The two-hour demo is the detail worth sitting with. Twelve hundred dollars of tokens went out the door in a single session [4], which is about 80% of what one engineer is now permitted for an entire month [2]. Nothing in that number was visible to anyone until the invoice arrived. The same arithmetic explains the headline: consuming a twelve-month budget in four months is a run rate three times the plan [5], and it took a quarter of the year to notice.
That is the governance failure, and it is separate from the value question. Deloitte's 2026 enterprise survey puts 74% of organisations hoping AI will grow revenue against 20% doing it today, a gap of 54 points [4]. Meanwhile more than a third are already banking efficiency gains [9]. Efficiency is the return most finance teams can actually document, and it is also the return a metered bill eats first, because savings and token spend land on the same side of the ledger.
Commonwealth Bank's roughly $2.4bn a year on technology and capability [10] is, on its own, a number that settles no argument. What makes it defensible is the line items underneath it: a generative tool called Compass that has handled over 500,000 banker queries and lets frontline staff answer complex questions about three times faster [12]. Chief financial officer Alan Docherty's stated logic is that productivity and better customer outcomes feed top-line revenue, which is what justifies the next round of investment [13]. That is a sentence a CFO can repeat in February and be held to in August.
There is a second test that has nothing to do with dollars. Harrison.ai trained its radiology models on more than a million clinical studies and 550 million expert annotations from over 140 radiologists [14], and independent testing has shown its chest X-ray tool lifting diagnostic accuracy by 45% [15], with more than half of Australia's radiologists now having access [16]. The asset there is the annotation corpus, not the model. Rob Versaw's argument is that an organisation whose AI reasons from the open web instead of its own context pays premium prices for generic output a rival can buy just as cheaply [19]. Worth noting where the argument comes from: Versaw works in product strategy at Dynatrace [18], an observability vendor, so per-workflow instrumentation of spend is a conclusion that suits his employer. It does not make the Uber figures less real.
The gap in all of this material is a denominator. There are caps, totals, monthly ceilings and query counts, and not one cost per merged change, resolved ticket or shipped feature [6]. Until an engineering organisation publishes that rate, "per-token value" stays an argument about which workflows get named, not a measurement anyone can check.
