Most organisations have now crossed the threshold of AI adoption. The harder question (i.e. the one that actually determines who benefits) is what kind of adoption they're doing.
Geoffrey Moore's ‘Crossing the Chasm’ described the gap between early adopters of a technology and the mainstream market that eventually follows. The argument was that getting early adopters excited is the easy part. Getting everyone else to follow requires something different: a more mature product, a clearer value proposition, a willingness to commit.
AI is crossing a different kind of chasm right now.
Not between companies that have adopted it and those that haven't, because that gap is closing fast, and access is no longer the constraint. The chasm that matters is between organizations using AI to make their existing model incrementally better and those using it to build something structurally different. Both groups are deploying AI. The outcomes they're heading toward are not the same.
Efficiency isn't the ceiling
The first wave of enterprise AI adoption has largely been about efficiency: summarizing documents, drafting emails, flagging anomalies, and accelerating workflows. These are real gains, and nobody should dismiss them, but they accrue within an existing structure. The organization underneath stays roughly the same. Functions stay separate. Decision-making stays where it was. Data still gets treated as a record of what happened rather than a signal for what to do next.
The more significant shift, and the one that tends to separate companies seeing compounding returns from those seeing incremental ones, is when AI starts to change the structure itself. Teams organized around functions start to compress because AI can hold context across them. The question stops being 'how do we do this task faster?' and becomes 'does this task need to exist?' Data stops being something you query and becomes something that actively informs how work gets done, in real time.
The companies where this shift is most visible tend to be technology-native businesses, where clean data and end-to-end digital workflows make structural change more tractable. The harder and more interesting question is what this transition looks like for industries where the underlying infrastructure is more complex, the regulatory environment more demanding, and the cost of error significantly higher. That's not a reason to move slower than necessary, but it is a reason to be precise about what you're building toward and why.
What makes this hard in financial services
In most industries, experimentation is relatively low stakes. A wrong turn in a consumer app costs you some engagement. A wrong turn in a platform that moves money on behalf of insurers operating in regulated markets is a different order of problem, and the data reflects this.
Gartner's 2025 AI in Finance Survey found that momentum in finance function AI adoption has slowed, not because enthusiasm has faded, but because complexity, data quality, and talent challenges are proving harder to resolve than originally anticipated. The World Economic Forum's 2025 report on AI in financial services, produced with Accenture and drawing on more than 100 executive interviews, points to the same underlying tension: the gap between AI ambition and operational readiness is widest precisely where governance and accountability matter most.
This creates genuine tension. The organizations that will be best positioned in five years are the ones developing AI capabilities now, not just deploying off-the-shelf tools, but building genuine depth in how AI integrates with their operations. For companies operating at the intersection of banking and insurance, where financial infrastructure underpins the movement of claims funds across regulated markets, that tension is particularly noticeable. The tolerance for error at the transactional layer is essentially zero, and when funds move, they have to go to the right place the first time. Probabilistic outputs and payment instructions don't mix well.
Navigating that tension requires being precise about where AI belongs at each stage. There are parts of our operation where AI can move fast: internal workflows, exception identification, documentation, operational support. There are parts where it has to move carefully and prove itself before it gets closer to the core. That's not a technology constraint but a judgment call about sequencing, and getting the sequencing right matters more than moving quickly.
Where we are
At Vitesse, I believe we're somewhere in the middle of this. We're investing in AI across the business and we've moved well beyond exploration, but we haven't yet made the structural changes that would put us firmly on the other side of the chasm we're describing. We know what that looks like, and we're building toward it.
What we're trying to avoid is the pattern that's caught a lot of organizations out: treating AI as a layer added on top of existing processes rather than a reason to rethink them. The companies we've watched struggle aren't behind on the technology — they're ahead of themselves on ambition relative to the foundations underneath, where the data isn't clean, the workflows aren't clearly defined, and the infrastructure isn't ready for what they're asking it to do.
Getting the foundations right is less exciting than announcing what you're building on top of them. But it's what makes the difference between AI that compounds and AI that stalls.
This is the second post in Vitesse's ongoing series on AI in financial infrastructure. Read the first post in the series. We're sharing how we're thinking through these questions, not because we have settled answers, but because we think the conversation is worth having in the open.
