AI costs are rising: not because AI is being used incorrectly, but because no one has decided what it is actually worth.
Last week, the CEO of Commonwealth Bank of Australia made a point that I have been expecting for some time: AI costs become harder to predict as companies try to automate more complex tasks. Reasoning models, agents using tools, longer contexts... they make token costs explode. This particularly affects organisations that budget for AI as if it were a software licence.
But I found his second point at least as interesting. He called it work slop: AI output that looks professional but adds little real value: summaries no one reads, unnecessary analyses that do not inform decisions, slides that already existed, ... You can multiply them exponentially if you want. It costs money, but adds nothing.
Many organisations still measure AI adoption in terms of usage: number of users, number of prompts, number of documents generated, ... But usage is not value, and more output is not the same as better output.
I think that is what the next phase of AI governance needs to address. Privacy, compliance and security are by now the obvious basics, but also: which tasks are actually worth automating? Who safeguards the quality of the output? When is human review not a luxury, but a requirement? And when do we simply stop pursuing a use case that delivers no value?
Do not treat AI as an unlimited productivity layer, but as an operational system with budgets, quality criteria and ownership. The biggest AI cost saving may not come from automating more, but from producing less useless output.
