We built the Tech Council of Australia an AI assistant that keeps their members' data in their own hands

The Tech Council of Australia is the peak body for Australian tech, they shape Australia’s digital future and work to unite the tech sector, drive jobs, growth and investment for all Australians. TCA sits between the private sector and government as a neutral party, so that companies with something to say to Canberra have a credible platform to say it from, much of its value to members sits in what it has written: submissions to government, position papers, and research built up over years.

Accessing that valuable policy data used to mean guessing. Members searched a list of document titles, picked one that looked relevant, downloaded the PDF, and read it to find out whether it was. Anything the organisation had said that was not spelled out in a title was invisible, and the categories and tags that helped narrow the list existed only because a staff member had applied them by hand.

Cloudflare AI Search changed the economics of fixing that. Pointed at the R2 bucket already holding the documents, it produced a working index within half an hour and handled the PDF and DOCX parsing that would otherwise have been days of thankless work. The first demo we ran for TCA asked what the sentiment around AI was across their own submissions. It answered in a couple of minutes, streaming back a summary drawn from their published positions and citing the specific documents it came from, TCA's own archive answering questions about itself, sources included.

Members now ask a question in plain English and get an answer drawn from TCA's own material, with links back to the documents it came from. A member wanting the organisation's position on AI regulation gets an answer just as quickly, sourced, instead of working through a reading list. Nobody at TCA has to tag another document for that to keep working.

Two constraints shaped how we built it. Cloudflare ships a set of web components for this, and they assume a public endpoint, which was never going to work for member-gated material. We used their source as a reference and built a version that sits behind TCA's own authenticated APIs. Getting citations to resolve correctly took system prompt work, because the index knows documents by their storage keys while members reach them through the app's document API.

The second constraint was AI data governance. Queries and documents never reach a frontier model. Retrieval and generation run on Cloudflare's open-weight models inside the same account as the rest of the application.

Thinkmill is a member of the Tech Council of Australia.