Artificial intelligence in your work

What to know and what you will need

Before you start: clients, the recommendation, the figures

Financial advice works with other people's money and choices, under precise rules. Five rules apply to every case below.

• Your clients' data stays yours. Before you upload a statement, a portfolio or a questionnaire, remove the name, National Insurance number, account numbers and sort code, and replace them with a label ("client A"). In the tool's settings, check whether conversations are used to train the models and switch it off; if your firm has a professional plan, read what it says about data (fact sheet "If the tool comes from your office"). Client personal data is covered by the UK GDPR, and the ICO publishes guidance on AI and data protection.
• The recommendation stays yours. The model prepares, explains, compares and checks; the suitability assessment, the personal recommendation and what the client receives with your name on it are decided by you. When an answer contains "I recommend buying" or "it would be better to sell", that is not the work you asked for.
• Calculations are done with a program. Returns, weights, costs over several years: always ask the model to calculate them with a program and show you the formulas. In the tests of the other fields, the first mistake with figures was a calculation done in its head without saying so.
• Every market figure has a source and a date. An index level, a rate or an inflation figure without a date is not used with a client. If the model cannot find a figure, it must write "not found".
• Responsibility stays yours, with the regulator too. The FCA has said that it does not plan extra rules for AI and applies the existing ones, the Consumer Duty and the accountability of senior managers among them. In "AI in financial services: good practice and areas for improvement" (19 October 2026) it writes that firms using AI-supported tools "remain responsible for consumer outcomes", and that they should tell consumers how they have used AI and its limits. Those examples do not create new requirements. If your client agreement says how you use AI, the line must be true.

1. Preparing the annual review with the client

Read first. Stages 3 (a document of your own), 7 (Excel: formulas, data and charts).

The situation. You are seeing the client in a week. You have the portfolio statement at the end of the year and the one from a year ago, and you want to arrive with one clear page.

The request. Upload the two statements and the year's transactions, without names, then:

"These are two statements of the same portfolio, one year apart, and the transactions for the year. Calculate with a program, showing the formulas: the weight of each holding and of each asset class (equities, bonds, cash, other) at the two dates; the return for the year, net of the payments in and withdrawals shown; the three holdings that contributed most to the result, positive and negative. Then one page for the meeting: what has changed and why, in plain words, with no recommendations."

To redo the test. The fictional documents used: case1_statement_2025-09-30.csv, case1_statement_2026-09-30.csv, case1_transactions.csv.

What to check. The return: if the client paid money in or took money out during the year, the simple return (final value divided by starting value) is wrong, and the model must say which method it used. Two weights redone by hand. The sum of the weights, which must be 100. The final page: no sentence like "it would be wise to reduce", which is a recommendation and is yours to make. In the test the model used the right method, Modified Dietz, with the formula (5.64%, where the naive calculation gives 21%), left the total row out of the holdings and reconciled the cash account on its own: £200 more than the transactions explain. That is where the coupons and distributions missing from the data show up. Always ask what is missing from the data you uploaded.

2. Explaining a fund from its product summary

Read first. Stages 3, 4 (the check).

The situation. The client asks what the fund you suggested actually is. The product summary is three pages long and they will not read it.

The request. Upload the product summary, then:

"Explain this product to a client who is not in the business, in half a page: what it buys, how much risk it carries (with the risk score in the document and what it means), how much it costs each year in pounds on £10,000 invested, how soon they can get out and on what terms. Use only what is in the document and give the section each figure comes from."

To redo the test. The fictional document used: case2_product_summary_alpha_balanced.txt.

What to check. Past performance: in the document it is history, and in the explanation it must not turn into "you could earn". The costs: the ongoing costs figure is the headline, but transaction costs and any performance fee are shown separately and add to it; the yearly total in pounds must come from the document's figures, and the model must say which totals it added up itself. The risk score, read in the document: product summaries use a scale from 1 to 10, while the KIDs and KIIDs that funds may still use until June 2027 use a scale from 1 to 7. Every figure with its section: if a figure has no section, find it yourself in the document. Two mistakes to look for: an average of past years presented as the return to expect, and the ongoing costs figure presented as the whole annual cost. In the test the model avoided both: £195 a year (ongoing costs, transactions and the performance fee average) and £395 in the first year with the entry charge, saying that it had added these up itself, and it read the score 4 on the right scale, from 1 to 10.

3. Comparing the costs of two funds over ten years

Read first. Stages 3, 7.

The situation. Two similar funds, different costs. The client wants to know how much the difference weighs over time.

