Financial Services in the Age of Agency

Executive summary

• Agents are already in production. DBS has put multi-agent credit tools in front of about 1,500 bankers and lets millions of customers complete transactions through its assistants. Goldman Sachs is testing Claude-based agents in accounting, compliance and onboarding. Allianz has launched agentic claims handling in Australia ^[3][4][2][5].
• Regulators are moving in step, but no agent-specific rulebook exists yet. MAS finalised its AI Risk Management Guidelines in October 2026 and will consult on agentic AI in 2027. The FSB finalised its 12 sound practices on 15 October 2026. The HKMA plans a discussion paper on governing AI agents. The EU AI Act's high-risk obligations for credit scoring and for life and health insurance pricing now apply from 2 December 2027 ^{[1][9][11][13][14]}.
• The board's job is changing. It now has to decide what agents may do, under whose authority, with what limits, and who answers when they get it wrong. That includes vendor models, where accountability stays firmly with the institution ^[1][10].

What's changing

Financial services is moving from AI that drafts and summarises to AI that carries out multi-step work within set limits. Four developments show the pattern.

1. Agents inside credit decisioning (Asia-Pacific). In August 2026, DBS expanded its agentic credit-assessment tools from a 150-person pilot to about 1,500 relationship and credit risk managers worldwide. Specialised agents handle more than 70 distinct tasks. They pull from annual reports, industry research and internal records to produce a review-ready draft credit memo. The bank aims to cut information-gathering and documentation time by at least 30%. Bankers keep full responsibility for the final decision ^[3].

2. Agents that act for customers. In July 2026, DBS upgraded its corporate and retail assistants from answering questions to completing selected banking requests. The upgrade reaches more than 10 million customers across Singapore, Hong Kong and Taiwan, starting with about 350,000 business clients in Singapore ^[4].

3. Agents in the back office (US). Goldman Sachs has worked with Anthropic for about six months to build autonomous agents for accounting, compliance checks, client onboarding, trade reconciliation and document review. Its CIO describes them as a digital co-worker rather than a replacement for staff, at least for now ^[2].

4. Agents in claims and advice. Allianz launched its first agentic system, Project Nemo, in Australia to handle food-spoilage claims, and plans a wider network of specialised agents ^[5]. In wealth management, Deloitte estimates that advisers spend nearly 70% of their time on work outside client relationships. Only about 6% of advisory firms use agentic tools so far, against 73% using AI in some form ^[6].

Across the industry, Accenture reports that 57% of banking executives expect agents to be fully embedded in risk, compliance, audit, fraud and transaction monitoring within three years. A further 56% expect broad use in credit, loan processing and KYC ^[8].

Through the Age of Agency lens

Age of Agency: from tools to delegated authority. The key change is not better analytics but delegation. When an agent assembles a credit memo, starts a claims payment or completes a customer request, the firm has handed over part of its operating authority. This changes three things.

• How firms decide. Decision rights need to be written down for each agent action, not just for each human role. The FSB's own recommendations already lean this way: define which actions are prohibited or need human approval, and use dual approval above set transaction thresholds ^[10].
• How firms delegate. The FSB suggests treating agents as "synthetic employees", with individual identifiers, access rights and assigned responsibility ^[10]. The HKMA's planned agentic framework points the same way, proposing a registry of material agents and intervention mechanisms ^[11]. Boards that already have mature delegation-of-authority frameworks for people can extend them. Boards that do not will find agents expose the gap.
• How firms compete. Deloitte finds that technology-stack readiness is the biggest factor in how much value agents deliver, and that value is capped by the weakest of three levers: adoption, process redesign and the tech stack ^[6]. Competitive advantage will therefore come from data, permissions and process architecture, not from access to models, which most firms buy from the same few developers ^[1].

Agentrification: who gains, who loses. As agents take over tasks, value and roles shift in predictable directions.

• Mid-office and preparation work shrinks first. Information gathering, reconciliation, onboarding documents and first-pass compliance review are exactly the tasks being targeted ^[2][3]. Roles built mainly around assembling material will contract or be redefined.
• Supervision becomes a role in its own right. A Capgemini study cited by Banking Dive found that nearly half of banks and insurers are creating roles to supervise AI agents ^[8]. "Agent supervisor" and "agent risk owner" are becoming real jobs, not side duties.
• Front-line judgement goes up in value, but so does pressure on fees. Deloitte expects advisers to shift towards judgement and client conversations. Lower cost-to-serve may widen margins at some firms but compress fees at others ^[6]. In wealth management, firms that turn freed capacity into more clients per adviser will gain. Firms that only take the cost saving may find competitors pass it on to clients.
• The value of pricing data changes. As agents make servicing cheaper, an institution's private data, risk appetite and client trust become a larger share of what it is actually selling.

