Client confidentiality, more than technology, is the safety question for an accountant: do you have the client's agreement, are you giving the assistant only what the task needs, and can you show later what it saw. If any of those is no, do not connect.
The tool matters less than the discipline around it. Your duty to the client does not change because an assistant is fast.
What should be settled before connecting anything?
- Written consent from the client for AI-assisted work on their records, ideally in the engagement letter.
- Which data is in scope: transactions only, or documents too.
- Where the records live: the client's own workspace, not a mixed pile.
- Who else at your firm can connect, and who reviews the output.
Note the product boundary honestly: SpendixAI Pro covers one personal workspace and is not a team, approvals or client-management tool. If a client keeps their own records there and chooses to connect an assistant, that is their workspace and their decision; do not build a multi-client practice around it.
Which scope suits an accountant's work?
Claude or ChatGPT can connect to SpendixAI through MCP with a scope you choose, and you can start read-only and disconnect any time. For review work such as spotting odd entries or summarising a quarter, read-only is enough. Reserve any write access for a narrow clean-up and check every change.
| Task | Suggested approach |
|---|---|
| Summarise a quarter's spending by category | Read-only |
| Flag unusual or duplicate entries | Read-only, then fix by hand |
| Bulk recategorise after a chart-of-accounts change | Write scope for that session only, then disconnect |
| Anything touching GST or tax treatment | Do it yourself; use the tool to list, not decide |
What about one client's data leaking into another's chat?
Use a separate chat and a separate connection per client, and never ask a question that spans two. Suppose you handle twelve small clients: twelve connections is more bookkeeping, but it is the only arrangement where a mistake cannot cross over. Disconnect each when the engagement task ends.
How do I treat the output?
Treat it as a junior's first draft. GST figures in the app reports are estimates to be checked against invoices and with a CA, which in your case is you. Reconcile any total the assistant gives you to the ledger before it goes near a return or a certificate. Rules and limits change, so confirm the current position from the official source.
What do I tell the client?
Say it plainly and in advance. Tell them which records the assistant will read, that you will begin read-only, that you will reconcile its figures to their ledger, and that you will disconnect when the task is done. Clients are usually comfortable when the answer is specific. A vague line such as we use AI tools is what makes people nervous.
- Name the tool and the purpose, for example reviewing a quarter's categorised spending.
- State what is out of scope: bank logins, other clients, anything unrelated.
- Offer to show them the reconciled figures.
- Confirm how long you will keep any working notes.
If a client says no, that settles it. You can do the same review from an export and your usual checks.
A short policy you can adopt
- No connection without client consent on file.
- Read-only by default.
- One connection per client, closed at the end of the task.
- No bank logins, OTPs or card numbers ever pasted into a chat.
- Every figure checked against the ledger before it is used.
Useful ideas start with a clearer picture of your own spending.
Explore your reports

