A system of record (SOR) is the single authoritative source for a specific type of business data, the place where, if a number or record disagrees between two systems, the system of record is treated as correct. For accounts payable and financial data, this is typically the ERP or accounting system, the definitive record of what invoices exist, what has been paid, and what the general ledger shows.
Most businesses run many systems that touch the same underlying data: an ERP, a CRM, a separate expense tool, spreadsheets used for planning, and increasingly, AI tools layered on top of all of them. When the same piece of information, a customer balance, an inventory count, a vendor's payment status, appears differently across these systems, there needs to be one designated answer for which version is correct. That designated system is the system of record.
Without a clearly established system of record, discrepancies between systems become genuinely unresolvable disputes rather than simple corrections, since no system has the authority to say which figure is right.
These are frequently different systems working together: a CRM might be the system of engagement salespeople use daily, while the ERP remains the system of record for the actual revenue figures once a deal closes and becomes a financial transaction. The two are connected, but they serve different purposes and are not interchangeable.
As AI agents increasingly read from, write to, and act on business systems, the question of which system remains authoritative has become newly important. An AI agent that quietly builds its own shadow database of vendor information, invoice status, or payment history, separate from the ERP, creates exactly the same risk as any other unauthorized parallel record: a second version of the truth that can drift from the actual system of record and create confusion about which figure is correct.
This is a genuinely current concern in enterprise software: whether an AI agent should be treated as a new system of record in its own right, or whether it should operate strictly within the existing system of record, reading and writing to it directly rather than maintaining a separate, parallel copy of the data it works with.
The safer architectural pattern, and the one that avoids creating a new reconciliation problem, is for automation to read from and write directly into the existing system of record rather than maintaining its own separate database of AP data. An AI agent that posts invoices directly into the ERP, using the ERP's own screens and data structures, keeps the ERP as the single authoritative source it has always been. An agent that instead builds its own database of invoices and payment status, syncing periodically with the ERP, creates a second system that can disagree with the first.
This distinction is exactly why computer-use automation, operating the ERP through its own interface rather than building a parallel system, preserves the existing system of record instead of creating competition for that role.
For most businesses, the ERP or accounting system is the system of record for invoices, payments, and the general ledger, even when other tools, an approval workflow app, a spend management dashboard, a reporting tool, are used alongside it day to day. Any of those supporting tools should ultimately defer to the ERP's data as authoritative, rather than maintaining figures that could diverge from what the ERP actually shows.
