AI on open data · Department of Finance office occupancy reports
Ten PDFs in. One open dataset out.
The decade, chart by chart
Every point is a number printed in one of the reports. Click a point, or any underlined figure on this page, to see the page it came from. Breaks in method are marked on the charts; the data detective explains each one.
Signals found automatically
Generated by rules over the extracted series (largest changes, records, turning points). Each sentence is templated from the data and cites the pages it used.
Places
States and territories, remoteness, lease expiry and how space is distributed across density bands. Pick a measure and a year: only the years a report published that breakdown are offered.
Entities
The largest portfolios over time, followed through renames and machinery-of-government changes, and the full entity-by-entity tables from the editions that published them.
What-if
Arithmetic on the published aggregates, recalculated as you move the sliders. Illustrative only: it shows what the published totals imply under a different benchmark, not what any change would cost or save.
Data detective
Reading ten reports side by side shows things no single report says: figures that were later revised, definitions that changed, and printed numbers that don't reconcile. The checks below were run automatically on the extracted data.
Automated consistency checks
How the method changed
Definitions and scope, in the reports' own words.
Restatements: the same year, different numbers
Ask the data
Type a question in plain English. This page answers offline with a built-in parser that calls the same tools the Desk Check connector gives Claude, and shows the calls it made. Answers cite their pages.
For AI and developers
The dataset is published so that other software, including other AI, can use it directly: a hosted MCP connector you can add to Claude by pasting one address, tidy CSV and JSON with a Frictionless data package, an llms.txt guide, and a context pack sized for a model's context window.