Skip to content
be-human.ai
← Back to the journal

Tools & integrations

Accounting Software Write Limits for AI Agents: Read, Draft, Approve

Which accounting records an AI agent should read, which it may draft, and which writes need a named human approver before anything posts to your books.

Be Human typographic cover with the title: Accounting Software Write Limits for AI Agents: Read, Draft, Approve
Conceptual illustration created with AI · Be Human

An AI agent should read most of your accounting records, draft a smaller set of entries for review, and never post, void, pay or change a ledger record without a named person approving that exact action. The safe boundary is not the software's permission screen alone. It is the list of record types, actions and amounts you decide on in advance, written down before the connection goes live. This guide shows how to draw that boundary for QuickBooks, Xero, Conta Azul or a similar system.

Key takeaways

  • Start with read access to reports, customer and vendor records, and open items, and add nothing new until that first reading job is reliable.
  • Let the agent draft bills, categorizations, reconciliation notes and journal proposals in a review queue, outside the ledger, until a person accepts them.
  • Require a named approver for every posting, voiding, payment, refund, tax setting change and chart of accounts change.
  • Give every write a stable key so a retry after a dropped connection cannot create a second bill or payment.
  • Log each record the agent read to support a decision and each approval it received, so an accountant can trace how a figure was produced.

Why accounting records need three permission levels

Accounting software is the system of record for money. A wrong bill, a duplicated payment or a reclassified expense does not stay inside one screen. It moves into the profit and loss report, the bank reconciliation, the tax summary and the month-end pack that someone else relies on. An agent that touches this data needs three distinct permission levels, and they should never be collapsed into one vague status such as connected.

The first level is read. The agent can see records and reports but changes nothing. The second is draft. The agent can prepare an entry, bill, note or proposal that sits in a queue, outside the ledger, where a person can inspect it. The third is write with approval. The agent may request a change to the ledger, but a named person must approve that exact change before the system accepts it.

The distinction matters because most accidents happen in the middle ground, where an agent is trusted to help and quietly ends up posting. A useful rule is that each permission must be granted separately and reviewed separately. For a broader view of this split, see read, draft and send: three different AI permissions.

Which records an agent should read

Reading is where an agent creates value early. It can answer questions that otherwise take an accountant an afternoon of exports: which invoices are past due, which bills arrived without a matching purchase order, which bank lines remain unmatched at the end of the week. Read access is also the cheapest permission to reverse, which makes it the right place to start.

A practical first read scope usually includes:

  • The chart of accounts and tax codes, read-only, so any draft uses account names that actually exist.
  • Customer and vendor records, limited to the fields the task needs, such as name, payment terms, currency and status.
  • Open invoices and open bills, with due dates and remaining balances.
  • Bank transactions and their reconciliation status, for the accounts the task covers.
  • Standard reports, such as profit and loss and balance sheet, for the periods the task requires.

Some records should stay out of the first scope. Payroll data and personal identifiers rarely add value to a bookkeeping task, and they increase what could leak if a prompt, a log or a summary is exposed. Bank credentials never belong in the agent's context at all. When a company runs several legal entities, keep read scope separate for each one, so a question about one entity cannot pull figures from another. Each read should also be recorded, because a later reviewer will want to know which report or ledger line the agent relied on.

Which records an agent may draft

Drafting is where the agent does the slow preparation work while a person keeps the decision. A draft is a proposed object with its reasoning attached. It lives in a review queue or a document, never in the ledger itself. Good candidates include:

  • Bill or expense proposals created from a received invoice, with the vendor match, proposed account, tax treatment and a reference to the source file.
  • Categorization suggestions for uncategorized bank lines, each with a short reason and a flag when the description is ambiguous.
  • Reconciliation exception lists that group unmatched items by likely cause, such as a timing difference or a missing document.
  • Journal entry proposals for accruals or reclassifications, marked unposted, with the calculation shown.
  • Reminder messages for overdue invoices, saved as drafts for a person to review and send.

Every draft should be checkable without reopening the whole source. It should state what it would change, which fields it fills, which it leaves empty and which assumption it depends on. A draft that says categorized as software tells the reviewer very little. A draft that says the vendor was matched to an existing supplier by tax number, the account was set to software subscriptions because the last three charges used it, and the payment term was left blank lets a reviewer decide in seconds.

Scanned invoices need extra care. Text extraction can misread a total or a date, so the draft should show the extracted values next to the source so the reviewer can compare them.

