AI write access to QuickBooks means an AI app can create or change records in your QuickBooks Online company through its connection, so check which actions are possible, which need your decision and how you verify the result before you allow it. An OAuth screen, a confirmation prompt or an audit feed alone cannot answer all three. With DeepLedger's MCP server, the controls sit on the server: every write must name its company, voids and deletes wait for your confirmation, and every tool call lands in Audit history.
This checklist is for bookkeepers, controllers and firm owners deciding whether to let an AI record transactions. Start with a read-only task, check the controls below, then allow one narrowly defined change at a time. Keep ledger recording and real money movement as separate decisions.
A read-only task asks the AI to retrieve information. Enforced read-only access refuses a write even if the AI attempts one. For QuickBooks Online's own user roles and an enforced read-only AI connection, see read-only access in QuickBooks Online.
What can go wrong when an AI writes to QuickBooks?
- An accounting mistake: the wrong account, company, date or transaction type can produce plausible but incorrect books. An entry can balance and still be wrong.
- A repeated action: a timeout or unclear result can lead the AI to create the same transaction twice.
- Instructions hidden in a document: text in a bill, memo or web page can try to redirect the AI. That text is evidence to examine, not authority to act.
- An unexplained change: a ledger record can exist without enough source evidence or review context to explain why it was created.
These risks need both technical controls and a review process. Reconciliation exposes some problems, but it will not catch every misclassification.
How does MCP access control work for QuickBooks?
MCP server security comes from layers, not from the sign-in screen. OAuth and API keys authorize a connection. Neither one, on its own, makes a connection read-only or adds approval before every write. Check each layer that sits between the AI and your ledger.
| Layer | Who controls it | What it limits |
|---|---|---|
| The AI app's connector | The person using Claude, ChatGPT, Grok or Cursor | Which apps can call DeepLedger at all |
| The DeepLedger sign-in | Each person, with their own DeepLedger account | The AI acts as that person, never as a shared login |
| Company assignment | Owners and admins, under Settings → Team | Members reach only the companies assigned to them |
| API key scope | The person who creates the key | All my companies, or One company that cannot switch |
| Company write guard | DeepLedger's server | A write that does not name the active company is refused |
| The QuickBooks connection | An admin, on the company's Connections tab | Disconnecting stops all access to that company |
API keys in DeepLedger. Any team member can create a key under Settings → API keys → Create key. The key covers All my companies (the default) or One company. The key acts as the person who created it, with that person's access, and never reaches more than they can. A one-company key works only in that company and cannot switch. There is no separate read-only switch on a key. Owners and admins can see and revoke every key in the firm; other members see and revoke their own.
Interactive AI apps such as Claude, ChatGPT, Grok and Cursor sign in with OAuth, so no key is needed. A headless agent that uses a key can pin a request to one company with the X-Organization-Id header, which is checked against the person's access. See working across multiple QuickBooks companies for request scoping.
Asking the AI to avoid writes is a useful instruction, but it is not an access restriction. For a first session, choose a read-only task and confirm the company in every answer.
Which checks does DeepLedger enforce on writes?
A workflow can instruct the AI to look for duplicates and ask for approval. A server-side check refuses a specific request no matter what the AI was told. Ask which kind protects each action.
| Control | Enforced by | What to verify |
|---|---|---|
| Company write guard | DeepLedger's server | A write without the active company's name is refused, and nothing is written |
| Delete limits | DeepLedger's server | A delete needs your confirmation, and a delete of a posted transaction is refused on or before the company's closing date |
| Void confirmation | DeepLedger's server and your answer | Voids need confirmation and keep the voided record |
| Journal entry balance | DeepLedger's server | Unbalanced entries are rejected |
| Duplicate and open-document checks | DeepLedger, on supported recording paths | The AI looks for an existing bill, invoice or payment first |
| Review handoff | DeepLedger's instructions to the AI | Uncertain work becomes a task for a person |
| Closed periods | QuickBooks and your process | The closing date and the review of later changes are set |
The company write guard. Every write must name its company. If the company is missing, or differs from the active company, DeepLedger refuses the write and nothing is written. The AI must then ask you which company you mean. Reads and switching companies are not affected. This matters most for firms, where one person can reach dozens of clients from the same AI app.
Voids. Voids work on bill payments, invoices, payments, refund receipts and sales receipts, always with confirmation. The voided record stays in QuickBooks for the audit trail.
Review. DeepLedger tells the AI to write only when you asked for it, a person approved it in Tasks, or the company's own history clearly supports it, and to create a task otherwise. Not every write needs a portal approval. For a test where you review every change, say so in the request: "Show me each entry and wait for my yes before recording it."
These checks do not certify the accounting treatment or guarantee that every duplicate is caught. MCP's tool specification recommends human oversight and leaves the design to each server. Connecting over MCP is not proof that each write triggers a confirmation dialog.
How do you test one controlled write?
