Ever asked an AI assistant to summarise a bank statement or explain a payslip? It's a natural use, and it can be genuinely helpful. But you're handing over sensitive information, so do it deliberately. Here are simple principles that let you enjoy the benefits without opening a door you didn't mean to open.
What to share and what not to
Usually fine to share
- Amounts and categories, without identifying details.
- A statement or payslip after masking your name, ID number and account numbers.
- General questions: 'how do management fees work?'
- Data from an account you connected with defined, read-only permissions.
Never share
- Passwords for your bank, an app or your email.
- One-time codes (OTP) and authentication methods.
- A full credit card number, CVV and expiry date.
- Documents with your full ID number and account details, unmasked.
One rule that never changes
No assistant, app or representative needs your password or one-time code to help you understand your money. If someone asks for them, that is a red flag.
Permissions, scopes and read-only
Instead of pasting data by hand, you can connect an AI assistant to a data source through a defined connection. In Shavve this works through MCP, a protocol that lets an AI client such as Claude or ChatGPT read data from your account. The key principles:
- 1
A personal key
You create a personal key just for this connection, and never share a password.
- 2
Choose scopes
You define what the assistant may read: transactions, pensions, insurance, net worth or debts. You can grant only some.
- 3
Read-only
The assistant can read. It cannot change anything, and there is no write tool.
- 4
An audit log
Every call to every tool is recorded, so you can see what was done and when.
- 5
Revoke in one tap
You can revoke the key at any moment, and access closes.
Red flags and judgement
- They ask for a password or one-time code. A legitimate service won't ask for these in a chat.
- Broader permissions than needed. If one question gets a request for all your data, grant less.
- No clear way to revoke. Don't enable a connection when it isn't clear how to end it.
- The ability to act on your account. An assistant that can move money calls for a completely different level of care. Prefer read-only.
- Pressure to act fast. 'Now or never' is a signal to stop.
An AI assistant can be wrong
Language models phrase things confidently even when they are mistaken. Don't rely on a number you haven't seen at the source, and don't make a big financial decision on an assistant's answer alone. This is general information, not advice, and for a professional decision, consult a licensed professional.
Before connecting an AI assistant to your financial data
0 of 7 checked
Your ticks are saved on this device only.
Three ways to give an AI assistant information, and the risk of each
What to prepare before pasting data
Before pasting into a chat
0 of 5 checked
Your ticks are saved on this device only.
How to check an AI assistant's answer
An AI assistant can be wrong, state something untrue with confidence, and rely on old information. With money that matters especially, because a decision based on a wrong number can be costly. So check every important figure against its source: the fund's report, the authority's site or the official calculator. When the question concerns a law or a ceiling, ask the assistant for a source and check it yourself. And questions that need a personal decision, such as moving between funds, are best taken to a licensed adviser.
- Ask the assistant how it reached the result, and work out the simple steps yourself.
- Make sure the date and the law are current: ceilings and rules change.
- Do not act on a recommendation without checking with an official body.
- If an answer is surprising or sounds too good, that is a reason to check twice.
Good uses, and uses to avoid
Good uses
- Explaining a concept such as management fees or open banking in plain language.
- Drafting questions for the provider or the agent.
- Organising a list of expenses and suggesting categories.
- Summarising a long report into key points you then check.
Uses to avoid
- Making an investment decision or moving a fund on one answer.
- Handing over bank or app login details.
- Relying on a tax or rights calculation without checking with the authority.
- Connecting broad access and forgetting it.
If you connect an AI assistant to your data, choose a service that offers limited scopes, read-only access and an audit log, and that revokes in one tap. This is explained in the guide to connecting Claude or ChatGPT to your Shavve data. And for choosing the app itself, see Which apps are licensed and how to check.
If you shared more than you meant to
- 1
Delete the conversation
If the assistant allows it, delete the conversation and check the privacy settings.
- 2
Revoke connections
If you connected access, revoke it in the service's settings.
- 3
Change the password
If a password or code was exposed, change it at once and contact the bank or the relevant service.
- 4
Learn for next time
Set yourself a rule: what you never share.
Three questions and the right way to handle each
An example of checking an answer
Suppose an AI assistant writes that the management-fee ceiling in a pension fund is a particular number. The right step is not to accept the number as it is: ask it for the source, open the authority's page or Kol Zchut's page, and make sure the number is the same and current. If there is a difference, the official source is the truth. That way the assistant is a starting point, and the official source decides.
How to word a good question without identifying details
You can get a useful answer from an AI assistant without sharing who you are. The rule: give the structure of the situation, not the identity.
- Replace names, ID numbers, account numbers and your employer's name with labels such as “Bank A”, “Fund 1” and “Employer”.
- Round amounts when exact figures are not needed, and ask the assistant to explain how it calculates, not just the result.
- State the assumptions yourself (time horizon, track, fees) and say they are illustrative.
- Ask the assistant to list what information it is missing, rather than letting it assume.
An example: “I have two funds. Fund 1 charges a higher fee on the balance than Fund 2. What data is missing to compare them, and how do I calculate the difference?”. Then check the answer against the provider's own documents.
Frequently asked questions
Not with Shavve. The connection is read-only and there is no write tool, so a connected assistant cannot change data or move money.
Pasting hands over everything in the document, including details you don't need. A scoped connection limits what is read, and you can see in the log what was read and revoke at any time.
Revoke the personal key in settings. In Shavve you can do it in one tap, and the assistant loses access.
You can, if you follow the rules: share little, grant narrow permissions, read-only, and check against the source.
Revoke the key immediately, check the audit log to see what was read, and create a new key if you want to continue.
Passwords, verification codes, card numbers and ID numbers. Remove names and account numbers before pasting too, and copy only what the question needs.
Check every important figure against the official source: the fund's report, the authority's site or the official calculator. Questions that need a personal decision are best taken to a licensed adviser.
It is better to remove names, account numbers and identifiers before pasting, and to copy only what is needed. Also check the assistant's privacy policy.
Your AI, your data, read-only
With Shavve you can connect Claude, ChatGPT or another MCP client to your data with a personal key, scopes you choose, an audit log and one-tap revoke. Join the waiting list.
Join the waiting list