How to Use AI to Manage Your Money

Ask a chatbot how to pay off three credit cards and it will tell you to attack the highest interest rate first. This is the avalanche method, it minimises total interest paid, and it is arithmetically correct.

It is also, for a lot of people, the wrong advice — and the reason why is one of the more useful findings in consumer finance.

In Can Small Victories Help Win the War? Evidence from Consumer Debt Management (Journal of Marketing Research, 2012), David Gal and Blakeley McShane analysed real customer data from a debt settlement firm. They found that the fraction of accounts a person had closed predicted whether they eliminated their debt — while the dollar balance of what they closed did not. Paying off small accounts, whatever the interest rate, tracked with going the distance.

The mathematically optimal plan is worth nothing if you abandon it in month four. A model optimising the spreadsheet cannot see that, because adherence isn't in the spreadsheet.

This is the general shape of the problem. AI is genuinely useful for money — and its characteristic failure is confidently optimising the wrong thing, in a voice that sounds like a professional.

Make it show both plans and argue against itself

The fix is to force the trade-off into the open rather than accepting the first recommendation:

Here are my debts: [balances, rates, minimums]. Show me both the avalanche and the snowball plan side by side. For each: months to payoff, total interest, and which debt disappears first. Then tell me what each plan costs me — avalanche in motivation, snowball in interest — and don't recommend one until you've asked me how previous payoff attempts went.

That last clause is the one that matters. The relevant input isn't in your balances; it's in whether you have quit before. This is why the personal finance pack puts both methods in front of you and frames the choice as the one you'll actually stick with.

Assume every number it recalls is out of date

A model's knowledge has a cutoff, and personal finance is dense with figures that change annually — contribution limits, tax brackets, standard deductions, benefit thresholds, rates. It will state last year's number in this year's confident present tense, with no signal that anything is stale.

It also defaults to US assumptions unless told otherwise, so a UK or Canadian user gets advice about accounts they cannot open.

Two habits handle almost all of this:

• Paste the number, don't ask for it. Look up the current limit on the official source, put it in the prompt, and let the model do the reasoning around it. Reasoning is what it's good at; recall of volatile figures is what it's worst at.
• State your country and currency in the first line. Every time.

And the guardrail:

If any figure you need is time-sensitive or jurisdiction-specific, do not state it from memory — mark it [VERIFY: ...] and tell me where to look it up.

Check the arithmetic, always

Models are far better at explaining compound interest than at computing it. Multi-step arithmetic across many rows — amortisation, a payoff schedule, a savings timeline — is where errors appear, and they appear inside beautifully formatted tables that read as authoritative.

Spot-check with the crudest possible sanity test. If a plan says £400 a month clears £14,000 in 30 months, multiply: 400 × 30 = 12,000. That's short of the principal before any interest, so something is wrong. That one check catches most of what goes wrong, and it takes five seconds.

Better still, ask for the formula and inputs alongside the result, then run it in a spreadsheet. Use the model for the setup and the explanation; use a calculator for the numbers.

Where it's genuinely strong

Explaining without selling. Almost every plain-language explanation of a financial product online is published by someone who profits from your choice. A model has no commission. Explain what this is, how it makes money, what it costs me, and what the catch is is a question with no unconflicted answer elsewhere on the internet, and this is the single best use case in the category.

Making the vague concrete. Save for a house becomes a target figure, a monthly number and a date. The arithmetic is trivial; it's the sitting-down that's hard, and a conversational tool lowers that barrier more than a spreadsheet does.

Finding the recurring drains. Categorising a year of transactions and surfacing the subscriptions you forgot is mechanical work that reliably finds real money.

Rehearsing the call. Practising how you'll ask for a lower rate, dispute a charge, or negotiate a bill — with the other side's likely responses.

A word about your data

Before pasting a bank statement into a chat window, strip it: account numbers, full names, addresses, card digits. You need the amounts, dates and merchants for the analysis; you need none of the identifiers. Check whether the tool you're using trains on your conversations, and turn it off if it does.

The line

AI can do the arithmetic, the explaining, the categorising and the planning. It cannot know your risk tolerance, your job security, what your family expects of you, or whether you'll still be doing this in June.

Any output that depends on those is a draft for you to judge — and for anything with real, irreversible consequences, a starting point for a conversation with a human who is paid a flat fee rather than a commission.

---

Source: Gal, D. and McShane, B. B., Can Small Victories Help Win the War? Evidence from Consumer Debt Management, Journal of Marketing Research, Vol. XLIX (August 2012), pp. 487–501.