Agents in finance are mostly bounded and human-supervised, adoption figures are not comparable, and supervisors are extending rules rather than replacing them. Evidence, gaps and three NKK models.
Written for: Chief risk officers, heads of AI and model risk, compliance and operational resilience leads, financial services CIOs
13 chapters, 50 cited sources, 24 models and charts. The first 2 chapters are open; the rest unlock with the form.
Headline numbers on agentic AI in financial services measure different things, and the best supervisory evidence points one way: agents are doing bounded, human-supervised work in fraud, research and coding, within guardrailed use cases. The sources reviewed show no evidence of agents making core credit, trading or payment decisions on their own at scale. The control problem is real but narrower: piloted agents are governed by frameworks written for models and human approvers.
This report is a synthesis of public research; it does not contain primary survey data. Sources were opened up to 13 October 2026; projections, self-reported cases and unread sources are flagged. NKK's frameworks are opinion built on cited prior art.
Start with a problem that distorts every statistic that follows: there is no shared definition. The Financial Stability Board (FSB) calls agentic AI "a less formalised concept" and describes "systems designed to autonomously perform complex tasks over extended periods, often making decisions and taking actions with limited human oversight" [9]. The Bank of England's Financial Policy Committee defined it in 2025 as "systems which can take autonomous action to achieve specified goals by utilising tools, learning from feedback, and adapting to dynamic environments" [10]. FINRA describes agents as "systems or programs that are capable of autonomously performing and completing tasks on behalf of a user" [11].
Two differences matter in practice. The first is technical versus organisational framing. The FCA's Mills Review, published on 6 July 2026, adopts a five-level autonomy spectrum from academic work by Feng and colleagues, in which the human's role shifts from operator to collaborator, consultant, approver and observer [6][12]. The question becomes where the human sits, not how clever the model is. For central-bank staff, the BIS makes a related split between copilots that augment people and agents that automate specific tasks and can substitute for roles [13].
The second is scope. MAS lists "AI agents (including multi-agent systems)" as in scope of its AI risk guidelines, and says rule-based RPA is out [14]. Its final guidelines add that formula-driven tools and RPA with fixed instructions would not ordinarily count as AI [15].
| Criterion | Emphasis | What it adds | Stated boundary |
|---|---|---|---|
| FSB (2025) | Autonomy over extended tasks | Notes the concept is less formalised and often built on generative AI | Calls it "a less formalised concept" |
| Bank of England FPC (2025) | Goal-directed action using tools | Learning from feedback and adapting to dynamic environments | Autonomous action "to achieve specified goals" |
| FINRA (2025) | Completing tasks on behalf of a user | Lists autonomy, scope and authority, and auditability as risks | Sits inside its generative AI section |
| MAS (2026) | Agents including multi-agent systems | Treats chained actions and tool use as a distinct risk set | Excludes rule-based RPA and fixed-instruction tools |
| FCA Mills Review (2026) | The human's role, in five levels | Frames accountability by where the human sits | Not a prediction that every service moves to full autonomy |
Every figure and regulatory statement in this piece traces to a numbered source below. Links open the original publisher in a new tab.
ISO 42001, NIST, the EU AI Act and MAS agree on the capabilities but not the operating model. NKK's tiering matrix, six pillars and evidence chain turn them into decisions, owners and proof.
Prompt injection remains a residual risk, so agent security is a privilege-and-containment problem. Evidence, controls and three NKK models: a Threat-to-Control Map, Autonomy Tiers and an Agent Identity Stack.
Dated, source-checked status of AI regulation across the EU, US, UK, India, China and Asia-Pacific as of 13 October 2026, with three NKK models for managing regulatory change.
