Last week, while talking with a CFO about a system upgrade, I mentioned that some ERP systems can now make payments directly within the system—no exporting files, no logging into online banking. He paused for half a second, didn't ask about features, and only asked one question: Then who can send it?
This question is far more important than "whether it is easy to use." On November 17, Odoo announced on its official website that its partnership with Salt Edge would expand from "retrieving bank data" to "initiating payments." Odoo became the first live customer of Salt Edge's bulk payment solution Bulk Payments, and the solution expanded from Italy to the wider EU and the UK. The announcement talks throughout about the experience: single payments can be initiated and verified directly in Odoo, while batches can be sent with a single authorization, with no need to export payment files or re-enter them in online banking.
The experience has indeed improved. But the finance director was asking about another layer: in the past, for a sum of money from application to disbursement, how many people it had to go through and how many traces were left; now a single authentication covers an entire batch—how many people are left on this chain.
1. What this announcement changes is not the interface, but the authorization chain
Looking at this on a timeline makes it clearer. Odoo and Salt Edge's banking data partnership began in 2021; in those years, what was solved was "visibility": bank transactions came in automatically, and reconciliation no longer required manual entry. What is being crossed now is "being able to pay out"; this step moves funds, not data.
Salt Edge's credentials were also disclosed: its AISP license under PSD2, ISO 27001 certification, and connections to more than 5,000 financial institutions in over 50 countries. These are necessary qualifications, but they all fall at the level of "whether one can connect," and cannot answer "who can issue."
The seven things not in the announcement are exactly what finance cares about most: which Odoo versions are supported, which countries other than Italy are already available, the maximum number of transactions per batch, whether there is two-person review, how partial failures in a batch are handled, how returned payments and fees are recorded, and how payment records after initiation are automatically matched with bank statements. The fact that these seven items are not written in the press release does not mean they have not been done, but when preparing a proposal, one must not make assumptions based on the wording of the press release.
Why do these seven items matter? Because the risk structure of batch payments is different from that of single payments. If a single payment is wrong, one payment is wrong; if a batch is wrong, a batch is wrong. Moreover, errors in a batch occur simultaneously—19 payments are made correctly, and the 20th is paid to the wrong account; the outgoing funds happen at the same time, but recovery does not. The larger the amount, the more deadly this difference.
II. A sum of money originally passed through four people, but now passes through one action
In traditional processes, the matter of payment is artificially split into several segments, each corresponding to a role; this is not process complexity, but internal control.
| Action | Who traditionally does it? | What does it correspond to in Odoo | What happens if this section is skipped? |
|---|---|---|---|
| Registration | Accounts Payable Accountant: Generate Payment Order from Invoice | Payment (account.payment), Batch Registration Wizard (account.payment.register) | Paid a bill that shouldn't have been paid, or paid twice |
| Review | Finance supervisor: check the amount, account, and purpose | Approval nodes are usually implemented in the system through status or a dedicated approval workflow. | Accounts were altered, amounts were entered wrong, and no one stopped it. |
| Send | Cashier: export files, log in to online banking, enter data, confirm | Payment method configuration, bank interface, or file export | Payee information does not match the actual details |
| Write-off | Accounting: Collect receipts and reconcile transaction records | Bank statement line (account.bank.statement.line) matched with journal entries | Paid-but-unsettled items sit on the books, and month-end closing drags on longer and longer. |
Moving the "send" step into the system has the most direct effect of removing one manual step; the most easily overlooked effect is that the boundaries between the first three stages have become blurred. Because on a single interface, after registering you can conveniently click send; with the same batch selected, one authentication, dozens of transactions go out at once. The authorization chain has not disappeared because of this, but it has changed from "between positions" to "between buttons"—and isolation between buttons relies on permission configuration, not on people's self-discipline.
The one place to watch most closely: If "Create Payment Order" and "Approve Payment" are given to the same role, then there is no review in the system. This is not an Odoo problem; it is a configuration problem. In the standard, these two actions belong to different permission groups, but as soon as someone, to save effort, puts the finance supervisor into a larger group, the separation is gone—and you cannot tell from the interface.
3. In the same quarter, two other things came crashing in
There is another layer to this. Odoo 20 was released in September this year. According to a third party's summary of the official release notes, this version made two changes on the finance side and one change on the AI side, and all three landed in the same quarter:
- Payment statuses have been renamed. The original "Processing" is now called "Paid", and the original "Paid" is now called "Reconciled". The semantics are more accurate, but all reports, automated actions, and external interfaces that hard-coded the old status values will silently calculate incorrectly after the upgrade—it won't report an error; it will just include a different batch of documents in the calculation.
- Bank consistency requirements are becoming stricter. Every journal entry involving a bank account must originate from a bank transaction. The direction of this rule is correct, but when combined with "initiating payments directly within the system," it creates a new reconciliation problem: when and in what form should a payment initiated by the system be written back as a bank transaction?
- The agent gained the ability to create and modify records. In the same version, the AI agent can create and modify records according to instructions, and can also be invoked by external assistants through MCP.
The first two matters are accounting problems, and the third is a people problem. Looking at them stacked together, a scenario that did not exist before emerges: an agent with write permissions, through a channel capable of initiating payments, operates funds on a shortened authorization chain. None of these three is new on its own, but together they constitute a new risk.
