By James Proctor, Co-Founder and Managing Director, The Inteq Group
Blast radius, in AI decision-making, is the scope, cost, and reversibility of the consequences if a decision is wrong. It is the measure that should determine how much authority an AI agent receives over any class of decisions, and it is conspicuously absent from most AI governance frameworks, which organize themselves around model characteristics, data sensitivity, and task complexity instead. Those dimensions matter, but none of them answers the question an operating executive actually needs answered: if this specific decision goes wrong, how far does the damage travel, what does it cost, and can we take it back?
What Are the Dimensions of Blast Radius?
Four dimensions do the analytical work, and none requires technical depth to assess. Reversibility: can the decision be undone, at what cost, and within what window? Propagation: how far do the effects travel before stopping, one transaction, one relationship, one regulatory posture, one public commitment? Detection speed: how quickly would an error surface, immediately, at month-end reconciliation, or at an audit two years out? And concentration: does the decision touch one case or reprice an entire portfolio at once? A decision that is reversible, contained, quickly detected, and case-level has a small blast radius no matter how sophisticated the reasoning behind it. A decision weak on any of the four deserves scrutiny regardless of how routine it looks.
Why Is Complexity the Wrong Proxy for Risk?
Because complexity measures how hard a decision is to make, and blast radius measures how much it matters, and the two are nearly uncorrelated in real operations. Here is the bias I see in risk register after risk register, and I will name it plainly: organizations fear the smart decisions and ignore the dangerous simple ones. The multi-variable pricing analysis gets three layers of oversight because it feels intelligent, while the mundane payment release, a decision a clerk executes hundreds of times daily, runs with minimal scrutiny because it feels trivial. Then the loss event arrives, and it almost never comes from the clever analysis. It comes from something simple, high-volume, and irreversible that nobody classified because nobody was impressed by it. If your AI risk register is sorted by sophistication, it is sorted by the wrong column.
“Complexity measures how hard a decision is. Blast radius measures how much it matters.”
How Do You Assess Blast Radius in Practice?
Commodity trading operations show what mature consequence classification looks like, because trading is one of the few functions that learned this discipline the expensive way. A trading floor does not govern activity by intellectual difficulty. The most analytically demanding work, market analysis and strategy formation, runs with wide freedom, while the simplest instructions, a settlement detail, a limit change, a wire release, sit inside hard controls, because a fat-fingered settlement instruction can lose in one afternoon what a bad analysis loses in a quarter. Position limits, delegation matrices, and four-eyes rules are blast-radius governance, refined over decades of incidents. The AI translation is direct: classify each decision class in a process against the four dimensions, in a working session with the people who own the consequences, and let that classification, not the impressiveness of the technology, set the autonomy boundary.
Running consequence-classification workshops, and converting the results into defensible autonomy designs, is core to our agentic AI consulting services. The classification typically takes days and reorders the risk conversation permanently.
Your organization already knows how to govern by consequence. It does it wherever money moves. Agentic AI simply requires extending that discipline to every decision an agent might touch, before the loss event teaches it the way trading floors learned.