The request. Upload the two product summaries, then:

"Compare the costs of these two funds: for each, entry and exit charges, the ongoing costs figure, transaction costs and performance fees, from the cost section. Then calculate with a program, showing the formulas, how much is left of £10,000 after 1, 5 and 10 years with the same gross return of 4% a year for both, and the difference in pounds. State at the top that the 4% is an assumption for comparing costs, not a forecast."

To redo the test. The fictional documents used: case3_product_summary_alpha_balanced.txt, case3_product_summary_beta_index.txt.

What to check. That the costs read are those of the document's cost section, in the same unit (yearly percentage or amount). The ten-year calculation: redo it for one fund with a formula in Excel. The sentence about the 4% assumption, which must stay at the top when the text goes to the client. Performance fees: if one fund charges them, a comparison at a fixed return can understate them, and the model must say so. In the test the figures added up, with the formula shown (a difference of about £2,450 after ten years, by a method slightly different from ours, which gives £2,523: what matters is that the method is stated), and the model noticed what the request did not say: at the same gross return the active fund does not beat its benchmark, so the performance fee would be zero; it gave both results, with and without.

4. The portfolio against the client's profile

Read first. Stages 3, 4, 7.

The situation. The client's profile says cautious, five-year horizon. After a year of markets, the portfolio may no longer be.

The request. Upload the portfolio and the profile summary (attitude to risk, horizon, objectives, agreed limits), without names, then:

"Compare the portfolio with the profile. Calculate with a program the weights by asset class and by single issuer, and flag every breach of the profile's limits, with the figure. Also list the holdings whose risk is not stated in the documents. Do not suggest any transactions: I only want the list of breaches and of missing information."

To redo the test. The fictional documents used: case4_client_profile_A.txt, case1_statement_2026-09-30.csv.

What to check. The limits: the model must use those written in the profile, not thresholds of its own. The holdings with no information on risk: that is the part usually missing, and it is where you will work. The final sentence: if it suggests transactions despite the request, remove them. The suitability assessment remains your act, with your own records. In the test the model found the certain breach (the structured product, a complex product the profile excludes), raised as a question whether to count the structured product among the equities (with it, equities go from 26.8% to 35.4%, above the 30% limit) and flagged that the profile was more than a year old. Read literally, it also said, the two funds would break the 10% single-issuer limit, before explaining that a fund manager is not the issuer of the securities the fund holds: how a limit is read is your decision, and it is worth writing it into the profile. For each holding it listed the risk information missing from the documents.

5. The note after the meeting

Read first. Stages 2 (the same request, done well), 5 (the finished work).

The situation. An hour with the client, notes taken in a hurry. The note has to be written today, while you remember.

The request. Paste the notes, then:

"From these notes write two texts. The first, for the client: what we discussed, what the client decided, what I will do and by when, in a friendly tone and without jargon. The second, for my file: the same, plus the new information about the client (family situation, objectives, horizon) and the questions still open. Do not add anything that is not in the notes; where a note is unclear, write [to be checked]."

To redo the test. The fictional document used: case5_meeting_notes.txt.

What to check. The decisions: the model can turn "I'll think about it" into "the client decided". Added sentences: in the tests of other fields the model made firmer a position that nobody had taken, and in a note to the client that becomes a commitment of yours. Dates and amounts, against the notes. The [to be checked]: if it never appears, read again more carefully. In the test the [to be checked] were there, also on "wants + bonds?", and "she'll think about it" stayed a non-decision. The text for the client, though, included a delicate sentence: that she does not remember choosing the structured plan. Said in a meeting, it is information; written by you in an email, it becomes a record. Sentences about disputed products or about the client's choices are yours to write.

6. The email to clients after a week of falling markets

Read first. Stages 2, 15 (when it searches instead of remembering).

The situation. Markets have lost a lot in a few days and clients are starting to call. You want to write to all of them before they do.

The request. With web search switched on:

"Find out what happened on the main equity and bond markets over the last five working days (the FTSE 100 and FTSE 250, the main US and European indices, ten-year gilts, US Treasuries and German Bunds): the change in the main indices, with source and date for every figure, and the reasons the sources give. Then write an email to my clients, one page at most: what happened, what it means for an investor with a long horizon, what I am doing (I will write this part: [...]), and that they can call me. No forecasts for the coming days and no promises."

What to check. The figures, each against its source and with its date: a fall "since the start of the week" and one "since the start of the month" are not the same thing. The reasons: the sources present them as hypotheses, and the email must not turn them into certainties. Generic reassuring sentences ("markets always recover"): remove them, they are forecasts. The part about what you are doing is yours to write. In the test, on a real week (2-6 November 2026), the model gave every figure with its source and date, flagged where the sources disagreed (the S&P 500 down 2.7% for one, 3.2% for another) and left the uncertain figures out of the email, saying which calculations were its own. Two things the request now fixes. Without a list of markets, from an Italian account it picked Wall Street, European indices and Italian and German bonds, with no FTSE 100 and no gilts: that is why the request names the markets. And with no list of sources it used blogs and websites that republish agency news, which it said itself: if you want only sources you know, write them in the request, as in case 7.

