Can an AI agent recommend a savings or loan product?

It depends on what "recommend" means. An AI agent can explain the bank's savings and loan products, compare them against needs the customer states, and show which ones the customer is eligible for. A personal recommendation that amounts to regulated advice stays with a qualified person, and a loan is never suggested to someone showing financial difficulty.

Where is the line between information and advice?

Information tells the customer what exists and how it works: this account pays this rate, needs this notice, allows this many withdrawals. A comparison filtered by what the customer says they want is still information: "You said you need access within a week, so these three accounts fit." The customer decides.

Advice is different. It is a personal recommendation that a specific product suits this customer's circumstances. For some products, such as investments and, in many markets, mortgages, giving advice is a regulated activity with qualification and suitability rules. Even where savings accounts and personal loans are not covered by advice rules, the bank's duty to deliver good outcomes still applies; in the UK, the FCA's Consumer Duty is the clearest example.

Why are loans riskier to suggest than savings?

A savings product that is not quite right costs the customer some interest. A loan that is not right can cost them far more. The danger is sharpest when an agent uses account data to target offers: the customers whose transactions show them running short each month are the ones most likely to respond to a loan offer, and often the least able to repay it.

So the agent treats loan suggestions with a second filter. Before showing any credit product, it checks for signs of financial difficulty: repeated overdraft use, missed payments, gambling transactions, payday lending. Where those appear, the agent does not suggest borrowing. The customer is offered help instead, and a person can follow up.

What may the agent do, and where must it stop?

• Explain any product the bank offers, from its approved product information, at depth 1.
• Compare products against needs the customer states, showing the trade-offs, at depth 2.
• Check eligibility using the bank's published criteria.
• Open a savings account the customer has chosen, at depth 3, after showing the key terms.
• Start a loan application the customer asks for, with the full process and checks that follow.
• Stop before saying "this one is right for you" on any product covered by advice rules.
• Stop before suggesting credit to a customer with difficulty signals.
• Stop before ranking products by the bank's margin rather than the customer's stated needs.

Running a savings goal once the account is open is covered in whether an AI agent can manage a customer's savings goals.

An example: a customer with spare cash and a credit card balance

Take a customer who asks, "I've got 3,000 spare, which savings account should I use?" The agent reads that they also carry a credit card balance of 2,000 at a high rate.

The agent compares the bank's savings accounts that fit the customer's stated need for access within a month, and shows their rates. It also states a fact from the bank's approved guidance: interest on credit card balances is usually higher than interest earned on savings, and some customers choose to pay down debt first. It does not tell the customer what to do. The customer decides to pay off the card and save the rest, and the agent opens an instant-access account for the remaining 1,000.

The record shows the products compared, the facts shown and the customer's choice.

What controls make it safe?

• Product facts arrive sealed. Rates and terms come from the bank's product catalogue, versioned, so the agent never quotes a stale rate.
• Needs stated by the customer. Filters use what the customer says, recorded as declared, not assumptions from their data.
• Difficulty screen before credit. Signals of financial difficulty suppress loan suggestions and route the customer to support.
• No advice capability. Certification includes customers pressing for "just tell me which one", and the agent must decline.
• Ranking by fit. Products are ordered by the customer's stated criteria, never by margin.

If the customer goes ahead with a loan, see processing personal loan applications with an AI agent. Customers already struggling with a loan need loan restructuring, not another product. The general limits are in AI agents in banking: what they may and may not do.

Where Pryme fits

In Pryme, explaining products and opening a chosen savings account are separate capabilities with separate depths. Attest replays the agent on the bank's own past product conversations, and the adversarial gate fails any run where the agent suggests credit to a customer showing difficulty, a state where the efficient answer is the wrong one. Product facts carry sealed provenance from the catalogue, so a quoted rate can always be traced to its version.