Brazil spends US$75 billion a year on public medicines, but local governments buy them through 263 different systems. Because buyers often record purchases in free text without standard details like dosage, comparing prices across the country has been difficult. To fix this, a coalition of government and civil society groups built a platform aggregating two years of data from the national procurement portal. They found that 30% of the original procurement processes had insufficient or incorrect descriptions. To standardise the database, the team used a large language model to read the unstructured purchase records and match each item to a standard federal product code.
The language model successfully matched 80% of the procured items to the standard catalogue, up from 5% during the team’s initial review. It classified 100 items in five seconds, a task that took a human 90 minutes. Procurement officials are now using the resulting database to filter searches, establish reference prices, and plan purchases. While improving data quality can increase the number of items the system classifies, the article argues this approach demonstrates it is possible to make procurement comparable across highly decentralised markets.
Using artificial intelligence to match messy local records to a central taxonomy solves a fundamental scaling problem in public administration. The open question is whether this approach works for public services that lack a highly structured target catalogue. Without a rigid national index waiting to catch the outputs, language models might struggle to organise municipal data.
Today’s links: Assorted links for 1 October 2026.