On February 28, two Tomahawk missiles struck Shajarah Tayyebeh Elementary School in the southern Iranian town of Minab on the opening day of a planned overwhelming initial assault. More than 150 people died; at least 123 were children. A Pentagon investigation, reported by Bloomberg in September, attributed the strike to three compounding failures — bad intelligence, institutional attrition, and what the report calls overreliance on Palantir’s Maven Smart System.

Maven is an AI-powered data integration and targeting platform. It was built to accelerate intelligence analysis — to take hours of work and compress it into minutes. That is what it did. It processed the Minab target, flagged it as a valid strike candidate based on U.S. database classifications, and recommended it as a first-day priority. U.S. databases still listed the site as an Iranian Revolutionary Guard facility, a classification that dated to 2003.

The building had been a school since around 2017.

An analyst had noted the discrepancy in a 2019 observation — logged the concern after reviewing commercial satellite imagery showing the site conversion — but that note lived in a disconnected system that was never linked to targeting records. Maven had no access to it. More importantly: Maven was never designed to check whether its target database was current, or to flag potential conflicts with information stored elsewhere. The operators who relied on it expected it to do those things. That expectation is what the investigation calls overreliance.

This failure mode deserves to be named precisely. The system performed correctly within the scope it was designed for. The problem was that operators attributed capabilities to it that it didn’t have, and the institutional processes that existed to catch this kind of error had been hollowed out. Centcom’s civilian harm mitigation team shrank from 10 people to one over recent years — part of a broader 90% reduction in DoD civilian harm staffing. No member of that team reviewed the Minab site before the missiles launched. The Trump administration’s demand for an overwhelming opening assault further compressed the confirmation window.

Three failures, none of which is straightforwardly an “AI failure”:

One: outdated data. A database entry from 2003 survived into 2026, attached to a building that changed purpose nearly a decade ago. Data hygiene is not glamorous work; it also isn’t optional when the database drives targeting.

Two: system capability misattribution. Maven can process intelligence at speed. It cannot validate the provenance or currency of that intelligence. Those are different functions, and confusing them in a high-stakes context is not a design flaw in the system — it’s a training and doctrine failure.

Three: missing oversight. The people whose job was to catch exactly this kind of error were not present, because their positions had been eliminated by policy.

The accountability question that follows is uncomfortable: if the system worked as designed, and the doctrine was wrong, and the oversight positions were eliminated by policy decision — who is responsible? The answers are not primarily technical. They involve procurement contracts, military doctrine, congressional oversight, and casualty standards that are set well upstream of any individual AI deployment decision.

That is precisely why this story matters for engineers who build and deploy AI systems in high-stakes contexts. The Minab case isn’t an argument against AI-assisted analysis — it’s an illustration of what happens when capability deployment outruns governance. The gap between what a system can do and what operators believe it can do is not stable: it widens under time pressure, in novel environments, and when the oversight structures designed to close it have been stripped away.

Building a targeting system is an act with consequences that extend far beyond the code. The question of who verifies the database, who reviews the edge cases, and who has the authority to pause a recommendation — those are design decisions too, even when they look like organizational ones.