Finance edition

Bloomberg published an investigation on 18 September into a deadly air strike on a school in Minab, Iran, on 28 February. 123 children died. According to the investigation, excessive trust in Palantir's Maven Smart System played a role. That system uses Anthropic's Claude to rank targets semi-automatically. US Central Command soldiers expected Maven to flag outdated or conflicting information. That didn't happen. Compiling target lists used to take hours. With the system, it took just a few minutes. Palantir says there's no evidence the software malfunctioned, and that the company isn't responsible for the underlying data. It has added new checks that should catch conflicting information. This isn't a finance story, and yet it is one. The pattern is exactly like yours: a system that delivers results faster than people can check them, users who assume the system itself will warn them about messy data, and a supplier who says the data isn't its problem. Write down today which AI output in your process nobody double-checks anymore, because the system supposedly catches it. The speed itself is the risk. When a list is ready in minutes instead of hours, the time in which someone might have spotted something odd disappears. At the same time, the result becomes more convincing, because it arrives neatly ordered and ranked. In finance it works the same way: an AI that matches invoices or flags discrepancies produces a tidy list, but it doesn't say whether its source is old or contradictory. What you're left checking isn't the sum, but whether the input is correct. The question of responsibility is sharpest for you. Palantir points to the customer's data, the customer trusted the software. That gap is exactly where things go wrong. For every AI feature in your financial systems, record who is liable for the quality of the input, and who signs off on the result. If that's written down nowhere, you'll end up responsible in the end.

Source →