Bloomberg has reported new details from officials involved in an unreleased Pentagon investigation into the February 28 U.S. missile strike on Shajarah Tayyebeh Elementary School in Minab, Iran. The strike killed more than 150 people, including at least 123 children. The reporting describes a chain of targeting and intelligence failures, including outdated imagery, gaps in underlying intelligence, cuts to civilian-harm review staffing and overreliance on Palantir’s Maven system.

The first point to keep clear is the status of the evidence. The Pentagon investigation has not been publicly released. Bloomberg’s account is based on officials involved in that internal review. That means the findings are important reporting, but they should be attributed to those officials rather than described as a published Pentagon report. Palantir has also said its software was not responsible for the underlying data or for identifying deficiencies in that intelligence.

The second point is what Maven did and did not do. Bloomberg describes Maven as part of an AI-assisted targeting workflow that helps combine and process information for military planners. The reporting does not establish that Maven autonomously chose the school as a target or authorised the strike. Instead, the reported failure involved people, data, staffing and process around the software. Some personnel relied on the system while working with information that was stale or incomplete and under compressed review timelines.

That distinction matters beyond this one case. Decision-support systems can make it faster to search, combine and rank information, but speed does not make the source data current or correct. If a workflow assumes that a tool has already checked for stale intelligence when it has not, the software can amplify a mistaken assumption rather than catch it. Human review only works as a safeguard when reviewers have enough time, authority and reliable information to challenge the output.

For organisations using AI in high-stakes settings, the practical lesson is about controls rather than slogans. Teams need clear records of where data came from, when it was last updated, what the system is designed to detect, and which decisions still require independent human verification. The more consequential the decision, the less safe it is to treat a confident-looking software output as proof that every underlying assumption has been checked.

What to watch next is concrete: whether the Pentagon publishes the investigation or responds formally to Bloomberg’s account, whether military targeting procedures or civilian-harm review practices change, and whether Palantir or the Pentagon provides more detail about Maven’s role in the workflow. Until then, the strongest conclusion supported by the reporting is not that an AI system independently made the strike decision. It is that officials involved in the review say overreliance on AI-assisted tools was one part of a broader cascade of preventable failures.