Skip to content
Behind the Bank

🤖 Agentic Banking

AI agents are moving from answering questions to doing things: moving money, drafting credit memos, triaging outages. Agentic Banking tracks where banks and fintechs actually deploy them, what the agent is allowed to do on its own, and what the firm says it changed. Public, linked sources only. Company claims are labelled as such; our read is labelled as ours.

Trust Bank · Fintech / digital bank · APAC · Singapore ·

Agents triage IT incidents before the engineer joins the call

What the agent does. Agents built on Amazon Bedrock AgentCore follow a runbook to pull logs and system data, then post a likely root cause and their reasoning in the incident ticket. They recommend; engineers decide the fix.

The numbers. The bank says triage fell from 15–20 minutes to about two minutes, and about 65% of the agents' root-cause analyses are actionable (the bank's own assessment).

Our read. The design choice worth borrowing is the split: deterministic runbooks collect the signals, the model only interprets them. Recommendation, not action, is how you get an agent past an operational-risk committee.

Sources: Computer Weekly, 30 Sep 2026 · Fintech.com.sg, 1 Oct 2026

DBS · Bank · APAC · Singapore / Hong Kong ·

Corporate-banking assistant starts completing tasks

What the agent does. DBS Joy can now complete tasks for 350,000 corporate users in Singapore, with a rollout to 100,000 users in Hong Kong that began in September.

The numbers. In Singapore in H1 2026, Joy's active users rose 61% and calls and emails to customer service fell 7% (the bank's figures).

Our read. AI assistants are common in retail banking and rare in corporate banking, let alone ones that can act. That makes this one to watch.

Sources: DBS newsroom · DBS Hong Kong newsroom

Starling Bank · Fintech / digital bank · Europe · UK ·

A customer-facing assistant that executes banking tasks

What the agent does. Starling Assistant launched for personal accounts in March 2026 and for all business customers in August. It takes voice and natural-language prompts and carries out tasks such as setting up saving goals and organising bill payments. Business owners can ask it to move a share of earnings into a tax Space, and it flags possible invoice fraud. Customers opt in.

Our read. The SME version is the more interesting money story: tax ring-fencing and invoice-fraud checks are jobs small businesses currently do badly or pay an accountant for. Built on Google Gemini and Google Cloud, per the bank.

Sources: Starling, 20 Mar 2026 · Starling, 24 Aug 2026

Nubank (Nu Holdings) · Fintech / digital bank · Americas · Brazil ·

Agents handle most customer-support conversations

What the agent does. Nu reports that AI agents handle more than 60% of customer-support conversations in Brazil "at or above human parity", and its in-house model NuFormer runs underwriting on three loan portfolios.

The numbers. 60%+ of support conversations in Brazil (company-reported, Q2 2026 results).

Our read. Servicing is where Nu's cost advantage transfers best to incumbents. See Issue #002 on its 19.5% efficiency ratio.

Sources: Nu Holdings Q2 2026 results

TD Bank Group · Bank · Americas · Canada ·

An agent writes the mortgage underwriting memo

What the agent does. TD's first agentic model classifies client documents, extracts key information, calculates income, checks figures against select policy requirements and consent, flags discrepancies, then writes a summary memo for the underwriter on mortgage and HELOC applications.

The numbers. TD says early results cut the process from an average of 15 hours to under three minutes (the bank's claim).

Our read. Pre-adjudication is a sweet spot: document-heavy, rules-bound, and a human still signs the credit decision. TD says it has mapped every step of the secured-lending journey for agents next.

Sources: TD Stories, 21 May 2026