Universal Basic Intelligence (UBInt): the family office gap. UBInt is directly relevant for smaller institutions. Citi reports that 22% of family offices use AI for operations or investment analysis, up from 13% the year before. However, 57% cite a lack of internal expertise as the main barrier, and privacy is non-negotiable ^[7]. Large banks can build agent platforms with frontier labs. Family offices, independent wealth managers and smaller insurers cannot. If capable, secure AI becomes basic infrastructure for the industry, the question is whether smaller firms reach it through trusted shared platforms, through their banks, or not at all. Firms left without it will see their operating costs and service levels fall further behind.

Risks and governance

Key risks

• Speed and compounding error. The FSB warns that agent risks can appear very quickly. An agent may take hundreds of intermediate steps before a human steps in, and connections to APIs, databases and other agents can turn one fault into a wider operational, cyber or conduct incident ^[10].
• Third-party and concentration risk. Most institutions rely on models from a small number of developers, yet accountability stays with the firm ^[1].
• Gaps in model risk coverage. In April 2026, the US Federal Reserve, OCC and FDIC replaced their 2011 model risk guidance. The new guidance explicitly excludes generative and agentic AI, which must instead be governed through broader risk management ^[12]. Firms should not assume their existing model validation function covers agents.
• Conduct and fairness. Agents used in credit, insurance pricing and advice carry discrimination and suitability risk that regulators already scrutinise ^[14][15].

Regulation and guidance by region

• Singapore. The MAS Guidelines on AI Risk Management apply to all financial institutions and all types of AI. They require board oversight, an AI inventory, materiality assessment of each use case, and lifecycle controls. Firms stay accountable for vendor AI and should limit, suspend or replace third-party services that fall outside their risk appetite. Key sections take effect from October 2027 and October 2028, and MAS will consult on agentic AI in 2027 ^[1].
• Hong Kong. On 15 October 2026, the HKMA launched its "ART" framework (accountability, resilience, trust). It includes updated supervisory guidance and a forthcoming discussion paper on governing agentic AI. More than 60% of use cases in its second GenAI Sandbox cohort explored agentic AI ^[11].
• Global. The FSB's final 12 sound practices cover board oversight, materiality assessment, explainability, human oversight, cyber risk and third-party AI dependencies. They are not binding, but the FSB strongly encourages boards to use them ^[9].
• European Union. Regulation (EU) 2026/1744 (the Digital Omnibus on AI) was published on 24 July 2026. It moves the deadline for stand-alone Annex III high-risk systems to 2 December 2027 ^[13]. Creditworthiness assessment and credit scoring of individuals, and risk assessment and pricing in life and health insurance, are high-risk. Claims management and fraud-only systems are generally outside scope ^[14].
• United States (insurance). As of 8 October 2026, 29 jurisdictions have adopted the NAIC AI Model Bulletin. California, Colorado, New York and Texas have their own insurance-specific rules ^[15].

Questions for boards and leadership

1. Which agents in our firm can take an action rather than only make a recommendation, and is there a single register of them with a named owner for each?
2. For each agent, which actions are prohibited, which need human or dual approval, and at what financial threshold?
3. If a vendor's model changes or fails tomorrow, what can we suspend, replace or run manually, and how quickly?
4. Our model risk framework may not cover agentic AI. Who validates and monitors agents, and against what standard?
5. Which of our agent uses fall into EU high-risk categories, or into MAS's material category, and will they meet the 2027 deadlines?
6. When agents free up staff time, where does that capacity go: more clients per banker or adviser, better risk work, or headcount reduction? Have we decided deliberately?
7. Which roles are we building to supervise agents, and are we hiring and training for them now?

How KIG helps

• KIG Advisory. We help boards and executive teams decide where agents should and should not act. We design delegation-of-authority and agent-governance models aligned with MAS, HKMA, FSB, EU AI Act and US expectations, and we redesign operating models so that freed capacity turns into competitive advantage.
• KIG Research. We apply the Age of Agency, Agentrification and UBInt frameworks to financial services. This gives leadership a shared language for decision rights, workforce change and access to capable AI, and we track regulatory developments across Asia-Pacific, the US and Europe.
• KIG Ventures. We bring practical lessons from building and backing AI companies: how agent products are actually designed, where they fail, and what an institution should demand from vendors before it hands them authority.