Agentic banking is the use of AI agents that do work inside a bank, not only answer questions. An agent reads documents and systems, prepares files, drafts decisions and, within limits the bank sets, takes actions, while people approve anything that commits the bank or its customer. It covers staff desks such as lending, compliance and fraud, and money advisors that customers meet in the bank’s app.

A short definition

An AI agent is software that is given a goal, a set of tools and a set of rules, and works through the steps to reach the goal. A chatbot answers the question in front of it. An agent can open a loan file, read the bank statements in it, compare the income against the application, draft the appraisal note and put it in front of the credit officer for a decision.

Agentic banking is what happens when a bank runs many of these agents across its desks, under one set of controls. The word “agentic” is about agency: the ability to act. The banking part is about where that agency is allowed to go, and who decides.

What an agent does that a chatbot does not

Three things set an agent apart:

  • It works in steps. It plans, calls a tool, reads the result, and decides what to do next. A file that needs ten checks gets ten checks.
  • It uses the bank’s systems. Through connectors to core banking, loan, CRM, KYC and document systems, it reads what it needs and, where allowed, writes back.
  • It hands work on. On a well-run desk, a lead agent splits a job into parts, gives each part to a specialist agent, and checks what comes back.

That is why the question for a bank shifts from “is the answer right?” to “what is this agent allowed to do, and who signs off?”. We compare the three tools banks already know in agents vs chatbots vs RPA.

Where agents work in a bank

Most of a bank’s time goes into preparing work for a decision. That is where agents earn their place first:

  • Lending. Document intake, statement and GST analysis, appraisal notes, credit memos, pricing, deviation checks, sanction letters, disbursal and early warning. See AI agents for loan origination and credit memos.
  • Fraud, disputes and KYC. Chargeback files, fraud claim investigations, mule account alerts, KYC and onboarding files verified at source, beneficial ownership and periodic KYC refresh. See AI agents for KYC, fraud and dispute operations.
  • Compliance. Circulars mapped to policies, regulatory returns prepared, AML alerts triaged and an inventory of the bank’s AI systems kept.
  • Trade and payments. Trade document checks, LC and guarantee validation, payment repair and reconciliation.
  • Wealth and treasury. Research, suitability checks, portfolio reviews, cash forecasts and client meeting packs.
  • Service and collections. Answers from the bank’s own documents, call wrap-up, complaint letters, contact plans and hardship files.
  • Legal and bids. Contracts checked clause by clause against the bank’s playbook, and RFP responses drafted from approved answers.

On the customer side, the agent is a money advisor in the bank’s app. It explains, plans and, inside limits the customer and the bank both set, prepares or carries out money actions. We cover what to look for in AI money advisors for bank customers.

The control layer is the product

A bank can buy capable models from many places. What makes agents usable in a regulated business is the layer that decides what each agent may do. In practice that layer has six parts:

  1. Autonomy levels. Each agent holds a level. Below it, the agent acts. Above it, a person decides. A person approves every step up.
  2. Limits and budgets. Caps on amounts for customer actions, and daily and monthly spend budgets for each agent.
  3. Approvals. A named person approves on a card, with the reason written down.
  4. Guardrails. What always needs a person, what each agent may see, and rules every agent follows.
  5. A kill switch. Any agent can be paused at once, and going over budget stops it too.
  6. A record. Every action and decision, with who asked, what ran and why.

We go through each one in how banks govern AI agents.

Teams, not single bots

Banking work is rarely one task. A corporate credit file needs financial spreading, borrower research, a memo and a check of the security documents. Agentic banking works best when agents are organised the way the desk already is: a lead agent for the desk, specialists under it, and a shared room where the desk’s people see the work, steer it and approve it.

That shared room matters. It keeps the case, the files and the decisions in one place, so the next agent and the next person start from what was already agreed.

What stays the same

Agentic banking does not replace a bank’s core systems or its people’s judgement. The core stays the system of record. Agents connect to it through APIs, scoped to what each one needs. Credit sanctions, fraud liability, suitability, filings and anything a customer will see remain a person’s decision. What changes is how much of the preparation is done before that person picks up the file.

How banks start

The banks that move fastest start narrow. They pick one desk or one customer journey, set the rules and limits, run the agents in a sandbox on their own documents, then open it to a small group with the record reviewed every week. Six to eight weeks is a realistic span for that first pilot. See how a pilot works.

Where Tuvel fits

Tuvel is an agentic banking platform for banks, NBFCs, wealth managers, insurers and fintechs. It has five products on one control plane: Tuvel Advisor for customers, Tuvel Agent Teams with 99 staff agents on 14 desks, Tuvel Rooms, Tuvel Control Plane and Tuvel Launch. Each bank gets its own private copy. Being deployed with a Fortune 500 US bank and one of India’s three largest banks.