All guides/QuickBooks and MCP

Is it safe to connect your accounting to an AI assistant?

Three checkable questions decide it: what gets stored, what the connection can do to your books, and who trains on the data. Here are the answers, including when not to connect.

Updated

It depends on three things you can actually check: what the connector stores, what it is allowed to do to your books, and whether the assistant's operator trains on what it sees. Anyone answering "yes, totally safe" without addressing those three is selling. Here are ours, in order.

What is stored, and what is not

The connector is a relay. A question goes to QuickBooks, the answer comes back to the assistant, and no synchronized copy of your ledger is kept in between. No cached reports, no customer or invoice list on our side.

What the connector stores is what the service cannot function without:

  • Your account email address, billing details and usage records.
  • The details of the companies you have connected.
  • Any bookkeeping rules or task instructions you save.
  • Your QuickBooks access tokens, AES-256 encrypted, on infrastructure hosted in Canada (AWS Montréal region).

Two optional features keep more. The document box stores the files you send it and the text read from them. Autopilot stores its instructions, conversations, run history and proposals, and that history can include QuickBooks data. Do not read any of this as "nothing worth stealing", though. The tokens are the sensitive part: whoever holds one can read the books it is bound to until it is revoked. That is why they are encrypted, and why revocation matters more than any other control on this page.

A token, not a password

We never receive your QuickBooks password. Authorization happens on Intuit's own screen, the same one every QuickBooks app uses. What comes back is an access key Intuit issues to the application, and you can revoke it at any time from QuickBooks. The connection dies on the spot. Nothing to reset, nothing to change on your side, no email to us.

If you take one operational habit from this page, take that one: you already hold the off switch.

You approve every change

Every connection starts with full access, so the assistant can read your books and create or change records. In Claude and ChatGPT, it asks you before each change: every tool that writes waits for your approval in the chat, unless you change that in the assistant's own settings.

Read-only is a switch per company, and the account owner flips it in the portal, at any time. On a read-only company the assistant can read anything and change nothing. A bookkeeper can keep nine client files read-only while working in the tenth. Nothing propagates. On a team, each person also has their own level for each company: Read or Write.

Even with full access, deleting, voiding and sending a document by email need an explicit confirmation of that specific record in the conversation, every single time. Before an email goes out, the assistant shows you the document and the address as QuickBooks holds them. An assistant cannot email your customer or void an invoice on its own momentum.

Model training

On Claude's Free, Pro and Max plans, Anthropic's article on model training says the raw content returned by connectors, remote MCP servers included, is not used to train its models. The conversation itself can be, if you allow it: your questions and Claude's answers, which quote your figures. Turn off model improvement in your Privacy settings before you connect a real company file. Team and Enterprise plans fall under Anthropic's commercial terms instead.

OpenAI differs by account type. Business, Enterprise and Edu are excluded. On a personal ChatGPT account, you turn it off yourself: Settings, then Data Controls, then switch off "Improve the model for everyone". Do that before connecting a real company file.

What we are not claiming

We are not claiming your numbers become untouchable. They travel to a model run by Anthropic, OpenAI or Microsoft, and they land in your conversation history, which lives in your account with that company under their retention rules, not ours. If a client contract or an engagement letter forbids that, no connector setting fixes it. That is a contract question, and it belongs to you.

We are also not claiming the assistant is always right. It can misread a report the same way a person skimming can. Anything headed for a filing, a lender, or a decision that matters gets checked against the report in QuickBooks. Accurate and safe are different properties and people conflate them.

Full detail on the architecture is on the security page.

If you are not comfortable with any of this

Two respectable answers, and neither of them is "get over it".

Switch the company to read-only and leave it there. You get all 35 reports, the whole general ledger, aged receivables, budget versus actuals, everything, and the connection keeps zero ability to change your books. Plenty of connections should stay exactly like this forever, and the pricing is the same either way. There is no upsell to full access, see the pricing page.

Do not connect at all. If you hold client files under terms that do not allow a third-party processor, or you would not want to explain the arrangement to a client, then the right move is to skip the connector and export what you need. Run the report in QuickBooks, hand the export to the assistant, get the analysis, keep the connection out of it. You lose live data and you lose write access. For a lot of people that trade is correct, and we would rather say so than sell a subscription that ends in regret.

Once you have decided, the setup guide walks through connecting a first company. If you want it read-only, switch it in the portal right after.

Try Caribooks for 14 days

→ Connector from $39/month. Full 14-day trial, no card required.

Read next