AI for Personal Finance Admin, Not Advice

Sorting your own records into categories is admin, and a model is good at it. Telling you which pension to choose is regulated financial advice, and nothing in a chat window is authorised to give it. That boundary is the entire structure of this page.

The line, and why it is a legal one

In the UK, the EU and most comparable jurisdictions, giving a personal recommendation about a specific investment, pension or insurance product is a regulated activity. Firms doing it need authorisation, they carry professional indemnity insurance, they are obliged to know your circumstances, and when they get it wrong there is a complaints process and a compensation scheme behind them.

None of that applies to a chatbot. There is no authorisation, no duty to you, no record of your circumstances and no redress. This is not a statement about how clever the model is; a very capable model in this role is more dangerous than a poor one, because it is more convincing.

A useful test: if the answer would change depending on facts about you that were never supplied, and it did not ask, it is not analysis. It is a plausible sentence.

What counts as admin

• Categorising transactions from a bank export into groups you defined. The method is in the next section.
• Extracting figures from statements and letters: annual charges, interest rates, end dates, notice periods — with the sentence each came from quoted, so you can check it.
• Explaining terminology in general. What an APR is, what a variable rate means, how a salary-sacrifice arrangement works in principle. General explanation is not a personal recommendation, though it becomes one the moment it says “so you should”.
• Drafting correspondence: a complaint, a request for a settlement figure, a cancellation, a subject access request. Ordinary letter-writing that happens to be about money.
• Preparing to talk to a professional. Producing the list of questions and the summary of your position to take to an accountant or an adviser. This is the highest-value use on the page, because professional time is expensive and most of it is spent establishing facts you could have arrived with.

Categorising your own transactions

Export a statement as CSV. Remove the account number and sort code first — you do not need them for this and they should not leave your machine.

Below are 300 transaction lines: date, description, amount.

Assign each to exactly one of these categories, and no others:
  housing, utilities, groceries, transport, insurance, health,
  childcare, subscriptions, eating out, one-off, transfers, unknown

Rules:
- Use "unknown" freely. A wrong category is worse than an unknown
  one, because I will not review the ones that look right.
- "transfers" is money moving between my own accounts. Do not count
  these as spending.
- Do not change, round or re-total any amount.
- Do not add commentary about my spending.

Output: the original line, unchanged, plus two columns - category
and confidence (high/low).

Then: review every low-confidence row and everything marked unknown, spot-check twenty high-confidence rows, and paste the categories back into your spreadsheet. The totals are computed by the spreadsheet, not by the model, for the reason in the next section.

The instruction against commentary is worth keeping. Unrequested observations about somebody’s spending are both unwelcome and, in the strict sense, uninformed — the model has one month of one account and no idea what else is going on in your life.

Arithmetic belongs in a spreadsheet

Never accept a total from prose. Not because the arithmetic is always wrong, but because a wrong total is indistinguishable from a right one and it is trivially easy to get a real one instead. In your spreadsheet:

=SUMIF(D:D, "groceries", C:C)          total for a category
=SUMIFS(C:C, D:D, "groceries",         a category in one month
        A:A, ">="&DATE(2026,7,1),
        A:A, "<="&DATE(2026,7,31))
=SUMIF(D:D, "transfers", C:C)          check this nets to ~0
=SUM(C:C)-SUMIF(D:D,"transfers",C:C)   real spending

The transfers check is the one people miss: if money moved between your own accounts is counted as spending, every total is inflated and the error is invisible. The wider set of checks for spreadsheet work is in making sense of a spreadsheet.

Understanding a document you were sent

Pension statements, mortgage offers and insurance schedules are written to be complete rather than legible. Extracting the facts is admin, and the quoting rule keeps it honest.

From the attached document only, extract:
- every charge or fee, with the exact wording and the page
- every rate, and whether the document says it is fixed or variable
- every date and deadline
- anything the document says I must do, and by when
- anything conditional: benefits or rates that apply only if
  something is true

Quote the sentence for each. If something is not in the document,
write "not stated" - do not tell me what is typical.

Then list the questions I should ask the provider, based only on
what is ambiguous or missing in this document.

“Do not tell me what is typical” blocks the specific failure here, which is a document-grounded extract with two invented industry-standard figures blended in. The question list at the end is what you take to the provider or the adviser.

An annual routine that takes an afternoon

The pieces above compose into something worth doing once a year, all of it on the admin side of the line, and all of it ending in a document you own rather than a recommendation you were given.

1. Gather the documents. The annual statement from every account, policy and pension you hold. This is the step that actually takes the time, and no tool shortens it.
2. Extract from each, using the quoting prompt above: charges, rates, dates, conditions, and what you must do by when. One extract per document, saved as plain text.
3. Build one table yourself with a row per product: what it is, what it costs a year, what rate it pays or charges, when it ends or renews, and what the notice period is. Your spreadsheet, your arithmetic.
4. Categorise a year of transactions and total each category with SUMIF, checking that transfers net out. You now have income, outgoings and the shape of both.
5. Write the question list. From the extracts, the things that are unclear, expiring, or apparently expensive relative to their siblings in the table. This is what a professional is for, and arriving with it converts an hour of their time into something worth paying for.

Notice that no step produces a recommendation. The output is a table, a set of totals and a list of questions. Every judgement about what to do with any of it stays with you or with somebody authorised to give it.

What not to ask, and what to ask instead

Your bank data is not ordinary data

A transaction history is one of the most revealing documents about a person that exists: where you live, your health, your relationships, your employer, when your house is empty. Treat pasting it as a decision rather than a convenience.

1. Remove account numbers, sort codes and card numbers before anything leaves your machine. Nothing on this page needs them.
2. Consider replacing merchant names with categories yourself where they are sensitive. A row that reads 2026-07-14 | clinic | 240.00 carries a great deal less than the clinic’s name.
3. Check whether the product uses your input for training and how long it retains it. These are two separate promises and most services make only one of them — the distinction is set out in using AI without handing over your data.
4. Never connect a tool directly to a bank account without reading what access you are granting and how it is revoked. Read-only access to balances and access to initiate payments are very different permissions.

If any of this feels like too much for the benefit, that is a reasonable conclusion. Categorising a year of spending by hand takes an afternoon, and there is a whole category of tasks where the honest answer is that the tool is not the right one — that list is in the tasks where AI is the wrong choice.