Illustrative scenario: a small services company asks an agent to prepare bill intake. In the first phase, the agent reads open bills and vendor records only. After a short comparison period, in which its proposals are checked against the accountant's decisions, it begins producing drafts for new bills, and the accountant approves each one. Write access comes later, and only for a short list of well-understood vendors. The sequence is a sketch of how the limits can be staged, not a measured result.

Which writes always need a named approver

Some actions should never run automatically, however well the agent performs in testing. Put these behind an explicit approval step:

  • Posting, editing or reversing journal entries and transactions.
  • Creating or sending invoices, and voiding or deleting either side of a document.
  • Creating payments, transfers or refunds of any amount.
  • Changing the chart of accounts, tax rates, tax codes or default accounts.
  • Changing bank connections, user access or integration settings.
  • Recategorizing transactions in bulk, or closing and reopening an accounting period.

An approval is only useful if it is specific. The approver should see the exact payload: the vendor, the account, the amount, the date, the tax treatment and the source document. A general message such as yes, go ahead in a chat thread is not an approval record. A sound approval names the approver, expires if it is not used within a short window, and binds to the exact request, so that a changed amount forces a new review. The check itself belongs in the integration layer, not only in the agent's instructions. The OWASP authorization guidance makes the same point in general terms: access decisions should be enforced server-side, where a clever prompt cannot override them.

Retries need the same discipline. If a connection drops after the accounting system accepted a bill but before the agent received confirmation, a naive retry creates a second bill. Each write should carry a stable key derived from the source document and the action, so the system recognizes a repeat instead of creating a new record. Stripe documents this pattern for its API as idempotency keys, and the principle carries over to any accounting integration that supports it. The practical detail is covered in idempotent tool writes that stop duplicate records.

A step-by-step plan for setting the limits

Work through the sequence in order, and do not advance a stage until the previous one has run without surprises:

  • Step one: name the single accounting task the agent will support, such as past-due follow-up or bill intake, and write its success criterion in one sentence.
  • Step two: list every record type that task touches, and mark each one as read, draft or write with approval.
  • Step three: create the connection with read-only scope for the entity and period involved. For connection details, see QuickBooks AI agent integration: read and draft, never post alone.
  • Step four: run the agent on past data and compare its output with the decisions the accountant actually made, noting every disagreement and its cause.
  • Step five: enable drafts into a review queue, with a named reviewer and a fixed daily review time.
  • Step six: add approval gates for each write type, recording who approves and how long an approval stays valid.
  • Step seven: test failure cases before any write is enabled, including a dropped connection, a duplicate upload and a changed vendor tax number.
  • Step eight: review permissions and logs every month, and remove any scope the task no longer uses.

The order matters because each stage produces evidence for the next. A draft queue that reveals repeated categorization errors is a reason to fix rules before enabling writes, not a reason to enable writes faster.

How Be Human handles write limits in managed operations

In a first engagement, Be Human defines the single role the agent will fill, the records and tools it needs, the read, draft and approval boundary for each one, and an observable success criterion that the business owner can check without reading code. Managed operations can then cover monitoring of the agent's activity, recurring evaluations against the decisions the accountant makes, permission review as staff and entities change, integration maintenance when an accounting vendor changes its API, and incident response when a connection expires or a write behaves unexpectedly. If you are not sure which accounting task should reach an agent first, the free Be Human AI diagnostic can help you find the first task worth handing over.

Frequently asked questions

Can an AI agent post journal entries on its own? Not in a sound setup. Posting is a write, and writes need a named approver who sees the exact entry. The agent can prepare the entry as an unposted draft with its calculation and source attached, and the accountant decides whether it posts.

Should the agent use the accountant's login or its own account? Use a separate integration identity with the smallest role the task needs. That keeps actions attributable and lets you revoke access without changing anyone's personal login. Confirm that the audit log shows the agent as the actor, so reviewers can separate its actions from human ones.

What happens if a draft is wrong? Nothing is posted, because drafts never enter the ledger. The reviewer rejects or corrects the draft, and that correction becomes evaluation data. If the same kind of error keeps recurring, adjust the agent's rules or its read scope before considering any write permission for that task.

Does the same approach work for Xero or Conta Azul? The three levels apply to any accounting system, although the available scopes and approval features differ by product. Check what each vendor's API actually permits, how it records the actor on each change, and whether it supports idempotency keys. Where the vendor offers no scope separation, enforce the limit in your integration layer instead.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →