work where events can be undone. It is not AI Governance. AI Governance is a named human who exercises a checkpoint and accepts, modifies, or rejects the output before it moves downstream.
That distinction sharpens one finding. The CIO or CTO is the most frequently named approver for agents with authority to act in production. Naming who approves an agent’s authority is not the same as naming who accepts or rejects an individual output. A business owner should ask for both names: who allowed the agent to act, and who answers for what it did this time.
Leehealey also said security teams cannot afford humans approving every move an agent makes. Basil agrees on the practice. Pre-approving every action turns the reviewer into a rubber stamp, and a stamp is not a checkpoint. The answer is placement. Put the named human where the action is hard to reverse or the exposure is high, and let reversible, low-exposure steps run under machine checks with sampled human review.
Steps and KPIs for an operations lead
These are Basil’s recommendations, not findings from Strike48’s report.
- Write the role table. For each step in one workflow, list the role (Researcher, Editor, Coder, Calculator, Liaison, Ideator, or Navigator), whether an agent or a person holds it, and the named person who accepts, modifies, or rejects the output at the end.
- Sort steps by reversibility. Mark each step reversible or hard to reverse. Hard-to-reverse steps get a named human checkpoint before the output moves on.
- Earn autonomy with a record. Start with human pre-approval. Move a step to sampled review only after a documented track record of accuracy in your own environment, which is the same condition 57% of Strike48’s respondents named.
- Check the data first. Strike48’s respondents ranked data infrastructure as the top barrier. Before an agent touches a function, confirm it can see the records that function depends on.
- Track the reviewer, not only the agent. Measure the share of outputs the reviewer modifies or rejects, review turnaround time, the share of agent actions with a complete audit record, and the reversal rate. Basil’s recommended target for audit-record coverage is 100% on anything that touches money, access, or customers.
- Watch for rubber-stamping. Basil’s proposed and untested trigger: if reviewers accept more than 95% of outputs for three straight review cycles, pull a sample and audit it.
The boundary this survey leaves open
Strike48’s respondents want a track record before they grant autonomy, but the release does not say how much record is enough, who certifies it, or what happens to accountability once a team crosses into high or full autonomy. Only 30% of teams with agents in production have crossed that line. Basil does not think a percentage of accurate outputs settles it. Where the output is hard to undo, a person’s name has to stay on it, and where a business owner decides a record is enough to remove that name is the decision no survey answers.
Sources
- – Puglisi, B. C. (2026). HAIA-RECCLIN. basilpuglisi.com. https://basilpuglisi.com/haia-recclin/
- – Strike48 Inc. (2026, November 18). Strike48 survey finds AI agent adoption in the SOC doubled in five months, but autonomy lags [Press release]. GlobeNewswire. https://globenewswire.com/news-release/2026/11/18/3405673/0/en/strike48-survey-finds-ai-agent-adoption-in-the-soc-doubled-in-five-months-but-autonomy-lags.html
FAQ
Is the 74% figure proof that AI agents work in operations?
No. It counts how many surveyed organizations have agents in production, as reported to a vendor. The release summary gives no accuracy, cost, or time-saved results, so it shows adoption and not outcomes.
What does a named human checkpoint add if there is already an audit trail?
An audit trail records what happened. A named person accepting, modifying, or rejecting the output decides whether it moves on. Basil treats the record as a Responsible AI control and the named decision as AI Governance.
Does every agent action need human approval?
Basil does not recommend it. Approving everything turns review into a rubber stamp. Place the named person where an action is hard to reverse or the exposure is high, and sample the rest.
Can a small business use this when the survey covers security teams?
Yes, as a pattern. Any function with intake, triage, action, and sign-off can use the same role table, reversibility sort, and reviewer KPIs, though the numbers come from security and IT leaders and should not be applied as benchmarks to finance or HR.
Which roles in a workflow can an agent take?
That depends on the step. Basil’s frame lists seven roles, and an agent can take the research, calculation, and first-draft work in many workflows. A person keeps final authority when consequences are hard to undo.
#AIg
Leave a Reply