An AI agent is software that pursues a goal on your behalf. It reads a request, works out what needs to happen, calls the systems that hold the answer, and carries the case through to an outcome. A model generates language. An agent acts on it.

An AI agent sits above the model that powers it. The model supplies the reasoning and the language; the agent supplies the tools, the permissions, and the stopping conditions that turn reasoning into completed work. An agent is not a chatbot with better phrasing, and it is not a workflow engine with a language layer on top. Both of those follow a path drawn in advance, while an agent decides its own.

How does an AI agent work?

An agent runs a loop rather than a single step. It reads the request, decides what the goal implies, picks a tool, acts, checks what came back, and repeats until the case closes or it decides a person should take over. Four capabilities make that loop possible:

• Reasoning: breaking a request into the steps that would satisfy it
  
  

• Tool use: calling APIs, databases, and internal systems to fetch facts or change state
  
  

• Memory: carrying context across turns, and sometimes across separate cases
  
  

• Stopping conditions: knowing when to act, when to ask, and when to escalate
  
  

A customer saying "I don't recognise this charge" shows what the loop buys you. A chatbot matches that sentence to a help article about disputes and sends it. An agent retrieves the transaction, checks the merchant record, confirms whether the card was present, opens the dispute, and tells the customer what happens next and when.

What can an AI agent do in financial services?

Financial services work splits into cases that end at a reply and cases that end at a system. Agents earn their keep on the second kind, where answering properly means reading a record, applying a policy, and changing something. The territory below is where most deployments start:

• Frontline support: card freezes, payment queries, and account questions needing a live lookup
  
  

• Disputes and chargebacks: intake, evidence gathering, classification, and scheme submission
  
  

• Lending and collections: affordability checks, hardship assessments, and promises to pay
  
  

• Onboarding and verification: document collection, KYC exception handling, and business verification
  
  

Language is rarely the constraint. What decides whether an agent is useful is whether it can reach the systems holding the answer, and whether your controls let it act on what it finds there. Our guide to deploying AI agents for customer operations covers that groundwork.

What does an AI agent need to run safely in a regulated business?

Autonomy raises the cost of every mistake. A wrong sentence can be corrected in the next message, and a wrong payment cannot, so regulated deployments treat the agent as a control surface rather than a feature. Four things belong in place before an agent reaches live customers:

• Guardrails on every turn, inspecting what the customer says and what the agent is about to send
  
  

• Action limits, naming which tools the agent may call and under what conditions
  
  

• Human oversight for anything carrying potential harm, including vulnerability, complaints, and financial difficulty
  
  

• An audit trail recording every action, source, and reason against the case
  
  

None of this shifts accountability. FCA guidance states that firms using AI remain responsible for complying with its rules, and PRA SS2/21 is explicit that outsourcing does not reduce regulatory responsibility. The agent is your control, and so is the evidence that it worked.

How do you measure an AI agent's performance?

Volume metrics flatter agents, but you should be discerning about those numbers. For example, a high deflection rate tells you conversations left the queue without reaching a human, not that customers got what they came for, which is why the number worth arguing over is resolution rate instead.

Four measures cover most evaluations:

• Resolution rate: cases closed end to end, with no human touch and no repeat contact
  
  

• Deflection or containment rate: cases kept out of the human queue, whatever the outcome
  
  

• CSAT: how customers rated the interaction itself
  
  

• QA score: how the work scored against your own quality bar
  
  

Collections is the exception worth knowing. Buyers there judge on recovery rate and total exposure touched, so a resolution-rate frame misses what they care about. Our resolution rate benchmark sets out how to compare vendor claims without taking the headline number on trust.

Frequently asked questions

Is an AI agent the same as an AI chatbot?

No. A chatbot retrieves and responds, working from content someone wrote in advance. An agent reasons, calls systems, and changes state, which is why it can close a case rather than describe how one gets closed. Our guide on AI agents versus AI chatbots works through which fits a regulated operation.

Do AI agents replace support teams?

Not in the deployments we see. Agents take the volume that follows a known shape, and teams move onto the judgment work: vulnerable customers, complaints, edge cases, and anything where the right answer depends on circumstances no policy anticipated. Headcount plans change, though the human queue does not empty.

What does an AI agent need access to in order to be useful?

Whatever holds the answer. That usually means the core banking or ledger system, the payments or card processor, the CRM or ticketing tool, and the knowledge base. An agent restricted to the knowledge base alone can only ever tell customers what your help centre already says, which makes it no more powerful than a chatbot.

Who is accountable when an AI agent gets something wrong?

The firm. FCA guidance states that firms using AI remain responsible for ensuring compliance with its rules, and outsourcing to a vendor does not change that. This is why the ability to inspect, change, and evidence what the agent did matters more than the length of a vendor's feature list.