In 2011, I built something at MasterCard that feels remarkably relevant to the agentic SOC conversation happening today. It was a Fraud Rules Manager platform that let fraud analysts - internal and external - write AI and machine learning rules in plain human language to respond to new flash-fraud attacks in near real time. No engineering ticket, no deploy cycle. A banking analyst saw a novel attack pattern and could put a defense in front of it in minutes.
I tell that story because the industry tends to talk about agentic AI as if we have never handed consequential decisions to machines before. We have. I have been doing it for fifteen years — first building one of the industry's earliest AI fraud detection systems, later deploying genAI security agents at Intuit that take real work off real engineers. The hard problems of the agentic SOC are not new problems. They are old problems wearing a new and far more capable coat.
So let me share what those years actually taught me, because the lessons transfer directly.
Lesson one: the bottleneck was never detection
The bottleneck was never detection. It was the human in the loop being asked to do something machines do better. That 2011 platform did not succeed because the AI was brilliant. It succeeded because it moved the decision to where the expertise lived and removed the latency between knowing and acting. The agentic SOC's real promise is the same - not smarter detection, but the collapse of the distance between signal and response. If your agentic deployment is not shortening that distance, you have bought a more expensive dashboard.
Lesson two: machines fail differently than people
Machines fail differently than people, and your controls have to account for the difference. A human analyst makes a handful of mistakes a shift, usually independent ones. An agent makes the same mistake ten thousand times in a minute, confidently and at scale, before anyone notices. I learned this in fraud, where a miscalibrated model does not make a small error - it makes a systemic one across every transaction it touches. The control for human error is review. The control for machine error is the circuit breaker: hard limits, rate caps, and automatic kill switches that trigger on anomalous agent behavior, not just anomalous attacker behavior. Most teams instrument the threat. Few instrument the defender.
Lesson three: the language barrier matters more than the model
The language barrier matters more than the model. The reason that the fraud platform worked was that it let domain experts speak to the system in their own terms, their own language. Two decades later, large language models have made that interface extraordinary — but the principle is unchanged. The teams getting value from agentic AI are the ones whose practitioners can interrogate, correct, and steer the agent in natural language, not the ones who treat it as a black box that emits verdicts. Explainability is not a compliance checkbox. It is the mechanism by which your experts keep the machine honest.
Lesson four: you cannot bolt AI onto a fragmented signal
You cannot bolt AI onto a fragmented signal and expect coherence. This is where cloud, SaaS, and identity break a lot of agentic ambitions. An agent reasoning over partial telemetry produces confident, well-articulated, wrong conclusions – which is worse than no conclusion at all. At Intuit, protecting tens of millions of businesses across Mailchimp, QuickBooks, Credit Karma, TurboTax, and our money-movement products meant the unglamorous work of unifying signal had to come first. I integrated more than a dozen acquisitions, including the $12B Mailchimp deal, into a coherent security framework precisely because fragmented visibility is where attackers - and now, hallucinating agents – thrive. The agent is only as good as the signal it reasons over. Unify first, automate second.
Lesson five: measure the outcome, not the activity
Measure the outcome, not the activity. When I reduced QuickBooks spam fraud by more than 98 percent, the metric that mattered to the business was not how many messages the models scored. It was deliverability restored, revenue protected, and customer friction reduced. When we cut fraud 40 percent year over year, we did it while making the experience better, not by adding security speed bumps. Hold your agentic SOC to that same bar. Faster triage is an input. Reduced material risk and improved resilience are the output. Confusing the two is how organizations spend heavily on AI and end up exactly as exposed as they started.
I am genuinely optimistic about where this is going — more optimistic than most, because I have watched this pattern pay off across fraud, payments, and cloud security for fifteen years. The agentic SOC will be one of the most consequential shifts our field has seen. But it will reward the teams who treat it as the next chapter of a discipline they already understand, not the teams who treat it as a clean break from everything that came before.
The machines have been making security decisions for a long time. The agentic SOC just gives them more reach. Govern that reach with the lessons we have already paid for, and you will move faster and safer than the teams learning them for the first time.
FAQ
How do AI agents fail differently than human analysts?
A human analyst makes a handful of mistakes a shift, usually independent ones. An agent makes the same mistake ten thousand times in a minute, confidently and at scale, before anyone notices.
Why does an agentic SOC need circuit breakers?
The control for human error is review. The control for machine error is the circuit breaker: hard limits, rate caps, and automatic kill switches that trigger on anomalous agent behavior, not just anomalous attacker behavior.
Does agentic AI work without unified telemetry?
No — an agent reasoning over partial telemetry produces confident, well-articulated, wrong conclusions, which is worse than no conclusion at all. Unify first, automate second.