On November 18, OpenAI launched ChatGPT for Accounting. The launch page listed eight partner connectors: Intuit QuickBooks, Xero, Oracle NetSuite, NetSuite QuickStart, BILL, FreshBooks, Rillet, Campfire. For accounting software without a native plugin, it directly operates its interface using computer use.
The list does not include Odoo, SAP, Kingdee, or Yonyou.
This list is worth a separate look. The feature list explains what it can do; the list shows where it intends to enter. For people working on ERP, the latter often carries more information.
1. The four things it does all fall within a finance manager's daily routine
First, make clear exactly what it does. The four workflows basically cover the most time-consuming parts of month-end.
| Workflow | What does it do | Where is it stuck? |
|---|---|---|
| Draft journal entries and invoices | Extract information from emails and receipts, generate journal entries and draft customer invoices, suggest accounting accounts, and suspend uncertain items. | Account recommendations rely on historical data; for new business without history, recommendations are basically guesses. |
| Bank reconciliation | Collect statements and payment records from emails, files, and connected systems, and cross-check them | The nature of the discrepancy still requires human judgment: whether it is a fee, the exchange rate, or the other party underpaid. |
| Working paper verification | Compare working papers with supporting documents, and flag duplicates, missing fields, and unexplained discrepancies | Can only point out the discrepancies, but cannot decide on behalf of accounting how this transaction should be handled. |
| Month-end closing | Save the closing instructions, templates, and review checklists as reusable skills, then keep using them next month | Templates become outdated. Policies change, subjects change, and no one updates the template, so it keeps using the old rules. |
These four things have one thing in common: they are all "preparation" work, not "confirmation" work. Preparation can be outsourced to machines; confirmation cannot.
There is another supplementary capability that is easy to overlook: for software without native plugins, ChatGPT can use computer use to operate its interface. For small vendors this is a good thing—it means they can be integrated without developing an API; for large enterprises it is instead a risk—interface operations bypass the permission checks and logs at the API layer, and the actor becomes a 'person simulating clicks' rather than an authenticated caller.
II. Three design choices that are more worth reading than the features
There are only three governance settings on the publishing page, but each one directly determines whether this thing can get into the finance department.
First, who can enable it. Workspace administrators decide which accounting apps can be connected. Business edition employees can install them themselves, while Enterprise edition requires administrator approval. This means the switch is in the enterprise's hands, not the employees'.
Second point: whose permissions are used. Users connect their own accounts, and ChatGPT uses the permissions that account already has in the other party's system. OpenAI's example is very straightforward: a bookkeeper who does not have posting rights in the general ledger will not gain posting rights just because they use ChatGPT.
Article 3: Who approves. Every draft journal entry must be manually approved before it is posted. Regarding data, Business and Enterprise information is by default not used to train models.
Between “Draft” and “Posted” stands a person. This is not a product design preference; it is a requirement of accounting standards. Once an entry is posted, it enters the voucher sequence; changing it requires leaving an audit trail, and reversing it requires a reversing entry. Handing this step over to a probabilistic model is equivalent to handing the audit trail over to uncertainty.
There is one more tip worth highlighting separately. In its operating guidelines, OpenAI writes: if the status of an entry is unclear, first verify it in the accounting software, then decide whether to retry, to avoid the same invoice or entry being recorded twice.
This statement shows that the product team itself anticipated that the connection would fail and the status would be unclear. It listed "duplicate posting" as a real risk, not a theoretical one.
III. The list is a channel, and also a boundary
Back to those eight connectors. Why these companies? Looking at their customer base makes it clear: QuickBooks, Xero, FreshBooks, and Rillet mainly serve small and medium-sized businesses and accounting firms; NetSuite serves mid-sized and larger companies; BILL handles accounts payable; Campfire leans toward startups. This is a list arranged according to 'cloud accounting + US market.' Intuit is only open to QuickBooks Online in the US region, and Xero's review workflow is also only available in the US.
Xero's approach is worth a closer look. On November 19, it announced expanded integrations: accountants can perform month-end close reviews, client portfolio reviews, and quarterly communication preparation in ChatGPT; but the actions to address issues once found must be done back in Xero. The books and change records remain in Xero; ChatGPT is just the front-end workbench layer.
The division of labor is clear: AI reads and drafts in the front end, while the ledger is stored and modified in the back end.
What about systems that aren’t on the list? Three paths.
- Build your own channel. Use MCP or your own API to expose the system as tools for enterprise agents to invoke. Odoo 20 comes with an MCP server; tools are exposed as server actions, read-only by default, and write operations must be individually enabled by administrators. This means "not on the list" does not equal "cannot connect," but channels, permissions, and logs all have to be built yourself.
- Bridged by the integrator.Add a layer in between to convert documents from System A into a format recognized by System B. Costs are controllable; the risk lies in the mapping rules—when fields don't align, errors occur silently.
- Bypassed. Export to Excel, upload to AI, and import it back after processing. This is the least troublesome and also the most dangerous: the data leaves the system, and approvals, audit trails, and permissions all disappear together; only the results come back.
A third path is actually happening in many companies. Employees use AI to process reports themselves, and their bosses may not know. It won’t throw an error or leave a trace, until one day two reports don’t match, and only then is a manual step discovered in between. Whether to choose the list is a strategic issue; whether to block this path is a management issue.
So when enterprises evaluate ERP, they can add one more question: Is this system on the list of mainstream AI entry points? If not, who will maintain the self-built channel, and against what standards will it be accepted? There is no standard answer to this question today, but there will be within two years.
