How to run income and expenditure assessments with an AI agent

The agent helps the customer complete the assessment and checks it; a person decides what is affordable. Using approved wording, it guides the customer through income, essential spending and other debts, compares the figures with account data where the customer agrees, and flags anything unusual without overriding it. A named person approves the assessment and the payment it supports.

What is an income and expenditure assessment, and who owns it?

An income and expenditure assessment, often shortened to I&E, sets out what a customer in financial difficulty earns, what they must spend on essentials, what they owe elsewhere, and what is left. It is the basis for a fair repayment plan or a decision that no payment is affordable for now. Done well, it protects the customer from promising more than they can pay. Done badly, it becomes a tool for squeezing a payment out of someone who cannot afford it.

Collections and financial support teams own the assessment. Many lenders use a standard financial statement format with reference figures for typical spending. Under the FCA Consumer Duty in the UK, and similar conduct expectations elsewhere, firms must be able to show that customers in difficulty were treated fairly. The agent's job is to make the assessment complete and accurate, not to make the number bigger.

What does the agent prepare, and what does a person approve?

1. Agent, depth 1: checks for an existing assessment. If a recent one exists, it asks the customer what has changed rather than starting again.
2. Agent, depth 3: guides the customer through the form. In the lender's digital channel, using approved wording to explain each question. It does not suggest figures.
3. Agent, depth 1: listens for vulnerability. Mentions of illness, bereavement, job loss, mental health or a crisis route the customer to a person straight away, with the conversation passed across.
4. Agent, depth 1: checks figures against evidence. With the customer's agreement, it compares declared income with credits in their account. Whether an agent can verify income from bank statements depends on consent and on the lender's rules.
5. Agent, depth 1: compares spending with reference figures. Figures well outside the reference range are flagged with a question, not changed.
6. Agent, depth 2: drafts the assessment summary. Income, priority spending such as housing and energy, other essentials, other creditors, and the disposable figure, with the options policy allows.
7. Person: approves. The assessment, any queries back to the customer, and the payment amount or a decision that nothing is affordable now.
8. Agent, depth 2: drafts the outcome message, including where to find free, independent debt advice.

Which controls apply, and what evidence do they leave?

Customer figures are not overridden. If a customer says their food costs exceed the reference figure, the agent records it and may flag it for a question. It never replaces the customer's figure with the reference figure. Evidence: the declared figure, the reference figure and any query, kept side by side.

Priority spending comes first. The summary lists housing, energy and other priority costs before other creditors, so the disposable figure reflects what is left after essentials.

Consent for account data. Comparing declared income with account data needs the customer's agreement. Evidence: the consent record filed with the comparison.

Approval by a named person. The assessment is filed as held until approved. The record shows who approved, any changes they made and why.

Vulnerability routing. Every routing to a person is filed with the trigger. If the agent kept going with a customer who mentioned a crisis, the record will show it, and the case goes into the next certification set.

Where does it usually go wrong?

Reference figures used as caps. A lender reduces the customer's stated spending to the reference figure to increase the payment. That turns a fairness tool into a collections tool. The agent's bounds should refuse any edit to a declared figure.

Irregular income averaged away. Shift workers, the self-employed and people on variable hours have income that moves month to month. A single average hides the bad months. The summary should show the range.

Stale assessments. An assessment from months ago is reused because it is on file. Circumstances change; ask what has changed every time.

Other creditors missed. A plan that is affordable for your debt alone may not be affordable once other creditors are paid. Ask about all debts.

Checking the wrong things. Checking income evidence is useful; checking spending against the customer's account in detail can feel intrusive and is often not needed. The approach to evidence in checking income and affordability evidence applies, but in collections the customer's own account usually comes first.

What should you measure?

• Plans set from agent-supported assessments that are kept. Broken plans suggest the amount was not affordable.
• Assessments changed by the approving person, by direction.
• Customers routed to a person for vulnerability. Too few may mean the agent is missing signs.
• Complaints about assessments.
• Cost per completed assessment. Useful alongside the wider question of how much it costs to run AI agents in a bank, but never as the main measure.

The general bounds for bank agents are in AI agents in banking.