Take the two ends of the Brink's range as separate propositions. A 40% cut is a third larger than a 30% cut [14], which is a wide band to hand a treasurer, and the midpoint of 35% [15] is the figure that will get quoted anyway. What Brink's publishes is a ratio [1], and a ratio without the vault balance it multiplies prices nothing.
The more interesting version of the claim shows up when you set that saving beside H2O.ai's accuracy. If a machine's stock has to cover a forecast that lands within roughly 15% on average [3], and the new stock sits 35% below the old [1], then the old stock was carrying about 77% more cash than the machine actually needed, with the band running from 64% at a 30% cut to 92% at a 40% one [16]. That arithmetic borrows from two companies selling different products [1][3], so read it as a sizing exercise rather than a finding. What it says is that the recovered capital comes out of worst-case padding, and a model only has to beat a fixed delivery schedule to collect it. The seven-to-ten-day look-ahead Brink's describes [2] is doing less clever work than the phrase AI forecasting implies; what it buys is lead time.
The machine estate itself is close to fixed, and the Federal Reserve's August diary is why: cash is 14% of consumer payments [4] while more than 80% of consumers used it in the past thirty days and 90% expect to keep using it [5]. Revenue tracks the first number and the obligation to keep a machine loaded tracks the other two, so the money goes into forecasting the float rather than shrinking the fleet, which is the expensive way round.
This is probably wrong in one specific place. Average accuracy is the wrong statistic for a cash buffer, because a buffer is sized by the bad day rather than the mean day [3], and if the 15% average conceals a fat tail then the residual pad stays fat and the achievable cut sits at the bottom of the advertised range [1].
The comparison that sets the ceiling comes from the other end of the same trend: 70% of firms surveyed by PYMNTS Intelligence in October already run at least one AI tool for cash flow management [8], and those using agentic AI have automated as much as 95% of accounts receivable against 38% for firms that have not, a gap of 57 points and a ratio of 2.5 [9][17]. Receivables are paperwork. An ATM's contents are insured, counted and driven around in a truck, which is why the number on offer there is 30% to 40% and not 95% [1][9].
