Agentic AI in banking is the use of AI agents that complete banking tasks, such as processing a loan application or moving money, rather than only answering questions. TD Bank Group's first agentic AI model, described in a May 2026 release, automates mortgage and HELOC applications, and Broadridge said its agentic AI, software that "autonomously analyzes, prioritizes, and resolves operational exceptions," was live in capital markets and wealth workflows (TD Bank Group; Broadridge).
​

This article covers where banks use agents, what agentic banking infrastructure looks like, and the controls that decide whether an agent can move money.
​

Agentic AI in banking refers to AI systems that take multi-step actions inside a bank's processes, such as gathering documents, checking rules, and completing a transaction. Earlier banking AI mostly classified or answered. Agentic systems act, which is why banks pair them with policy checks, audit trails, and human review for high-risk steps.
​

Industry coverage shows banks moving from pilots toward scale in 2026 (Banking Dive). Tearsheet's survey of bank approaches describes Grasshopper Bank planning agents that operate beyond the rules-based automation bots it already uses (Tearsheet).
​

Banks use AI agents in four main areas: lending operations, operational exceptions, customer servicing that executes transactions, and agent-initiated commerce. Each area has a different risk profile, so banks deploy agents first where errors are reversible and reviewable, then extend to areas where agents move money.
​

Agentic banking infrastructure is the layer that lets an AI agent take banking actions safely: authenticated requests, a policy engine that checks each action, connections to the core banking system and payment rails, and a complete audit trail. Without that layer, an agent is a chatbot with access, which no bank risk team would approve.
​

Payman AI's description of its product is a clear example: the agent "acts only when an authenticated customer or member asks," an "immutable policy engine checks every action before it reaches the core," and the institution holds the kill switch (Payman AI). See Payman AI explained.
​

Banks control AI agents with layered limits: scoped permissions, policy checks before execution, human approval for high-value or unusual actions, monitoring, and the ability to stop agents instantly. The principle is the same one that applies to crypto agents: limits that must hold are enforced outside the model.
​

Onchain, that enforcement can sit in the transaction itself. Stablecoin payments made by agents can carry conditions checked at execution, which is the approach covered in AI agent spend controls and stablecoin payments for AI agents. For the treasury side, see AI agents for stablecoin treasury.
​

Agentic AI in banking and fintech differ mainly in where the agent sits. Banks deploy agents inside regulated core systems with institution-owned policy. Fintechs often build agents on top of bank partners or onchain rails, where controls live in their own software, partner agreements, or smart contracts.
​

For wealth and corporate finance uses, see agentic AI in wealth management and AI agents in corporate finance.
​

Agentic AI works in banking by connecting a language model to the bank's own systems through a governed layer that checks every action before it reaches the core. A customer or staff member asks for something in plain language, the model works out which steps are needed, a policy layer checks each step against the bank's rules and the person's permissions, and only then does the system execute a payment, transfer or update.
​

Payman AI's public description of its product shows this layered design in concrete terms. It lists "Authentication: Every action tied to a verified human," "Policy: Immutable layer the model cannot override or reason around," a tamper-evident audit trail, and a bank-held kill switch (Payman AI). It also states that the regulated institution keeps custody, license and KYC, and that the product runs on the institution's existing rails rather than replacing its core.
​

The pattern generalizes beyond any one vendor. Every agentic banking system needs answers to four questions:

Implementing agentic AI in a bank works best as a staged rollout that starts with low-risk, read-only tasks and adds money movement only after controls are proven. A practical sequence looks like this:

The order matters because each step produces evidence that the next one is safe. Skipping straight to execution means the first real test of the controls happens with customer money.
​

Agentic AI success in banking is measured by whether tasks get done correctly, safely and with less effort, not by how many conversations the agent has. A useful scorecard covers four areas:
​

Two measures are easy to miss. The first is the count of actions the policy layer refused: zero refusals over a long period may mean the policy is too loose, not that the agent is perfect. The second is the time staff spend reviewing agent work, which should fall as trust builds; if it does not, the scope may be wrong.
​

Agentic AI use cases in banking split into customer-facing, staff-facing and back-office work. The risk profile differs in each, so controls should differ too.
​

Agentic banking and stablecoins meet where an agent needs to move dollars outside bank hours or across borders. Bank rails have cutoffs and settlement windows; stablecoins settle on blockchains that run continuously. A bank that offers stablecoin services can let an agent use either rail depending on the task, but the control questions stay the same: who asked, is it allowed, does a person confirm, can it be traced.
​

Onchain, part of the policy can move into the transaction itself. A program can check the amount, the destination and the minimum delivered, and refuse to settle if any condition fails. Eco's Sauce SDK is built for that: the launch post describes institutions writing requirements such as minimum delivered amount and approved destinations into the transaction (Eco Sauce launch post). That does not replace the bank's own policy engine; it adds a check at the last possible point, inside the transaction, on one chain. For more on rails, see agentic payments.
​

The most common mistake is putting the rules in the prompt. A model told "never move more than the customer's daily limit" can still be confused or steered; the limit has to live in a policy layer the model cannot edit. The second is launching with money movement before read-only use has shown how often the agent misunderstands requests. The third is measuring volume instead of outcomes: many conversations with many handoffs to staff is not success.
​

Other mistakes show up later. Logs that record only what the agent did, and not what it proposed and was refused, hide the near misses that matter most for risk review. Approval steps that fire on every action train customers and staff to click through without reading. And a kill switch that nobody has tested is not a control. Each of these is cheap to fix before launch and expensive to fix after an incident.
​

Fintechs tend to adopt agentic AI faster than banks because they own fewer legacy systems and often build on modern APIs. Banks tend to move slower because each new system must pass model risk, compliance and vendor review. Many deployments end up as partnerships: a fintech or vendor provides the agent layer, and the bank keeps custody, licensing and the customer relationship, as in the Payman model described above. Whichever side builds it, the infrastructure needs the same parts: identity, policy, approvals, logging, and a way to switch it off.
​

Yes. TD Bank Group launched its first agentic AI model in May 2026 to automate mortgage and HELOC applications, and Broadridge describes its agentic AI as live in production (TD Bank Group; Broadridge).
​

Banking in which AI agents complete tasks such as payments and transfers under institution-controlled policy.
​

Start with one narrow journey, write the policy before connecting the model, run read-only first, then add prepared actions, then limited execution with full logging.
​

Track task completion without handoff, error rates, actions blocked by policy, and customer outcomes such as complaints and time to resolution.
​

No. Agent layers typically sit on top of existing cores and payment rails and call them through governed interfaces.
​

The bank. In vendor models such as Payman's, the regulated institution holds funds, license and KYC; the agent acts on its systems under its policies.
​

A chatbot answers questions. An agent can also take actions on the customer's behalf, such as sending a payment or moving money between accounts, which is why it needs authentication, a policy layer, approvals and an audit trail that a chatbot does not.
​

Sources: company announcements and industry coverage linked inline, accessed November 2026.
​