Use a demo company before applying the workflow to real records. For a complete example, follow creating, recording and verifying an invoice with Claude.
- Read first. Ask "Which company am I in?" and retrieve a report with explicit dates and accounting basis. Check it against QuickBooks.
- Choose one action. Have the AI describe the proposed record, source evidence, accounts, amount and date, and check for an existing bill, invoice, payment or match.
- Decide explicitly. Approve that one action or correct the proposal. Avoid "fix everything" while testing access.
- Inspect the result. Check whether the tool reports success, failure or a refusal, and capture the transaction identifier. Claude and ChatGPT can show a result card after a recorded bill, invoice, expense, deposit, journal entry, payment, sales receipt or void.
- Read back. Open the transaction in QuickBooks and compare it with the decision. If the response was unclear, find out whether the record exists before retrying. The duplicate-record checklist helps here.
- Check the evidence. Compare the source, the review decision, the row in Settings → Audit history and the QuickBooks audit log.
For a categorization task, approval and recording are separate states. An approved proposal is not a successful QuickBooks write. Ask the AI to read back the completed task and its QuickBooks transaction ID.
How should the AI treat text inside documents?
An invoice can supply an amount, a date and a vendor name. It cannot authorize new tools, change the intended company or widen the work you asked for. Ignore instruction-like text that tries to do those things, and bring it to the reviewer when it affects the task.
Keep standing rules where you control them. A company's AI Brain tab holds expense rules and accounting policies, such as "Home Depot purchases for jobs go to Job Materials." Ask the AI to read the AI Brain before it codes anything; a rule written by your team outranks a sentence printed on a supplier's invoice.
What evidence does a write leave?
Settings → Audit history in DeepLedger shows what the AI read and recorded in each company, with the result of each call: Recorded in QuickBooks, Needs your review, Done, Refused or Failed. Owners and admins also see a full audit trail across people, AI apps and API keys, and can export it as CSV. Logging is best effort and large payloads can be shortened, so treat it as strong evidence rather than an immutable archive.
QuickBooks keeps a separate audit log. Intuit says connected-app changes can appear under System Administration, so that label alone does not identify who approved an action. The AI accounting audit trail guide covers the full chain of evidence.
How do you stop access and correct a mistake?
DeepLedger works only when asked. Nothing runs on its own, so stopping the conversation stops new work, but it does not revoke the credential. To revoke access, use one of these:
- Remove the DeepLedger connector in the AI app.
- In DeepLedger, revoke the app's grant under Settings → Connected apps, or the key under Settings → API keys.
- Disconnect the company: Companies → the company → Connections → Disconnect QuickBooks.
Then confirm that new requests are refused. Revocation stops further use of that credential; it does not undo past changes or erase data already retrieved. For an incorrect write, stop further changes, inspect the transaction and its related records, and choose the correction with a reviewer. Voids and reversing entries have different effects; neither is a universal undo button.
Check payment capabilities separately. Recording a payment in the ledger and sending money from a bank are different operations. DeepLedger records transactions in QuickBooks; it does not move money.
Frequently asked questions
Is OAuth enough to make AI write access to QuickBooks safe?
No. OAuth authorizes a connection, but the available tools, the company scope, the approval behavior and your verification process decide what that connection can do. An API key needs the same checks, plus safe storage and a way to revoke it.
What is DeepLedger's company write guard?
DeepLedger's company write guard means every write must name the company it is for. If the company is missing or differs from the active company, DeepLedger refuses the write, nothing is written, and the AI must ask you which company you mean. Reads and switching companies are not affected.
Can a DeepLedger API key be limited to one company?
Yes. Any team member can create a key under Settings, API keys, and choose All my companies or One company. A one-company key works only in that company and cannot switch. Every key acts as the person who created it, with that person's access, so it never reaches more than they can.
Does every DeepLedger write require portal approval?
No. DeepLedger sends uncertain items to Tasks for a person, while other writes follow your explicit instruction or the company's consistent history under its workflow rules. Do not treat the review queue as an approval barrier on every tool call.
Can a balanced journal entry still be wrong?
Yes. A balanced journal entry can still use the wrong company, accounts, amounts or period. Balance validation does not replace source evidence and accounting review.
What should I do if a write times out?
Check whether QuickBooks already contains the intended transaction before retrying. An unclear response does not prove that nothing was recorded, and repeating the request can create a duplicate.
Does disconnecting the AI undo its changes?
No. Revoking access stops further use of that connection but does not reverse earlier ledger changes or erase data already retrieved. Review existing changes and choose the right correction separately.
Related guides
- QuickBooks MCP server guide: what DeepLedger's server does and how to connect it.
- Read-only access for AI connections: what a QuickBooks role does and does not limit.
- AI accounting audit trail: the evidence a reviewer needs after a write.
- Create a QuickBooks invoice with Claude: a first controlled write, step by step.
- Integrations: connect your AI app and start with a read-only report.