Parametric insurance has been quietly operating for years with no human in the loop. A trigger event occurs, the system validates it against predefined conditions, and a payment goes, with no approval queue or manual review. Across the industry, billions of dollars move this way, and insurers have built entire business models around it. It works because the rules are explicit, the conditions are defined in advance, and the system executes them exactly as written.
Nobody calls it AI, and nobody loses sleep over it.
So why does the conversation change so significantly the moment an AI agent enters the picture? The automation isn't new. The absence of a human in the loop isn't new. What's new is that the rules are no longer fully explicit, and that's a different kind of problem entirely.
A familiar pattern
When a new technology arrives faster than the frameworks to govern it, the natural response is to treat the technology itself as the risk. That's not irrational, but it tends to obscure where the real exposure is, and in regulated industries, that has a cost.
A useful parallel: the early days of motor cars. People understood that cars were faster and more capable than horses, but not yet what made them dangerous or under what conditions. The vehicle arrived before the infrastructure did. Traffic lights, road rules, speed limits: none of that existed yet. The discomfort wasn't wrong, it was just ahead of a clear understanding of where the real hazards lay.
We're in a similar moment with AI. The infrastructure around it is still being built, which means some of the caution is well-founded, even if it isn't always aimed at the right thing. The organisations that navigate this well will be the ones who can distinguish between AI being involved and the actual risk changing.
Where the risk actually sits
The technical distinction that matters here is determinism. Parametric payment systems are built on it: given these inputs, this output follows, every time. The system doesn't approximate or interpret; it executes a defined rule.
However, AI models work differently. They are fundamentally probabilistic, making predictions rather than executing rules, which is a different kind of reliability guarantee entirely. For most applications, that distinction is small and tolerable. For payment instructions in regulated markets, it isn't. When funds move, they have to go to the right place, first time, and probabilistic outputs and payment instructions don't mix well.
The question for any regulated business isn't whether to use AI, but where.
There are functions where variability in output is either low-stakes or easily caught, internal workflows, exception identification, documentation, operational support, and there are decisions where the output needs to be fully deterministic before AI gets anywhere near them. Getting that sequencing right matters more than moving quickly.
The more useful question
For companies working through their own AI adoption, the instinct to ask "is AI involved?" isn't wrong, but it's incomplete. The more useful question is whether the output is deterministic enough for what you're asking it to do.
Some decisions require absolute repeatability, while others tolerate interpretation. A lot of the anxiety around AI in regulated industries comes from applying the standards of the first category to situations that actually belong in the second, and vice versa. Getting that mapping right is harder than it sounds, and the frameworks to do it consistently are still being developed across the industry.
The organisations that will use AI well in financial services won't be the ones that move fastest or the ones that hold back longest. They'll be the ones that develop a clear-eyed view of where determinism is required, and build their AI adoption around that line.