IV. The hand that posts the entries—why can't it be handed over to the model
In an accounting system, "Draft" and "Posted" are not an interface state; they are a boundary of responsibility.
Take Odoo as an example. The journal entry object account.move has several clear states: Draft, Posted, and Cancelled. A posted entry cannot have its amount and accounts changed directly; to correct it, you can only create a reversing entry (reverse entry), and both the original entry and the new entry remain in the books. The system also supports a lock date: once a lock date is set, no one can post any voucher before that date, including administrators.
Together, these three things constitute an unavoidable path: to change an entry that has already been posted, you can only add another entry, and you cannot make the original entry disappear.
| segment | How does the system block it? | How might AI get around it? |
|---|---|---|
| Modify posted journal entry | Direct editing is not allowed; only reversal is permitted. | Cancel it and recreate one; no trace will be visible in the accounts. |
| Record to locked period | Reject writes after the lock date | Change the date to next month, keep the amount the same |
| Posting by a person without posting rights | User Group Permission Verification | Borrow a service account with permissions |
| The same invoice is recorded twice | Duplicate supplier bill number alert | Change the number, change the supplier's wording, and get around it |
The column on the right isn't alarmist fearmongering—it's the branch you have to think through in advance for every integration. When an agent hits a "system rejection," it won't stop to reflect; it will try the next path. If you don't clearly spell out which paths are off-limits, it will choose for you.
The most common mistake AI makes in the books is not choosing the wrong account, but recording the same entry twice. A wrong account makes the reports look bad, and someone checks at month-end; a duplicate posting makes the reports look good, and then no one checks. The former is an error; the latter is an incident.
5. Permissions are not for AI; they are for people first
"Reusing existing permissions" sounds convenient, but in reality it just carries the old problems over as they are: if people's permissions were already a mess, AI just amplifies that mess all over again.
Start with accounts. There are two reasonable approaches: a read-only account, given permissions to query data and run reports; and a write account, allowed only to produce drafts, not to post transactions. Both accounts must be included in the permission matrix, both must have an owner, and both must be able to be disabled within ten minutes.
Next, look at five of the most error-prone places. These are not issues of model capability, but of business structure.
| Error-prone areas | Why is it wrong | How to intercept in advance |
|---|---|---|
| Multi-company | The default company was selected incorrectly, so the document was assigned to another entity. | Use record rules to restrict visible companies; a service account is bound to only one entity |
| Multi-currency | The conversion exchange rate used the payment date rather than the document date | Hardcode the exchange rate date source in the rules, and uniformly revalue at month-end |
| Tax | The tax position (fiscal position) did not match when crossing borders. | Build mappings by customer, supplier, and country; validate tax rates order by order after generation |
| Analyze Account | The journal entry has been generated, but the cost was not allocated to a project or department | Make the analysis dimension required; without it, you cannot save a draft. |
| Attachment | The journal entry exists, but the source documents are not attached. | Require the number of attachments to match the number of entries even at the draft stage |
Also, distinguish two things: record rules determine “which rows can be seen,” and access permissions determine “which operations can be performed.” These two things should be tested separately, because AI makes mistakes in different ways—most of the time it will not perform unauthorized operations, but it will see rows it should not see, and then write information that should not appear into the remarks of another ticket.
6. If You Only Do Three Things
1. First, organize the ledger into a machine-readable version. The chart of accounts has no duplicate definitions, tax codes have no legacy issues, bank reconciliation rules can automatically match 90% of transactions, and attachments correspond one-to-one with journal entries. These four things have nothing to do with AI, but they determine whether integrating AI will save effort or cause trouble.
2. Narrow posting rights to specific roles, and give AI only a read-only account.Posting is the most expensive action in an accounting system; it should not be an incidental capability. Who has the right to post should be written into the job description, not left to the default value of a permission group.
3. Treat "duplicate entry" as a dedicated check item. Supplier bill numbers are unique, import interfaces are idempotent, bank statements have deduplication keys, and duplicate vouchers are checked at month-end. If these four are done solidly, you won't lose out whether AI is integrated or not.
AI accounting will not make accounting disappear. It will make the question of "who has the authority to post entries" visible—because from now on, drafts may be written by machines, but a human still has to sign.
The list will change. This year it is these eight; next year there may be a few more, or a few fewer. The only thing a company can control is one thing: make its books so clean that no matter who comes in, they can read them, change them, and stop them.
Data sources in this article: OpenAI's 2026-11-18 ChatGPT for Accounting launch page and product documentation (eight partner connectors, four workflows, administrator enablement and inherited permissions, drafts require human approval, Business and Enterprise data not used for training by default, duplicate entry prompts); Xero's 2026-11-19 integration announcement (review in ChatGPT, actions back in Xero, currently limited to the United States); Intuit's limitation to QuickBooks Online in the United States; OpenAI's published customer cases (Lisa Kholdi saves 5 to 6 hours per month; Jessica Dean takes 17 to 20 minutes per company to prepare journal entry upload files; HSP GRUPPE's real estate investment analysis drops from about 9 hours to about 2 hours), all provided by vendors and not independently verified; Odoo 20 documentation on the built-in MCP server, tools exposed as server actions, and read-only by default. Odoo's state machine and permission groups are exemplified by the public documentation for versions 17 to 20; details vary slightly by version, so please check the documentation for the corresponding version before implementation.
Zhongdi Information · Observations on Enterprise Management Digitalization · The enterprise cases in this article are cited from public reports and are not accompanied by third-party audits.
