There’s a line from the Apple Card issuer announcement that I keep coming back to: a promise to keep delivering “a best-in-class experience and exceptional customer service with Apple Card,” paired with the assurance that existing users would not see any visible change. That second part is the whole ballgame. Somebody stood up in front of a press release and committed to the idea that swapping the bank underneath millions of active credit card accounts would feel like nothing at all.
As someone who spends his weeks pulling apart AI toolkits and watching them break during far smaller transitions, I find that claim both admirable and slightly outrageous.
Twelve million accounts and nobody is supposed to notice
The numbers matter here. Apple Card had 3.1 million users in the US by March 2020, and roughly 12 million by early 2024. That’s not a pilot program. That’s a customer base with statements, disputes, autopay schedules, credit reporting histories, and a decade of muscle memory about where the buttons live.
In 2026 that whole thing moved from Goldman Sachs to JPMorgan Chase, with a 24-month transition period attached. Two years of runway to move an operation that big, and the stated goal is that the people using it every day feel nothing.
Compare that to the last time an AI tool you depend on changed something underneath you. A model got deprecated with a few weeks of notice. An API version bumped and the output format shifted just enough to break your parsing. A pricing page got “simplified” and your monthly bill doubled. A vendor got acquired and the roadmap quietly evaporated, along with the support channel you actually used.
The AI tooling space treats migration as an event the user is expected to absorb. Apple and its banking partners treated it as work the provider is expected to eat.
Why the 24 months is the actual headline
Every tool I review has some version of a transition plan, and almost all of them are aspirational. The pattern goes like this: ship fast, get users, defer the hard integration work, then discover that the hard integration work is the product.
A 24-month transition window is a confession that the unglamorous parts take real time. Account records, regulatory filings, servicing systems, statement generation, dispute handling. None of it makes a good demo. All of it determines whether the thing works.
When I score an AI toolkit, I’ve started weighing this kind of thing more heavily than feature lists:
- How much notice do you get before a model or endpoint goes away, and is that commitment written down anywhere binding?
- Can you export your data, your prompts, your evaluation sets, and your logs in a format another tool can actually read?
- When the vendor changes its underlying provider, does your integration code change too?
- If the company gets acquired, what happens to your account and your history?
- Does the pricing model survive a change of ownership, or does it reset to whatever the new owner prefers?
Most tools fail three or four of those. Some fail all five and make up for it with a very nice dashboard.
Boring is a feature, and it’s expensive
The reason this story is interesting fifteen years into Apple’s history with cards, going back to the iCards service from the late 2000s, is that it shows what continuity actually costs. Years of negotiations with various lenders before a deal landed. Two years of transition work after. An enormous amount of effort spent so that the product on your phone stays the same.
That’s a tradeoff most AI tool vendors haven’t had to make yet, because most of them are too young to have accumulated the kind of user base that makes breakage expensive. But that’s changing. The teams building on these toolkits now have production systems, compliance requirements, and internal users who will absolutely file a ticket when the output format shifts.
The vendors that survive the next few years will be the ones who figure out how to change their guts without changing their surface. Not because it’s a nice engineering value, but because their customers will start pricing instability into purchasing decisions.
What I’d actually ask a vendor
My new favorite question in a demo call is simple: tell me about the last time you changed something underneath your users, and what they had to do about it. The answer is more revealing than any benchmark. Teams that have done a clean migration talk about it with specifics. Teams that haven’t get vague fast.
A credit card changing banks without anybody noticing is not exciting. It’s also a standard almost nothing in AI tooling currently meets, and one worth borrowing.
🕒 Published: