> ## Content Index > Fetch the complete content index at: https://nextwith.ai/llms.txt > Use this file to discover other available public pages before exploring further. # RBA warns AI spending may raise costs before productivity arrives - URL: https://nextwith.ai/rba-warns-ai-spending-may-raise-costs-before-productivity-arrives/ - Published: 2026-09-22T10:10:01.000Z - Updated: 2026-09-22T10:10:01.000Z - Description: Australia's central bank says AI investment can add to inflation before productivity gains appear. For companies, the practical response is to measure finished, quality-adjusted output alongside total human time and cost. - Author: NextWith.ai Editorial Desk - Tags: Business, Analysis Australia's central bank is drawing a line between spending on artificial intelligence today and productivity that may arrive later. At a September 22 event in Sydney, Reserve Bank of Australia governor Michele Bullock said the bank had not yet seen evidence that AI was making the economy more efficient, while data-center investment was adding to demand and inflation, according to [the Guardian's report from the event](https://www.theguardian.com/technology/2026/sep/22/ai-could-be-a-bubble-and-is-not-yet-making-australia-more-productive-michele-bullock-says?ref=nextwith.ai). Bullock did not declare that AI is a bubble. She said central banks are watching that risk and that she did not have a settled view. The distinction matters: the warning is about timing and uncertainty, rather than a forecast that AI investment will collapse. ## The near-term cost comes before the measured gain The RBA had already described the underlying mechanism in an [official August 11 media-conference transcript](https://www.rba.gov.au/speeches/2026/mc-gov-2026-08-11.html?ref=nextwith.ai). Bullock said data-center construction could draw scarce workers away from other building work in Australia. She also pointed to strong global demand for semiconductors and chips as another source of price pressure. In that account, AI can raise demand for labor, electricity, construction and hardware before businesses reorganize enough to produce more with the same resources. That is different from saying AI has no productive use. In a separate [official March 3 RBA transcript](https://www.rba.gov.au/speeches/2026/sp-gov-2026-03-03-q-and-a-transcript.html?ref=nextwith.ai), Bullock described the technology as potentially useful for automating lower-level work so employees can concentrate on analysis. She also cautioned that businesses need time to reorganize processes before the effect reaches the wider economy. The September remarks add a current checkpoint. The Guardian reports that Bullock cited research in which employees using AI produced the same output while working fewer hours. That may be valuable to employees and firms, but it does not automatically appear as higher measured output per worker if organizations keep their process and output unchanged. ## What an AI team should measure inside the business NextWith's assessment is that companies should separate three ledgers: adoption, operating cost and useful output. Adoption records licenses, active users or tasks that involve an AI tool. Operating cost records model fees, review time, infrastructure and rework. Useful output records the completed result that matters to the organization, with a quality threshold attached. Consider a hypothetical support team that introduces an assistant for drafting replies. Counting generated drafts would measure tool activity, not productivity. A more useful test would compare resolved cases of equivalent difficulty, customer-rated quality, corrections and total staff minutes. If the assistant shortens drafting but creates more review work, the time saving and the rework belong in the same result. The same logic applies to software development. Prompt volume, accepted suggestions and generated lines can explain how a tool is used. They do not establish that a release arrived faster or with fewer defects. A bounded trial can instead choose one recurring change, hold the acceptance test constant and record elapsed time, review effort and defects found before release. ## Do not hide the implementation gap A team may find that workers finish the same workload sooner but that management has not changed scheduling, service levels or the number of tasks attempted. That is still a real result, but it should be named accurately. It is available capacity or reduced effort, not yet evidence of higher organizational output. There can also be a temporary decline while people learn a tool, revise roles and discover where human review is needed. A short test should therefore record setup and training separately from steady use. Otherwise, an early comparison can make a useful tool look permanently expensive, while a promotional comparison can omit the adoption work and make it look immediately productive. Infrastructure decisions require another boundary. A supplier's expansion plan can signal confidence in future demand, but it does not prove that a particular customer's workload will become cheaper. Teams should compare a dated service offer with their actual workload, including quality, latency, support and total cost. Macroeconomic optimism and a procurement decision answer different questions. ## A six-week checkpoint is more useful than a slogan A small business can start with one workflow that occurs often enough to observe repeatedly. For the first week, measure the current process without changing it. Introduce the AI step with a clear review rule, then measure the same outcome for several weeks. Keep unusual cases rather than removing every failure from the sample. At the checkpoint, classify the result plainly. Continue if the same-quality output rises or total effort falls without a new safety or reliability problem. Revise if the benefit appears only in one step but moves work to reviewers. Stop if the trial cannot preserve the required quality, privacy or accountability. An inconclusive result is also legitimate when the sample is too small. The RBA's warning is useful because it makes the lag visible. Investment can be real, adoption can be widespread and productivity can still remain unproven at the economy-wide level. Companies do not need to wait for national statistics, but they do need measurements that connect a tool to a finished result. Pick one recurring workflow and record quality-adjusted output, total human time and total cost for six weeks before calling the AI rollout a productivity gain.