What Have We Authorised When AI Takes Action?
Agentic AI executes financial actions, so delegated authority must become machine-readable, bounded by limits, supervised, and auditable.
Key Takeaways
- Agentic AI in finance shifts AI from giving advice to taking actions, making delegated authority the central oversight question.
- Human review of every intermediate step may be unrealistic, so firms need accountability, controls, and oversight across the full workflow.
- Authority must become machine-readable: AI mandates need clear boundaries such as objectives, limits, accounts, products, counterparties, validity periods, and escalation rules.
- Institutions must connect mandate, decision, action, and evidence so they can reconstruct whether AI actions stayed within authorized limits.
Meta description: Agentic AI in finance can take actions, not just give advice. SFC CEO Julia Leung says oversight matters, so delegated authority must become auditable data.
Featured image: abstract friendly illustration of financial authority becoming a data trail.
For financial institutions, agentic AI in finance creates a new question: what was the agent authorised to do? AI has so far mainly supported people by analysing information, preparing research, answering customer questions, detecting unusual activity and helping with decisions. Usually a person or an established business process takes over at a clear point. Agentic AI changes that: an agent can be given an objective, decide how to pursue it, use tools and systems, and carry out actions without asking a person to approve every intermediate step. The shift is from recommending an action to taking one. Much discussion focuses on model risk, explainability and keeping humans involved; the deeper question is how to establish the authority delegated to the agent. This may turn authority itself into a data problem. In finance, agentic AI may mean moving funds, managing liquidity, rebalancing investments, responding to collateral requirements or interacting with other automated services. Speaking at Hong Kong FinTech Week 2026, SFC CEO Julia Leung noted that human review of every intermediate step may not be realistic as firms adopt more autonomous AI applications, and highlighted the need for human accountability and oversight throughout the workflow.
From transaction approval to process approval
Financial institutions already represent authority through authorised signatories, approval matrices, dealing and position limits, system entitlements, customer-approved payments and investment mandates. Agentic AI differs because a person may not authorise a particular transaction. Instead, the person may authorise an objective and allow the agent to decide the actions needed. A company could instruct an AI treasury agent to maintain sufficient liquidity for the next 30 days, invest surplus cash in approved low-risk instruments, and obtain additional funding if projected cash falls below an internal buffer. That instruction contains many decisions: sufficient liquidity level, cash-flow forecast, qualifying low-risk instruments, funding trigger and acceptable funding source. Financial systems are good at recording who approved a transaction or who has access to an account, but less accustomed to representing a delegated machine mandate that may produce a series of financial actions not individually specified. The company has delegated a decision process within boundaries.
When authority becomes machine-readable data
A broad natural-language instruction may be understandable to a person but not precise enough for an autonomous agent. Terms such as "sufficient liquidity", "low risk", "reasonable cost" or "when necessary" require interpretation, and two agents could differ because of models, assumptions, market data or business rules. For meaningful autonomy, key parts of the mandate may need to be represented in a form systems can understand, apply and test. This makes agentic AI, at least partly, a data architecture issue. It does not require code for every instruction or a new regulatory data standard for every application. Product restrictions, transaction limits, counterparty limits, account permissions, approval rules and risk parameters already exist. The change is connecting those controls around an agent that may operate across an end-to-end process. A mandate could specify activities, accounts, products, financial limits, counterparties or markets, validity period, human-approval circumstances, whether delegation to another service is allowed, and how authority can be suspended or revoked. Technical access and financial authority are not the same. Identity and access controls can show the agent is recognised; the harder test is whether the type, size and circumstances of an action remain within authority. Establishing that a transaction came from an approved agent may not be enough; institutions may also need to test whether the action was consistent with the authorised decision process and limits. One mandate may produce many different actions over time.
Evidence must connect mandate, decision and action
A useful sequence is: Mandate -> Decision -> Action -> Evidence. The mandate defines objective and boundaries; the decision translates it into a course of action; the action happens in the financial system; evidence lets the institution reconstruct the connection. A technically valid transaction does not prove the agent stayed within authority. In the treasury example, an agent might identify a liquidity shortfall, sell part of a short-term investment portfolio and arrange additional funding. Each transaction might pass technical checks, use an approved connection, access permitted accounts and transact through recognised institutions. If the outcome creates a liquidity or concentration problem, the final transactions alone do not tell the whole story. The institution may need to establish which mandate applied, what liquidity threshold triggered action, whether cash-flow data were current, which limits were checked and whether human intervention should have occurred. This does not mean storing every internal calculation. The useful objective is decision-useful evidence: reliable information on authority, material conditions, controls, execution and any exception or human intervention. Not all of this belongs in a regulatory return, and supervisors do not need it continuously. The evidence should exist in a connected form so activity can be reconstructed for internal control, audit, customer redress or supervision. This echoes Transactions Without Infrastructure: a visible transaction is not necessarily institutionally complete. The surrounding relationships, rights and responsibilities still matter. A transaction executed by an AI agent may be accurately recorded while the record itself says little about the authority under which the agent acted.
Human oversight should focus on decision points
The idea of a "human in the loop" also needs redesign. Human oversight matters where AI can take actions with material financial consequences, but simply placing a person in the process does not show whether the control works. Approving every action creates a clear control point but removes much of the agent's benefit. Reviewing outcomes much later may leave little value for intervention. A practical model identifies decision points needing human judgement. A treasury agent might manage investments within approved products and counterparty limits, but stop if projected liquidity falls below a defined level or if it needs a new borrowing relationship. A portfolio agent might rebalance within agreed limits but require confirmation before entering a new asset class or materially increasing concentration. Oversight depends less on the number of approvals and more on whether the right issue reaches the right person at the right time. This connects with 24/7 Markets, 9-to-5 Supervision? Continuous supervisory awareness does not require watching every transaction; it requires signals showing when human attention is needed. Human accountability does not require approving every machine action. It requires clear boundaries, escalation points and enough information for meaningful intervention.
Agent chains make authority harder to trace
The issue becomes more complicated when an agent uses other agents or external services. A corporate treasury agent might use a specialist service to compare funding options. That service could obtain information from several financial institutions, while another system executes the selected transaction or manages collateral. The original mandate could pass through several technical and institutional layers before the final action. Long chains, delegation and outsourcing are familiar. What may change is the speed and flexibility with which agents choose and use services while completing a task. This creates an authority chain. If one authorised agent calls another, the question is how much delegated authority follows the task through that chain.
Tags: agentic AI, financial services, AI governance, Julia Leung, Hong Kong FinTech Week 2026
Comments (0)