An internal one that reads and drafts for staff. The best first agent works on a high-volume process with well-recorded past cases, produces output a person checks before anything happens, and touches no money. Payment investigations, a procedures assistant for frontline staff, and complaint file preparation all fit. Customer-facing and money-moving agents should come later.
What makes a good first job?
A first agent has two purposes: to save time on real work, and to build the bank's ability to test, approve and run agents. The second matters more. A first job that teaches the bank how to do this well pays off for every agent after it.
Score candidate jobs against these tests:
| Test | Why it matters for a first agent |
|---|---|
| High volume, repetitive | Enough cases to test and enough saving to notice |
| Past cases with known outcomes | You can replay them and score the agent |
| Output a person checks before it takes effect | Mistakes are caught cheaply |
| No money moves, nothing hard to undo | Lower validation burden, faster approval |
| Internal users, not customers | Staff can give fast, specific feedback |
| One owning team | One approver, one place to measure |
The general version of this reasoning is in which process you should give an AI agent first.
Which candidates usually win?
A procedures assistant for frontline staff. Branch and contact centre staff spend time finding the right procedure for unusual requests: a power of attorney, a deceased customer's account, a foreign cheque. An agent that reads the bank's policy library and answers with the relevant section, cited, saves time on every such case. It acts on nothing. It also helps resolve more queries on the first contact, which is the measure explained in what first-contact resolution is.
A payment investigations assistant. It reads the payment record and the inquiry, drafts the trace request or reply, and a person sends it. Cases are well documented, outcomes are known, and the work is tedious enough that staff welcome help.
Complaint file preparation. It gathers the account history, contact records and relevant terms for each complaint, and drafts a summary for the handler. The handler decides and writes the response.
All three are depth 1 or 2: read, or draft for a person to send.
Why should frontline staff come before customers?
Because staff can tell you when the agent is wrong, and customers mostly cannot. A branch adviser who sees a cited procedure can check it in seconds. A customer told the wrong thing may act on it. The case is made in full in frontline staff are the first users of bank agents.
Staff use also builds the evidence base for a customer-facing agent later. The questions staff ask, and where the agent's answers needed correcting, become test cases.
What should a bank avoid first?
- Credit decisions on individuals, which carry legal effect and, in the EU, sit in the high-risk category under the AI Act.
- Anything that releases payments.
- A general-purpose assistant connected to many systems with no defined scope.
- A customer chat agent with no proven grounding in the bank's own data.
- Financial crime decisions, such as closing alerts or deciding on reports.
Each of these can be a good later agent. None is a good first one.
An illustrative example
A mid-sized bank shortlists three first agents: a mortgage pre-approval chat, a fraud alert triage agent and a procedures assistant for its contact centre.
The mortgage chat is customer-facing and touches credit, so it is set aside. Fraud triage has good data but sits close to decisions on customer funds and regulatory reports. The procedures assistant wins: high volume, a policy library it can cite, staff who will flag errors, and no action on any account. It goes live in one contact centre team, and its answers are tested against questions the team logged over the previous quarter. Six months later the bank uses the same test methods, approval route and record to deploy the payment investigations agent.
What stays with people in the first agent?
Everything that changes an account, every message that goes to a customer, and every decision. In the first deployment, the agent's job is to make a person's next step faster and better informed. A matrix of which bank operations suit which depth, useful for planning the second and third agents, is in the bank operations depth matrix for AI agents.
How long until the first agent is live?
A read-and-draft internal agent moves fastest through access, testing and approval. The factors that stretch or shorten it are set out in how long it takes to deploy an AI agent in a bank. The wider limits are in AI agents in banking: what they may and may not do.
Questions people ask
Is a customer chatbot a sensible first agent?
Rarely. It is visible, its mistakes reach customers directly, and it needs grounding, disclosure and escalation all working at once. Start internally and bring the lessons across.
Should the first agent be built or bought?
Either works if the bank can test it on its own cases and read every decision it makes. The first agent is as much about building that capability as about the agent.
How do we know the first agent succeeded?
Staff use it, its drafts need less correction over time, and the bank has a tested route for approving the next one.