27 Jul 2026

The reason fintech hasn't produced a trillion-dollar company isn't a mystery. It's regulation and product depth, or rather the lack of both working in your favour at the same time.

Stripe, Revolut, and Nubank are genuinely impressive businesses. But they've been built in an era where crossing a border meant rebuilding your compliance stack from scratch, and where customer lock-in depended on adding product after product to justify staying. That's an expensive way to grow, and it caps how fast you can scale globally.

The AI and tokenisation argument in the original piece is where it gets interesting for those of us running UK credit operations. If the cost of launching a new financial product drops sharply, and if infrastructure starts to become genuinely portable across jurisdictions, the competitive logic changes. Right now, a UK consumer credit broker competes on distribution, pricing, and customer experience within a fairly contained regulatory perimeter. That perimeter has always been a moat. If it becomes less relevant, the moat shrinks.

I'd push back slightly on the breathless framing around which company 'wins'. The trillion-dollar question is a bit of a distraction. The more pressing question for UK financial services leaders is what happens to mid-tier operators when the barriers that protect them start eroding.

• Regulatory fragmentation has always slowed the largest fintechs down, but it's also protected domestic incumbents
• AI lowering product launch costs doesn't just help Stripe, it helps anyone with distribution and data

The companies best placed for hyperscale aren't necessarily the ones with the best technology. They're the ones that accumulated real customer relationships before the cost curves changed. In UK consumer credit, that's a prompt to think hard about what you actually own in your customer base, and whether it would survive a world where switching costs keep falling.

The framing around SRE AI agents usually focuses on speed — faster incident resolution, less toil, engineers freed up for meaningful work. That's fine as far as it goes, but it misses the more uncomfortable implication for anyone running financial services infrastructure.

These agents learn from historical operations data. Which means the quality of what you feed them determines everything. Most organisations in UK consumer finance have accumulated years of inconsistent runbooks, undocumented tribal knowledge, and incident post-mortems that were written to satisfy an audit rather than to actually capture what happened. An AI agent trained on that data doesn't get smarter. It gets confidently wrong.

This is the real preparation work, and almost nobody is doing it. Before you think about autonomous incident resolution, you need to ask:

• Are your runbooks actually accurate, or do engineers quietly improvise around them?
• Do your post-mortems capture genuine root cause, or just timeline reconstruction?
• Is your observability data clean enough to be meaningful as training signal?

In a credit brokerage context, this matters beyond operational efficiency. The FCA expects firms to understand and control their material systems. If an AI agent is making autonomous decisions about production infrastructure that processes customer credit applications, the question of who set the guardrails and on what basis is a governance question, not just an engineering one.

The shift from engineers executing runbooks to engineers setting agent guardrails sounds like an upgrade in job quality. It probably is. But it also means the judgement that used to live in someone's head at 2am now needs to be made explicit, documented, and defensible before the incident happens.

Most teams aren't ready for that conversation. Are yours?

Consumer Reports has put its finger on something the UK financial services industry should be taking seriously right now. As AI-powered financial assistants move from novelty to infrastructure, the question of whose interests they actually serve is becoming urgent. And the honest answer, in most current deployments, is the provider's.

The framework Consumer Reports proposes centres on a fiduciary-like duty of loyalty. That framing is deliberate and important. Fiduciary duty is a well-understood legal and ethical concept in financial services. Applying it to AI behaviour gives regulators, firms, and consumers a coherent standard to argue from, rather than the vague "responsible AI" language that currently dominates the conversation.

For UK consumer credit specifically, this matters enormously. We already operate under FCA rules requiring firms to act in customers' best interests, and the Consumer Duty has sharpened that obligation considerably. But Consumer Duty was written for human decision-making with human accountability. An AI assistant that nudges a customer toward a higher-margin product, through subtle framing or selective information presentation, may never trigger a traditional conduct review. The behaviour is too granular, too fast, and too embedded in the interaction design to catch with conventional oversight.

Two things should concern technology leaders building or buying these systems:

• Engagement optimisation is not neutral. A model trained to maximise session length or product uptake will find ways to serve that goal, even when the outputs look helpful on the surface.
• Explainability requirements under UK GDPR give consumers the right to understand automated decisions, but that right is nearly impossible to exercise meaningfully when the AI is shaping a conversation rather than making a binary credit decision.

The FCA is watching this space. Its recent work on AI and its ongoing Consumer Duty supervisory focus will eventually converge on exactly the question Consumer Reports is raising. Firms that wait for regulatory clarity before thinking about AI loyalty are making a bet that the regulator will be slow and lenient. Given the current political environment around consumer protection, that seems like a poor wager.

The real question is whether "duty of loyalty" can be operationalised in model design and governance, or whether it remains a principle that sounds good in a framework document and disappears in production.

The Visa and Airwallex freight partnership is worth paying attention to, and not because of the names involved. It's worth paying attention to because it targets a sector where the pain is genuinely acute: 42-day settlement cycles and payment admin eating up a fifth of transportation costs. That specificity matters.

Most embedded finance announcements are solutions looking for a problem. This one starts from operational misery and works backwards to the product. That's the model that actually gets adoption.

For UK consumer credit, the lesson isn't about freight. It's about where your own payment friction lives and whether you're treating it as a cost of doing business rather than a product problem to solve.

Consumer lending has its own version of this. Loan disbursements that don't reach customers on weekends. Repayment failures that trigger arrears processes before anyone checks whether the bank details simply changed. Manual reconciliation between origination platforms and payment providers that adds days to funding cycles and headcount to operations teams.

The freight sector tolerated 42-day settlements for years because no single player had the incentive to fix the whole chain. Embedded finance changes that equation by putting the financial tooling inside the software the industry already uses daily.

That last point is the one UK fintech and lending leaders should sit with. The competitive advantage isn't the payment capability itself. Visa and Airwallex aren't winning here because they can move money internationally. Plenty of providers can do that. They're winning because they're meeting logistics companies inside the workflow, not asking those companies to adopt a separate financial product.

Where in your own origination or servicing stack are you still asking customers or operations teams to step outside the workflow to handle money movement? That's your 42-day settlement equivalent.