Finance's AI problem isn't the model. It's the foundation.

Photo by Erik Mclean on Pexels

How Upflow built the first AI-native Financial Relationship Management platform and why the quality of AI agents in finance comes down to what sits underneath them.
-- Alex Louisy, CEO, Upflow

Every company I speak to right now is somewhere on the same journey: they've added an AI layer to their finance stack, they've automated a few reminders, maybe an approval workflow, and they're calling it intelligent collections. Some of them are right to be excited. Most of them have just made their existing problems run faster.

That's the thing nobody is saying clearly enough about AI in B2B finance. The quality of what an AI agent does is entirely determined by the foundation it operates on. And for most companies, that foundation was never built with relationships in mind.

The category that has dominated B2B finance software for the past decade is AR Automation. It's a real category solving a real problem: companies send invoices and don't get paid on time, so you automate the reminders. Faster emails. Batched dunning sequences. Rules-based escalation. It works, up to a point.

What AR Automation was never designed to answer is a different question: what about the relationship the invoice lands inside? The sales rep who just closed that customer. The CS manager negotiating a renewal. The VP of Finance who told their biggest account they'd sort out the billing issue personally. AR Automation doesn't know any of that. It sees a payment date and a balance.

When you bolt an AI agent onto that context-free infrastructure, you get an agent that sends the right message at the wrong time, to the wrong person, with no awareness of what's actually happening. That's not a technology failure. It's a foundation failure.

At Upflow, we've been thinking about this problem since before AI agents were part of the conversation. We call the alternative Financial Relationship Management, and it rests on three things.

The first is context. The full picture of a customer relationship has to live in one place, clean and current, before any automation acts on it. Not just payment history, but CRM notes, email threads, call recordings, CS interactions. Finance teams today spend a shocking amount of their time pulling this together manually before they can make a single decision. The infrastructure has to do that work instead.

The second is coordination. Collections is no longer a finance-only job. Sales and customer success are both involved, because they own the relationship and they're increasingly accountable for whether cash gets collected. Leadership is in it too. The result is that customers regularly get three different messages from three different teams, with three different tones, in the same week. That's not a communication problem. It's a coordination problem, and it has to be solved at the infrastructure level before AI touches it.

The third is trust. The customer experience of paying a B2B invoice in 2026 is, in most cases, still worse than buying something online in 2012. A PDF attachment, a no-reply sender, a generic portal with no branding. Meanwhile the same customer uses seamless payment tools in their personal life without a second thought. That gap matters. Getting paid is a touchpoint in the customer relationship, and it should be treated like one.

This is where the FRM vision and AI intersect in a way I find genuinely interesting.

We've seen agents dramatically raise the bar on what can be automated. A few years ago, automating a reminder was table stakes. Today an agent can handle an entire collection cycle autonomously, including handling disputes, escalating to the right person at the right moment, and adapting its approach based on the customer's history and current relationship status. The technology is real.

But that autonomy can only be trusted if the underlying infrastructure is trustworthy. An agent operating on bad context will make confident mistakes at scale. An agent with no coordination layer will damage relationships it has no visibility into. An agent pushing payments through a broken experience will collect money while destroying goodwill.

What we're building at Upflow is what I'd call supervised autonomy. Finance teams set the parameters, the agents execute, and every action is fully auditable. The CFO can see exactly what happened with every customer, why the agent took the action it did, and where a human needs to step in. That's not a limitation on the AI. It's how you make it actually usable in a function where trust is the whole game.

One CFO I spoke to recently put it simply. His company did everything perfectly for the customer experience, everything except one thing: when they sent an invoice, it was a miserable experience. That gap is what FRM is designed to close, across collections, across payments, and eventually across the financing decisions that sit alongside them.

The companies getting AI right in finance right now are not the ones adding the most impressive model to their stack. They're the ones who built the foundation first.

Alex Louisy is the CEO and co-founder of Upflow, the Financial Relationship Management platform. Upflow helps B2B companies manage their customer financial relationships across collections, payments, and beyond.