If you're about to make a billion calls per visa card issuer then you can afford a small amount of engineering effort to break that specific captcha. Most individual captchas are pretty rudimentary and can be attacked with simple image processing techniques and pattern matching.
This one in particular is simple enough that nearly any technique you wanted to throw at it would succeed with minimal fine tuning. I'd be shocked if it took an afternoon even if you'd never broken a captcha or done any image processing before, and that's without borrowing an off-the-shelf ML solution.
Plus, even if you had a 50% failure rate you'd just need twice as many calls. That's not trivial, but it doesn't really affect the viability of the idea.
This one in particular is simple enough that nearly any technique you wanted to throw at it would succeed with minimal fine tuning. I'd be shocked if it took an afternoon even if you'd never broken a captcha or done any image processing before, and that's without borrowing an off-the-shelf ML solution.
Plus, even if you had a 50% failure rate you'd just need twice as many calls. That's not trivial, but it doesn't really affect the viability of the idea.