AI agents are moving from chatbots to transaction executors, and the infrastructure gap is becoming painfully obvious. Fireblocks is positioning itself as the solution for institutions that want autonomous systems handling real money.

The digital asset custody firm laid out its vision for “agentic finance” in a detailed framework published this week, arguing that AI agents capable of treasury rebalancing, cross-border payment routing, and liquidity management need fundamentally different infrastructure than conversational AI models.

The Core Problem

Here’s what Fireblocks identifies as the disconnect: language models reason and generate text, but financial AI agents need to execute. That means touching real assets under real regulatory constraints.

Three infrastructure gaps stand out. First, transaction authorization—AI agents can’t operate with static API keys granting blanket access. A treasury agent moving $10 million between accounts requires different authorization than one processing a $50,000 supplier payment. Dynamic, policy-governed controls are non-negotiable.

Second, multi-party orchestration. A B2B cross-border payment doesn’t hit a single API. The agent must coordinate liquidity across venues, route through compliant rails, settle with counterparties, and reconcile across ledgers simultaneously.

Third, auditability. When regulators ask why an AI rebalanced treasury positions or chose a specific payment route, every decision point needs documentation. Not for debugging—for compliance.

Why Finance Breaks Generic AI Infrastructure

Most AI infrastructure conversations focus on model performance, training costs, and inference speed. Fireblocks argues that’s missing the point for financial applications.

An agent routing stablecoin payments across borders for a remittance company isn’t just optimizing for speed and cost. It’s ensuring every transaction meets AML/KYC requirements, passes sanctions screening, and satisfies travel rule obligations across multiple jurisdictions. That compliance layer needs to live in the infrastructure itself.

Counterparty risk adds another wrinkle. AI agents making liquidity decisions need real-time access to settlement windows, credit limits, and exposure thresholds—dynamic state that shifts with market conditions, not static configuration files.

Then there’s the custody question. These agents don’t just read data; they move assets. That demands cryptographic controls, transaction signing protocols, and security policies that respect organizational limits on approvals and thresholds.

The Stack Fireblocks Is Pitching

The company outlines five infrastructure components for agentic finance:

Security architecture with distributed key shares that never expose private keys to the agent or any single party. A policy engine defining what agents can and cannot do. API-first wallet infrastructure supporting multiple blockchains and both traditional and digital asset rails. Network access to exchanges, OTC desks, DeFi protocols, and payment networks through single integrations. And compliance systems ensuring agent-initiated transactions face the same scrutiny as human ones.

Fireblocks makes a pointed observation about market approach: “Most companies building AI infrastructure for finance are starting from the intelligence layer and working backward. The ones succeeding are starting from the transaction execution layer and working forward.”

What This Means for Institutions

The pitch is clear—general-purpose cloud APIs won’t cut it for production financial AI. Whether that’s true remains to be tested at scale, but the regulatory logic is sound. Financial regulators haven’t historically been impressed by “the AI decided” as an explanation.

For institutions exploring autonomous treasury operations or payment routing, the framework highlights questions worth asking any infrastructure provider: Where do keys live? How are policies enforced? What’s the audit trail when something goes wrong?

The AI agent race in finance is just starting. The infrastructure layer underneath may determine who actually ships production systems versus who stays in pilot purgatory.

Image source: Shutterstock