What's worth isolating here is who holds the estimate, not migration itself. For most of the past decade a mid-sized buyer could not afford to find out what leaving actually cost, which meant the incumbent's account of that cost was the only number in the room. What the column describes [1] is a fall in the price of the estimate, which is not the same thing as proof that the price of the move has fallen. Leverage moves the moment the buyer can produce an independent figure, whether or not the figure is precise.
A coding agent's scope of a core system migration is worth roughly what you paid for it, a real limit worth naming. But the negotiation does not require the scope to be right. It requires it to be defensible enough that the vendor cannot price on the assumption that it does not exist. The other half of the column's buyer advice [2] is a cleaner test, because it returns a binary answer: either your data comes back in an open format or it does not, and whichever way it lands tells you what the renewal is worth before you argue about the discount.
Then there is the clock. The founder quoted in the piece put a pre-AI core banking exit at dedicated teams working through 18 to 24 months or more [4], which is 1.5 to 2 years [1]. A buyer who signs a three-year renewal this quarter and points an agent at the stack the following morning is still not credibly able to walk until well inside the next term. That makes term length, not headline price, the variable this quarter's signature actually decides.
The column's evidence for the AI side is capability rather than outcome: AI can write COBOL, and IBM sells a code assistant pitched at translating mainframe COBOL into Java [6]. What's missing is a completed migration on a shortened clock; the record shows only capability [3]. We do not know yet whether 18 to 24 months becomes nine or fifteen, and the piece is a single contributor's argument built on one founder's testimony [12], including his description of core banking as very sticky and very hard to move away from [7]. Treat the direction as plausible and the magnitude as unmeasured.
It also helps to notice which buyers this describes. The Reuters count the column cites has COBOL, a 1959 language, underpinning 43% of banking systems [5], which by the same count leaves 57% that are not [2]. Most renewal fights are CRM-shaped rather than mainframe-shaped, and those were always the more movable end. The founder's own surprise at how many manual tasks a client still performed [10] points at process debt rather than code depth, and process debt does not migrate on a compiler.
Which is why the second half of the argument matters more than the first. The alternative to a bad renewal is not building your own core system, since the author expects software costs to stay well above zero and buyers to keep wanting a vendor [9]. It is a smaller vendor willing to absorb the migration as part of the deal [3]. An incumbent facing a costed exit and a challenger quoting the switching work is negotiating against two numbers instead of none, and the length of the term you sign now determines which quarter you get to use them.