This is not to say that this situation will definitely happen, but that it is now technically possible, while most enterprises' permission models are still designed around "people" and leave no place for "agents acting as people."
IV. Four Actions, Authorized Separately
If this upgrade is going to touch the payment flow, I suggest first splitting permissions by action instead of packaging them by role. For how to split them, refer to this table:
| Permission point | Who is it recommended for | Supporting constraints |
|---|---|---|
| Create Payment Slip | Accounts payable accounting, scoped by company/journal | Accounts outside the bank journal cannot be selected; changes to supplier bank accounts require review by another person |
| Approve Payment | Finance supervisor, tiered by amount | Cannot be the same person as the creator; if the threshold is exceeded, it is mandatory to escalate to the next higher level |
| Initiate/Send Payment | Cashier, independent account | Upper limits on single-transaction and single-batch amounts; payee account whitelist; abnormal accounts must be manually confirmed |
| Write-off and Adjustment | Accountant, excluding send permission | Reversals must include a reason; cross-period adjustments require approval |
Add one more measure: give the agent a dedicated account, granting only "read + draft," not "create payment orders," and certainly not "send." That way it can still help you organize a batch of invoices that are nearing their due dates and line up payment recommendations, but the money cannot go out. How far to open it up is something to discuss after you have run it for a few rounds on real data and know where it will make mistakes.
An actionable acceptance step: Disable the cashier's account. If the payment flow immediately stops, it means permissions are designed around the action; if the process is still running, it means someone else (or another account) can also send money—that means it wasn't designed.
V. Sending became faster, but the redemption side did not
There is another consequence that proposals usually don't mention: payment is the first half of the chain, and reconciliation is the second half. The first half is compressed from "a day" to "a minute," but the work in the second half hasn't decreased at all—receipts must be matched, discrepancies investigated, and outstanding items cleared. The time saved will pile up unchanged in the week before month-end closing.
The Odoo 20 rule that "every journal entry involving a bank account must come from a bank transaction" is pushing reconciliation forward: entries cannot be created out of thin air; they must be tied to a real bank statement line. The rule itself is not wrong, but it shifts the pressure from "payment" to "whether bank data is fast enough and complete enough." If statement lines are synchronized only once a day while payments are issued in real time, then money paid out in the afternoon of that day cannot be found in that day's bank statement, and a batch of "paid but no statement line seen" records will appear in the books.
These records are not errors; the trouble is that no one is managing them. At month-end closing, they turn into a pile of suspense items that require manual judgment, and the larger the amount, the harder they are to explain. So when payments go live, a rule must be put in place at the same time: for unmatched payment records at T+1, who checks them within what time limit, to what extent they count as cleared, and to whom they are escalated if overdue.
VI. Before launch, first use a small amount of money to run through the entire process.
1. Use a small real payment to go through the five nodes: "registration—approval—sending—receipt—write-off." The point is not the successful attempt, but the failed one: deliberately enter one account incorrectly, deliberately let the bank reject once, and then see what records are left in the system, who can see them, and how to correct them. All the risks of batch payments are hidden in the failure path.
2. Before upgrading, check every hard-coded status value in reports. Especially those management reports, automated actions, and external interfaces that filter by "Paid". Renaming a status won't throw an error—it will only silently swap the definition. Compare an old report from last month with the results produced by the new system line by line, and the differences will jump out on their own.
3. Establish a separate process for "supplier bank account changes." After payment channels become faster, the most valuable target is no longer invoices, but receiving accounts. Changes must be confirmed by two people, the old account number must be retained, there must be a cooling-off period, and it is best to also require the supplier to call back through another channel to confirm. This process does not require system modifications; it needs to be written into the policy and accompanied by a mandatory field in the system.
That finance director said one last thing: In the past, it wasn't that we didn't want to be fast; we didn't dare to be fast, because no one had our back at that step of going fast. That remark hit the crux—when it comes to payments, the value has never lain in 'being able to send it out,' but in 'being able to get it back if it's sent wrong.'
The system has compressed those intermediate steps, which is a good thing. But if what was compressed is approval rather than data entry, then what is saved is not efficiency but control. Which one it is depends on how those few checkboxes in the configuration are ticked.
Data sources in this article: the Odoo official blog post 'Odoo expands partnership with Salt Edge and is the first to go live with Bulk Payments' (2026-11-17, including statements by Olivier Maessen and Vladimir Modirchi, Salt Edge's PSD2 AISP license and ISO 27001 certification, coverage of more than 5,000 financial institutions and more than 50 countries, and the path for expansion from Italy to the EU and the UK); the changes in Odoo 20 regarding payment status naming, bank consistency, and agent creation/modification records are quoted from a third-party service provider's compilation of the official release notes (checked in November 2026); before implementation, please refer to Odoo's official release notes as the authoritative source. Odoo model and permission group names use versions 17 to 19 as examples; they may be adjusted in version 20.
Zhongdi Information · Enterprise Management Digitalization Observations · Configuration details not disclosed in vendor announcements are subject to official documentation and confirmation by the local implementation provider.
