An AI budget assistant is most useful when it describes the records you provide, shows its assumptions, and leaves decisions with you. Treat a fluent answer as a draft, not proof. In a 2023 report on chatbots used by banks and other financial companies, the CFPB highlighted risks of inaccurate answers, privacy failures, and blocked access to human help.
Use prompts that name the date range, input fields, desired output, missing data, and verification step. Never ask an assistant to invent transactions, guarantee savings, move money, or replace legal, tax, credit, investment, or safety advice.
Define the job before writing the prompt
Choose one bounded task: summarise categories, find duplicate-looking entries, compare a planned budget with actual totals, or list questions for a professional. A narrow job is easier to verify than “optimise my finances”.
Use a prompt with five explicit parts
| Part | Example | Why it helps |
|---|---|---|
| Scope | Household expenses, 1–30 June | Stops an open-ended answer |
| Input | Date, category, amount, note | Shows what the model can see |
| Task | Flag entries over €100 | Creates a testable output |
| Uncertainty | List missing or ambiguous rows | Prevents false precision |
| Check | Return source transaction IDs | Makes review possible |
Prompt patterns that stay bounded
- Summary: “Group these dated transactions by category; show totals, count, and rows you could not classify.”
- Comparison: “Compare June with May; show absolute and percentage changes, then list the transactions behind each change.”
- Question list: “From this budget, write five questions to ask a tax adviser; do not answer them.”
- Correction: “Show which rows need a human category decision; do not change any amount.”
Supply the minimum data
Remove passwords, full account numbers, identity documents, precise addresses, and secrets. Use a date, category, amount, currency, and a short neutral note. Keep the original file so you can compare the assistant’s summary with the source.
Verify the answer against the records
Check totals with a calculator or spreadsheet, open every flagged transaction, confirm the date range, and ask the assistant to list missing rows. If the answer cannot point back to source data, treat it as an unverified suggestion.
Recognise red-flag questions
A prompt about debt collection, fraud, benefits, taxes, a credit decision, investment, legal rights, or urgent safety needs a qualified human or official channel. An assistant can help you prepare facts and questions; it should not make the consequential decision.
What an AI assistant cannot guarantee
Fluent language is not accuracy. A model may misread a category, omit a row, use stale rules, expose sensitive data, or present an association as a cause. Do not rely on it for autonomous transfers, guaranteed forecasts, legal conclusions, or regulated advice.
- Never paste credentials, one-time codes, or private keys.
- Do not accept an unexplained total or forecast.
- Keep a human review for high-impact decisions.
Choose a non-AI workflow when it is clearer
A spreadsheet is often better for a small, auditable budget. A bank or official service is better for account disputes and rights. An AI assistant is optional when it saves explanation time without replacing the source record.
How Lumy fits the verification loop
Lumy, which publishes this guide, has its own AI chat, so read this section as a vendor's description. Lumy has no bank connection and never asks for banking credentials. Its AI chat reads the transactions, budgets, and goals already in your account before it answers. Lumy gives information, not financial advice, and does not make decisions for you; it cannot guarantee a generated summary, forecast, or recommendation. On the free plan, the AI chat gets 10 requests a month (a typed message, a voice message, or a receipt photo each uses one), and Lumy Pro removes that limit. Check its answers the way you would check any other assistant's.
Your next step: test one prompt on ten rows
Redact ten ordinary transactions, write the five-part prompt, and compare the result with a hand-calculated total. Record one correct observation, one omission, and one question that needs a human.
Sources and methodology
The prompt patterns are practical recommendations. The CFPB's 2023 report on chatbots in consumer finance supports the cautions about accuracy, privacy, security, and human escalation. The European Data Protection Board's ChatGPT Taskforce report (2024) notes that ChatGPT may produce biased or made-up outputs that users are likely to take as factually accurate, and that OpenAI stated it used prompts and uploaded files to train and improve the model, with an option to opt out.
- CFPB — Chatbots in consumer finance
- EDPB — Report of the work undertaken by the ChatGPT Taskforce (2024)
Prepared and reviewed: 31 July 2026.