How to integrate AI into your existing ERP stack

The ERP is the system of record for financial data. Chart of accounts, GL, AP, AR, fixed assets, inventory, and the audit trail on every posted transaction. The category was built around exactly that, and the modern entrants in the space (NetSuite, SAP S/4HANA, Sage Intacct, Dynamics 365, Oracle ERP Cloud, QuickBooks Online, Xero) all do it well. The ERP knows what was posted and when, and it gives the auditor the answer when the auditor asks.

The ERP is not for the manual operations that surround the data. It does not run the bank reconciliation, draft the variance commentary, chase the missing receipt, follow up on the AP exception, or assemble the close pack. Those are operator workflows that happen between the ERP and a dozen adjacent systems, and the operator running them is mostly the controller plus the AP clerk plus the FP&A analyst. The hours add up to roughly forty percent of the finance team's week, and they show up nowhere in the ERP's feature list.

The agent connects to the ERP through the same OAuth or token authorization the controller already grants third parties. It reads the GL, the open AP queue, the AR aging, the bank feed, and the payroll journal. It builds the close pack, identifies variances, drafts the commentary, prepares the journal entries the controller would have written by hand, and writes back through the standard import templates. Nothing post silently. The controller approves, and the JE posts on approval.

The integration pattern is on top, not in place of. The customer's audit posture stays exactly where it was, the customer's chart of accounts stays exactly where it was, and the customer's existing finance team keeps doing what they were hired to do, which is the judgment work, not the data entry.

First objection: data security. The agent reads the same fields the controller reads, under the same authorization the controller grants, with row-level security enforced at the database level. The customer can revoke the authorization in one click. Second objection: audit. The agent writes one hash-chained audit log row per action, and SOC 2 evidence packs lift directly. Third objection: AI hallucination. The agent does not hallucinate journal entries; it drafts entries against the controller's chart and the controller approves before posting. Fourth objection: change management. The controller's day does not change shape. The agent removes the manual reconciliation hours, not the controller's authority.

Connect the ERP. Pick one workflow (autonomous month-end close is the highest-leverage starting point). Run it once for the prior month, in a sandbox if the customer prefers. Compare the agent's output to the controller's actual close from that month. Calibrate the prompt for the customer's variance commentary voice. Switch to live for the next month. Measure the close days. The number should drop on the first cycle.