Agentic Payments: How AI Agents Became Autonomous Financial Actors

# Agentic Payments: How AI Agents Became Autonomous Financial Actors

For the past three years, AI agents were useful for drafting emails and summarizing calls. Now they're moving money.

HSBC UK is trialing AI agents inside its payment infrastructure. ConnectOne Bank is deploying them across commercial lending. nCino is rolling out AI agents to handle underwriting decisions in real time. These aren't chatbots answering customer questions — they're systems with the authority to initiate transactions, approve credit, and route capital without a human in the loop.

This is agentic commerce: the shift from AI as a decision-support tool to AI as a financial actor.

What agentic payments actually means

An agentic payment is any financial transaction initiated, routed, or completed by an AI system acting on delegated authority — not by a human clicking a button.

The distinction matters. Automated payments have existed for decades. What's new is that agents aren't just executing pre-programmed rules. They're making contextual decisions: when to pay, how much, which vendor, under what conditions. They adjust based on inputs. They escalate when thresholds are breached. They operate continuously, not just when triggered.

In commercial lending, nCino's agents are assessing loan applications against live financial data and issuing preliminary decisions. At ConnectOne Bank, the agents are embedded in the credit workflow — not advising a loan officer, but acting as one.

The difference between a rules engine and an agent is judgment. Agents have it.

Why banks are moving now

Three pressures converged in 2025 that made agentic payments viable at scale.

First, model reliability improved. GPT-4-class models had hallucination rates that disqualified them from financial decisions. The generation of models deployed in 2025 — including fine-tuned vertical models trained on financial data — operate within acceptable error tolerances for structured decisions like payment routing and credit assessment.

Second, regulatory sandboxes opened. The UK's Financial Conduct Authority and the EU's AI Act both created frameworks for supervised agentic systems in financial services. Banks can now run live trials with liability guardrails, rather than waiting for full regulatory clarity.

Third, the cost case became obvious. A human loan officer costs a bank roughly $80,000–$120,000 per year in salary and benefits, handles 200–400 applications annually, and works 40 hours a week. An agent costs a fraction of that, handles thousands of applications, and runs around the clock. For commercial lending volumes at mid-market banks, the unit economics are hard to argue with.

The architecture behind it

Agentic payment systems typically run on a layered stack:

• Perception layer: ingests documents, transaction histories, credit bureau data, market signals
• Reasoning layer: applies decision logic, policy constraints, and risk models
• Action layer: interfaces with payment rails, core banking systems, or lending platforms
• Audit layer: logs every decision with a traceable rationale for compliance review

The audit layer is non-negotiable. Regulators don't object to agents making decisions — they object to decisions that can't be explained. Systems that produce structured decision logs are getting approved for production. Systems that can't are stuck in sandbox.

For founders building AI-native companies, this stack is increasingly available off the shelf. Stripe, Plaid, and Modern Treasury all have API surfaces that agent frameworks can call directly. The infrastructure is there. The question is whether your product logic is worth wrapping around it.

What this means for solo founders

Here's where it gets interesting for a smaller operator.

Agentic payments aren't just a bank product. They're a capability that any sufficiently technical founder can now deploy in their own operations — and eventually, in their product.

Consider what a solo founder running a SaaS business currently does manually: chases invoices, approves vendor payments, monitors for fraud, reconciles accounts. Each of those is a repeatable decision workflow. Each of them can be handed to an agent.

A founder operating at one-person-unicorn scale — generating millions in revenue without headcount — can't afford to spend time on financial operations. Agentic payment infrastructure is what makes that math work. The agent handles the money movement. The founder handles the product.

This connects directly to how revenue per employee is becoming the defining metric for AI-native startups. When agents absorb the operational overhead of financial workflows, headcount stays flat while throughput scales. That ratio is the whole game.

Julien de Waal, who runs Sprinkal, Sonscape, and Nova Labs under the Waalhalla holding structure, builds with this constraint in mind. Across multiple ventures, agentic systems handle the repetitive operational layer so the founder layer stays focused on decisions that actually require judgment.

The risks nobody is talking about clearly enough

Agentic payments create two failure modes that don't exist in manual workflows.

Cascading errors. An agent operating across thousands of transactions can propagate a bad decision at scale before any human notices. In a manual process, a bad call affects one transaction. In an agentic one, it can affect ten thousand in the time it takes to review the first alert.

Delegation ambiguity. When an agent initiates a payment, who is legally responsible? The bank? The software vendor? The founder who configured the agent? This is unresolved in most jurisdictions. Until it isn't, every agentic payment system needs a clear human-in-the-loop escalation path for edge cases — not because the agent can't handle them, but because the legal framework assumes a human did.

Founders building on agentic payment infrastructure should treat the audit layer as a product requirement, not an afterthought. Your future compliance review will thank you.

The next 18 months

HSBC, ConnectOne, and nCino are early movers, but the pattern will generalize fast. Expect:

• B2B payment agents that handle procurement approvals, vendor payments, and contract-linked disbursements autonomously
• Lending agents at fintechs that underwrite, disburse, and service small business loans without human review below a threshold
• Treasury agents at mid-market companies that optimize cash positions, FX exposure, and short-term investments in real time

For founders learning how to build a one-person startup with AI, agentic payments are worth treating as a near-term operational tool, not a future abstraction. The infrastructure exists. The APIs are live. The only missing piece is your willingness to delegate financial workflows to a system you trust enough to configure properly.

That trust, and the discipline to build the guardrails around it, is the real skill.

---

Is your company eligible? Submit to the leaderboard → onepersonunicorn.co/submit