Over the past two years, when we talked about agentic ERP, it was mostly about: having AI generate a purchase requisition in the system, draft a reconciliation email, flag an abnormal inventory item. Humans were still in the loop, and that final “Confirm” button was in the employee’s hands.
In early November 2026, a section of this circuit was cut off.
At Hong Kong FinTech Week, Sunrate, together with Mastercard, released the world's first B2B agentic payment white paper, and launched the Sunrate Agentic Card—a virtual card designed primarily to be invoked by AI agents, which has been certified through Mastercard AgentPay and is already up and running in high-frequency cross-border scenarios such as OTAs. The agent is no longer just “recommending a payment in the system,” but actually holding the card, following permissions, and making a payment.
From “writing documents” to “moving funds,” it seems like only one step apart, but in reality they are risks of two different orders of magnitude. This leap directly turns the “governance gate” from an optional item into a mandatory one.
1. Sunrate lets an agent 'pay with a card' for the first time
First, let’s be clear about what this card can do. Enterprises can, in tasks such as procurement, travel, and supplier payments, generate virtual cards in real time, set custom spending limits, and set validity periods according to approved workflows. Authorized agents complete card creation, configure payment parameters, initiate payments, and query results; merchant category, amount caps, currency, geographic scope, and valid time windows can all be precisely defined, with identity and permissions fully traceable and auditable throughout.
The white paper breaks down the entire B2B payment lifecycle into 16 pain points that can be systematically solved by AI, and provides 13 actionable use cases, covering supplier sourcing, accounts receivable and payable automation, dynamic payment routing, exchange rate monitoring, automated reconciliation, and anti-money laundering screening. The fastest to implement is the OTA industry: marketing spending and supplier payments are scattered, with high risks of overspending and non-compliance; issuing virtual cards in real time for each campaign, making payments according to permissions, and screening anomalies in real time precisely address the problem.
Here is a detail worth remembering for everyone building enterprise systems: the prerequisite for an agent to move money is not how smart it is, but that permissions are cut finely enough and actions are recorded completely enough. In Xunhui's design, what is truly valuable is not “AI can make payments,” but that “the boundaries, limits, and audit trails of payments are all predefined and auditable.”
II. The gate is more important than the accelerator: SAP's governance gateway
Once an agent can "write" and even "pay" across systems, the biggest fear is not that it does one thing wrong, but that there is no unified place to keep watch over it.
SAP's Enterprise AI Gateway, announced at TechEd 2026, does the same thing: it consolidates authentication, policy enforcement, routing, monitoring, usage tracking, and auditing into one place. It is based on the open A2A and MCP protocols, allowing third-party agents to call Joule Agents through a "governed gateway" and enforce access control. It is accompanied by the AI Governance Assistant, designed specifically to help customers comply with the EU AI Act.
You can think of it as a three-in-one combination of the "corporate access control + monitoring room + audit department" of the agent world. Without this gateway, each agent connects to systems on its own and keeps its own logs; when something goes wrong, no one can tell whose hand moved the money. SAP is platformizing this set of capabilities, which shows the industry has reached a consensus: before agents scale up, install the brakes first.
This has a very direct implication for our ERP implementation: don’t rush to let the client experience the thrill of “AI automatically running processes”; first lock down these three things in the system: “who can adjust what, what gets recorded after adjustment, and how to roll back when something goes wrong.” If the brakes are not adequate, the more responsive the accelerator, the more dangerous it is.
III. Compliance Execution Must Also Enter ERP: Odoo 20's Multi-Country Localization
When an agent moves funds, compliance is unavoidable. And when it comes to compliance, what companies fear most is cobbling it together themselves.
The localization package in this Odoo 20 release is a good example: Mexico's CFDI / SAT; Saudi Arabia's ZATCA (directly integrated in the sending wizard); the UAE's VAT audit files and IFRS chart of accounts; Hong Kong's HKFRS reports and significantly upgraded payroll (including MPF and IR56); and China's VAT and surtax return forms and batch voucher printing. The upgrade platform is expected to open at the end of November 2026, and customers with custom modules installed must first run a test upgrade.
Odoo's approach is clear: make compliance a platform capability, rather than having enterprises find plugins themselves, connect to tax authorities, and gamble on API stability. For growing companies that want to go global or have multi-country entities, this means that when an agent generates cross-border documents or triggers payments, it can directly land within an already-validated compliance framework, rather than relying on manual after-the-fact addition of tax IDs and reports.
This also adds a natural guardrail to 'agent moving funds': when payment actions must pass through a localized compliance structure, the room for agents to misbehave is compressed.
IV. Three Gates for Growth-Oriented Enterprises
Back to implementation. Whether you deploy SAP, Oracle, Odoo, or some local agent platform, as long as you let the agent touch funds or key documents, I recommend first putting these three gates in place:
Permission granularity is the gate. Don't give the agent a master key for “company-wide payments.” Like Xunhui, issue virtual cards by task, set limits and validity periods by scenario, and make them invalid immediately after payment. In Odoo, you can use user groups, record rules, and approval workflows to split “can view,” “can create,” “can approve,” and “can pay” into different roles; the agent operates under the strictly constrained identity.
Audit completeness is the gate. Every step of the agent—card creation, configuration, payment, and query—must leave a trace and be traceable. An agent without audit logs is like driving blindfolded. Odoo's audit logs, email gateway records, and reconciliation receipts from external systems must link into a complete chain of evidence.
The gate of the rollback mechanism. If a person clicks wrong, it can still be undone; if an agent runs wrong, who will stop it? SAP's gateway and Xunhui's limit failures are all some form of "circuit breaker." Your own system must also have: abnormal thresholds trigger manual review, over-limit automatic interception, and incorrect payments can be returned via the original route or frozen. Rollback is not a post-event remedy; it is a must-answer question before going live.
From 'AI drafts an order for me' to 'AI pays the money for me,' what lies between them is not technology, but trust. Trust is not built on slogans; it is built on permissions fine enough to reach every pore, audits complete down to the byte, and rollbacks fast enough to be measured in seconds.
Only after crossing this step will enterprise digitalization truly transform from a 'system of record' into a 'system of execution.' But remember: being able to execute means being able to make mistakes. Install the gates first, then let the agent in—this is the sentence that, in Q4 2026, every enterprise preparing to let AI touch money should have carved on its wall.
