An AI-Powered Text Analytics Tool for the Reserve Bank of Australia's Business Liaison Program

Since 2001 the Reserve Bank of Australia's liaison program has sent staff to interview around 22,000 firms, industry bodies and community organisations, building a 25-year archive of qualitative notes on wages, prices, investment and business conditions used to inform monetary policy. Manually searching this large, free-text archive limited how systematically the RBA could draw signals from it.

RBA Research Discussion Paper 2025-06 (Gray, Lattimore, McLoughlin and Windsor, August 2025) describes building an NLP system to search the full liaison archive on demand, classify the topic and tone of each note, and extract precise numerical figures — such as a firm's self-reported wage or price growth — directly from free text.

The team validated the tool's extractions against human-coded benchmarks, then tested whether the resulting liaison-based indicators improved wage-growth nowcasting models when added to best-practice machine-learning and Phillips-curve approaches. The RBA published its code openly so other central banks could adapt the approach.

Adding the AI-derived liaison signals to wage-growth nowcasting models significantly reduced nowcasting errors by meaningful magnitudes compared with baseline approaches. As of the paper's release, no other central bank had yet reported adopting the released code.

The authors stress that despite strong quantitative performance, it will always remain critical for economists using the tool to interrogate, validate and apply judgement to its outputs rather than treat them as ground truth. This is a single institution's internal research evaluation, not an externally replicated trial.