7. The monthly market commentary

Read first. Stages 15, 4.

The situation. Every month you send clients one page on the markets. Every month it takes half a day.

The request. With web search switched on:

"Write the market commentary for [month and year] for the clients of a financial advice firm. Fixed structure: equities, bonds and interest rates, currencies and commodities, the events of the month, what to watch next month. For every figure the source and the date, in a note. Use only these sources: [the Bank of England, the ONS, the index providers' websites, the news agencies you choose]. If you cannot find a figure, write 'not found'. Sober tone, no forecasts and no recommendations."

What to check. The dates of the data: last month's inflation figure presented as this month's is the most common mistake. Three figures checked against the source. The "not found": they are a good sign, and you fill them in. "What to watch next month" must list dated events (a Bank of England rate decision, a data release), not opinions on how they will go. In the test, on October 2026 and with the sources in the request, the model kept to the list and wrote "not found" for many figures: the ten-year gilt at the end of the month, sterling on 30 September (the Bank of England database refused it access), Brent, gold, the FTSE 250 and the US Federal Reserve decision, because the Fed was not on the list. It said that it had read Reuters and AP through websites that republish them, and which attribution it could not confirm. A narrow list of sources gives a commentary with gaps: better that way, because you see the gaps and fill them yourself. Add to the list the central banks you want covered.

8. A new rule: what changes for the firm

Read first. Stages 3, 4.

The situation. A rule or guidance from the regulator comes out that affects your work. The text is long and you want to know what you have to do.

The request. Upload the official text, downloaded from the regulator's website or from legislation.gov.uk, then:

"Read this text and tell me what changes for an independent financial adviser: new obligations, when they apply from, what I need to change in client documents and in procedures. For every point, the paragraph or section it comes from, quoted. Keep what the text says separate from what is your interpretation."

To redo the test. The fictional document used: case8_notice_fictional_body.txt.

What to check. Every obligation with its paragraph, opened and read. The application dates: there is often more than one, and a rule published today may apply in a year. The separation between text and interpretation: if it is missing, ask for it again. On doubtful points, the view of your trade body, your compliance support or your legal adviser counts for more than the model's. In the test, on a fictional text with five different dates, all the dates came out right, every obligation with its paragraph, and the interpretations kept separate from the text, including a list of what the text does not say (penalties, format of the register, how long records are kept).

9. The firm in a project

Read first. Stages 12 (the first project), 14 (a custom assistant).

The situation. Every note, every email, every commentary starts from scratch: tone, warnings and format have to be explained again each time.

The request. Create a project for the firm (stage 12), with no client data, with these instructions:

"You work for an independent financial advice firm. Write in UK English, in a friendly and sober tone, short sentences. You never give buy or sell recommendations and you never make forecasts about the markets. Every figure has a source and a date. You do calculations with a program and show the formulas. At the end of every text for clients, add this warning: [yours]."

In the project, upload the firm's templates: the meeting note format, the monthly commentary format, the warning. Then do cases 1, 5 and 7 inside it.

What to check. That the instructions are followed in the project's first chat: ask for a trial meeting note and look for the warning at the end. That no document with a real client's data ends up in the project for convenience: project documents stay there for every later chat.

10. The Monday briefing

Read first. Stages 15, 20 (the task that comes back on its own).

The situation. Every Monday morning you want to know what happened the week before on the markets and in the rules, without opening twenty websites.

The request. As a scheduled task (stage 20), every Monday at 7:

"Briefing on last week for a financial adviser: the five most relevant market events, with figures, source and date; news from the FCA, the Bank of England, the Financial Ombudsman Service and HMRC (ISAs and pensions), with the link to the document; the events of the week ahead (central banks, economic data). One page, no comments."

What to check. For the first three weeks, every link: it must open the document cited, not a generic page. News from the authorities: if there is none in a week, the briefing must say so, not fill the gap with old news. After a month, if a source never appears, ask yourself whether the model can read it. When the model cannot find the official page of a document, it should say so and give the link it used instead: those are the links to check first. In the test, on the week of 2-6 November 2026, the news from the FCA, the Bank of England, the Ombudsman and HMRC came with links to the official pages, and the model said that the Ombudsman had published nothing that week and that the HMRC newsletters were from 29 October, outside the week. The market figures and the calendar, though, came from minor financial websites, and the weekly close of the FTSE 100 was not found: for those, name the sources you trust, as in case 7.