Palantir Technologies

Contribution to AI

Intelligence work after 2001 failed less from a shortage of data than from data sitting in incompatible systems that no analyst could query together. The answer was to make integration itself the product: pull in intercepts, informant reports, travel records and transactions, resolve the entities that recur across them, and hand the analyst objects and links rather than tables — an ontology, in the company's usage, later stretched to cover factories, cargo and customer orders. Two consequences outlived any particular contract. One was a way of selling software: engineers were posted inside the customer's building to rebuild the tool around the work they actually found, and the forward deployed engineer became a job title used across the industry, borrowed openly by AI firms now selling models into large institutions. The other was architectural. Permissioning and logging were pushed down into the data layer, so that every query was attributable and every field restricted by role and stated purpose, which put a concrete engineering proposal behind the claim that surveillance tooling can be bounded from within. Whether the bound holds has been contested without pause — over immigration enforcement work, predictive policing trials, and pandemic health-data agreements in Britain and the United States. That argument is itself the measure of the contribution: a great deal of state and industrial decision-making now runs through one vendor's account of what an organisation's data means.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).