What agent-ready really means for financial services

Financial services firms have good reason to be careful with AI agents. They hold sensitive information, their customers rely on them for consequential decisions, and they answer to regulators for how they behave. Caution is appropriate.

Caution is not the same as waiting, though. The firms making steady progress are not the ones with the largest technology budgets. They are the ones that have worked out what readiness means in their context and built towards it deliberately.

This article sets out how we think about that readiness. It is a business perspective, not legal or regulatory advice. Every firm should take its own guidance from its legal, risk and compliance teams on the specific obligations that apply to it.

Constraints are inputs, not obstacles

In many programmes, compliance arrives at the end. The agent is designed, a pilot is ready, and the risk function is asked to approve it. The answer is often no, or not yet, and months of work are revisited.

The alternative is to treat compliance, consent and privacy as design constraints from the first day. They shape what the agent may do, what data it may use and where a person must be involved. Designed in early, they narrow the options in a useful way. Added late, they break things.

In a regulated business, the constraints are part of the brief. An agent designed without them has not been designed.

Four dimensions of readiness

We look at readiness across four dimensions.

Data and consent. Does the firm know what customer data it holds, where it sits, and what each customer has agreed to? An agent that personalises a conversation needs to respect the purposes for which the data was collected. Consent and preference records need to be accurate and available at the moment the agent acts, not stored in a system it cannot see.

Boundaries of action. Is there a clear line between what an agent may do alone, what it may prepare for a person to approve, and what it must never do? In financial services, the line between giving information and giving advice deserves particular care, as do any actions that move money or change a customer's terms.

Explainability and record. Can the firm show, after the event, what the agent did and what it relied on? A full record of inputs, actions and handoffs supports complaints handling, internal audit and continuous improvement. If a decision cannot be explained, the agent should not be making it.

Accountability. Is there a named owner for each agent, an agreed oversight model, and a route for concerns to be raised and acted upon? Existing governance structures can often be extended to cover agents. Leaving agents outside those structures should not be an option.

A firm does not need to be strong in all four across the whole organisation. It needs to be strong in all four for the specific use case in front of it.

Where to begin

The sensible starting points are tasks with real value and contained risk. Typically these are informational or administrative, and many are internal.

• Answering routine servicing questions: balances, statement requests, the status of an application.
• Preparing a relationship manager for a client meeting with a summary of recent activity and open items.
• Gathering and checking documents during onboarding, with a person making the decisions. None of these asks the agent to exercise judgement on a customer's financial position. Each takes administrative effort away from skilled people, and each produces evidence about how agents perform in the firm's own environment.

Tasks involving advice, credit decisions or vulnerable customers call for much more caution and, in our view, for a person to remain the decision-maker.

Privacy as a matter of design

Privacy is easier to design in than to retrofit. A few principles carry most of the weight.

1. Minimise. Give the agent the data the task requires and nothing further.
2. Limit by purpose. Data gathered for one purpose should not quietly be used for another.
3. Control access. An agent should see no more than a person in the equivalent role would see.
4. Keep customers informed. People should know when they are dealing with an agent and how to reach a person.

Platforms such as Salesforce provide controls for access and permissions. The decisions about how to set them belong to the firm.

Readiness is a position, not a project

Readiness is not a state that is reached once. Products change, customer expectations shift, and the rules a firm operates under develop over time. What matters is having the structures to adapt: clear ownership, current data and consent records, an oversight model that is actually used, and a working relationship between the business, technology and risk functions.

Firms with those structures can move at a measured pace and keep moving. That is the kind of readiness we help leadership teams assess, one use case at a time.