AI & Financial Systems

AI becomes part of financial infrastructure when its outputs influence obligations, access to money and the records institutions rely on; that transition requires explicit authority, reliable settlement and accountable operations.

Why this matters

For a bank, asset manager or corporate treasury, AI can help staff interpret documents, investigate exceptions and prepare decisions. Giving a system authority to act introduces a further choice. A purchase, transfer or change to a financial record needs a clear principal, an approved purpose and evidence of what occurred. Institutions must assess the entire financial process around the model, including the people who remain responsible when it fails.

Public developments make these questions practical. DBS and Visa reported agent-initiated food and beverage purchases in February 2026, using cards through issuer-controlled flows in a joint pilot. Coinbase's x402 protocol offers a different mechanism for requesting stablecoin payment for web resources. These sources establish reported pilots and available technical approaches. They do not establish general adoption, comparative savings or the reliability of a particular institution's deployment.

Programmable finance also has an operating agenda beyond AI. MAS's BLOOM initiative explores settlement using tokenised bank liabilities and well-regulated stablecoins, with agentic payments among its wholesale areas. Tokenising an asset or automating an instruction still leaves questions about custody, liquidity, completion, accounting and recovery. An attractive demonstration can leave those dependencies unresolved.

Our work in this area examines those dependencies together. Singapore provides concrete sources and implementation context; regional operation requires each ASEAN route to account for its participants, currencies, providers and applicable requirements. The aim is to help financial and technology leaders choose bounded uses of AI and programmable infrastructure that they can fund, govern and operate.

Questions guiding this agenda

1. What financial authority may an agent exercise? Identify the person or institution delegating authority, eligible counterparties, exposure limits, approval checkpoints and the action that creates commitment. Separate a recommendation from a payment or contractual instruction.
2. What does identity prove? Connect an authenticated agent and payment credential to an accountable principal and current mandate. Examine how customer due diligence, supplier approval and ongoing transaction controls work when software initiates the request.
3. Which settlement route fits the obligation? Compare cards, bank payments, tokenised bank liabilities and stablecoins using actual participant access, legal rights, funding needs and recipient outcomes. A programmable interface does not make these assets equivalent.
4. Can operations establish the financial state after failure? Trace payment, custody, ownership and accounting records through duplicates, missing responses and disputed transactions. Define who can stop new commitments and manage completed ones.
5. Where does value appear in the whole workload? Measure correct fulfilment and resolved exceptions, including review, control and reconciliation effort. Supplier capability and faster ledger movement do not establish a cash saving.
6. How could shared dependencies magnify disruption? Examine concentration in models, cloud services, custodians, payment facilitators and settlement networks. Institutional contingency planning should consider failures that affect several processes or counterparties simultaneously, without asserting that a particular systemic event is inevitable.

Analysis and resources

When AI agents control money proposes a delegation record and tests for bounded spending authority. Stablecoins as settlement infrastructure for the agent economy compares payment routes and follows the exposure from funding to usable recipient money.

For tokenised assets, Tokenisation after the demo maps the operating work between instruction, settlement, custody and the books. Its exception tabletop supplies a clearly fictional transaction for teams to rehearse. The reconciliation evaluation protocol and evaluation pack help teams design a controlled test rather than mistake a matching demonstration for measured performance.

These materials are sourced analysis, proposed methods and working resources. They do not claim WCAI has conducted a banking benchmark, operated a payment network or delivered the third-party transactions they discuss. The questions remain open to challenge by financial, technology and risk practitioners.

Sources that shape the decisions

Singapore's agentic AI governance framework provides deployment guidance on limits, human accountability and controls. MAS's August 2026 reply on agentic finance describes runtime safeguards as a potential implementation approach. Neither source is a substitute for assessing a specific regulated activity.

For settlement choices, the BIS's 2025 analysis challenges assumptions about stablecoins' monetary role. MAS's September 2026 stablecoin consultation concerns proposed implementing legislation and further safeguards, with feedback due in October. Keeping policy proposals distinct from current permissions is part of the analysis.

Discuss a financial operating decision

A private executive briefing can examine a proposed agent payment, an AI-supported operations workflow or a programmable settlement route. Bring the business purpose, participating entities, proposed authority, current process and unresolved questions. That provides a concrete starting point for assessing the design, evidence required and work needed before deployment.