AI Guides › Playbooks
By Nigel Guy · 8 min read
The usual way to learn "how the rich use AI" is a post about a named billionaire's secret prompt, or their chief of staff's clever bot. Those stories are close to impossible to check, and when you go looking, you find they rarely trace back to anything the person said on the record. What you can check is less glamorous: dated surveys of family offices and wealthy investors, published by the banks and consultancies that serve them. Read side by side, they point somewhere quieter than the anecdotes, and the useful part is something you can copy this weekend.
The rule: let AI do the reading and the reporting, keep the deciding for yourself, and decide what it is allowed to see before you give it anything.
I couldn't verify a single named example of a wealthy individual's private AI habits, so this guide names none. These are the three sources it rests on, with dates and sample sizes, so you can weigh them yourself.
| Report | Published | Who was asked | The AI-relevant finding |
|---|---|---|---|
| UBS Global Family Office Report 2025 | 2025 (survey 22 January to 4 April 2025) | 317 family offices, average USD 1.1 billion managed | Asked how they were likely to use AI in their own operations over the next five years: financial reporting and data visualisation 69%, text analysis (such as summarising legal documents and financial statements) 64%, portfolio analysis 62%, content creation 54%, customer experience 13%. Separately, offices that keep tasks in-house cited privacy (63%) and operational control (63%) as reasons. |
| J.P. Morgan Private Bank 2026 Global Family Office Report | 2 February 2026 | 333 family offices in 30 countries, average net worth USD 1.6 billion | 65% intend to prioritise AI as an investment theme. 32% name cybersecurity as their greatest service need. 80% outsource some part of portfolio management. |
| Capgemini World Wealth Report 2026 | 4 June 2026 (HNWI survey fielded January 2026) | 6,510 high-net-worth investors, plus 1,317 relationship managers | Only 17% of wealthy clients say their advice experience feels seamless and personalised. 76% of relationship managers want AI systems that automate routine work and surface client insights. |
Alongside the J.P. Morgan report, the bank's Head of Cyber Advisory published a piece on 14 April 2026 about family offices adopting AI. Its advice: govern the data before adding tools, start on a ring-fenced set in read-only mode, restrict connectors, switch off model training where possible, and keep account numbers, tax records, medical data, personal identifiers and deal terms out of prompts entirely.
Two caveats. The UBS figures are stated intentions over five years, not observed behaviour. UBS has since published a 2026 edition; I could only check the AI-use table in the 2025 edition, so those are the figures used here.
None of these reports set out to describe how the wealthy use AI, but together they line up:
So the quiet signal isn't a secret tool. It's an order of operations: decide what data AI may touch, then use it to read and summarise, then make the decision yourself.
One table for your own finances or small business. Each row is one recurring piece of paperwork; fill it in once, use it monthly.
| Column | What goes in it |
|---|---|
| Task | The recurring reading or reporting job, such as "pension statement" or "monthly spending summary" |
| Signal | Which pattern above it matches: reporting, text analysis or portfolio overview |
| Data tier | Red: never goes into a chat (account numbers, National Insurance number, passwords, medical details). Amber: goes in only after redaction. Green: safe as it is |
| Prep | What you strip out or replace before upload, such as "replace names with Person A, cut account numbers" |
| AI job | Summarise, tabulate or compare. Never "decide" |
| Human check | The one thing you verify against the original before acting |
Build it in five steps:
Fill in the document type and your question, then paste the redacted text where shown.
You are a careful document analyst helping a private individual in the UK understand their own paperwork. You do not give regulated financial, tax or legal advice.
Document type: [DOCUMENT_TYPE, e.g. pension annual statement]
What I want to know: [MY_QUESTION, e.g. what changed since last year and what fees I paid]
Steps:
1. Read the document below in full before answering.
2. List the key figures (amounts, dates, rates, fees) in a two-column table: item | value as written in the document.
3. Summarise in no more than five plain-English bullet points what the document says about my question.
4. List anything unclear, missing, or that I should confirm with the provider, as questions I could send them.
Rules:
- Quote figures exactly as they appear. Do not estimate, round or fill gaps.
- If my question can't be answered from this document, say so and stop. Do not guess.
- Do not recommend what I should buy, sell or move.
- If the document type or question is missing above, ask me for it before starting.
Before you reply, check that every number in your table appears word for word in the document.
Document:
[PASTE_REDACTED_TEXT]
Fill in the month and paste a redacted export of your transactions or account balances.
You are a bookkeeping assistant producing a one-page monthly money report for a household or sole trader in the UK.
Month: [MONTH_AND_YEAR]
Categories I use: [MY_CATEGORIES, e.g. housing, food, transport, subscriptions, business costs]
Data: [PASTE_REDACTED_EXPORT]
Steps:
1. Assign each line to one of my categories. Put anything you can't place in "Unsorted". Do not invent a new category.
2. Produce a table of category | total in £ | number of transactions.
3. List any subscriptions or repeat payments you can see, with amount and date.
4. Note up to three things that look different from a normal month, phrased as questions for me to check, not as conclusions.
Rules:
- Use only the data provided. If the data is incomplete or the columns are unclear, ask me before producing totals.
- Do not give investment, tax or debt advice.
Before you answer, check that your category totals add up to the total of the data provided, and say so.
Priya is invented for this example: she's self-employed in Leeds, with a workplace pension from an old job, a Stocks and Shares ISA and one rental flat. Her ledger has eight rows. The pension statement is amber: she removes her name, address and plan number, then runs Prompt 1 asking "what fees did I pay and how has the value changed". Her bank export is amber: she deletes the account number column and the payee names of individual people before running Prompt 2. Her tenancy agreement is red because it contains the tenant's personal details, so it stays out entirely. The model's output tells her what to ask the pension provider. It doesn't tell her whether to transfer the pension; that stays with her, and with a regulated adviser if she wants one.