Month-end close: let an agent chase the evidence, not approve the journal

Give finance an accrual preparation pack that connects each purchase order to its responsible owner, delivery evidence and applicable procedure. Let an agent assemble that pack and flag missing information. Keep accounting judgments and journal-posting approval with finance. That is the bounded pilot we recommend—not autonomous accounting.

The starting point is concrete. In an account published on August 13, 2026, AWS and Rivian describe purchase-order accrual work involving ERP extracts, spreadsheet models, emails to owners confirming delivery dates, and accrual calculations. Evidence collection was part of that workload, rather than a separate administrative exercise. Read the AWS and Rivian implementation account.

Automate the chase, not the accounting conclusion.

What Rivian’s workflow establishes

The documented implementation goes beyond gathering confirmations. Its agent retrieves procedures from a knowledge base, routes purchase orders according to materiality, interacts with SAP and email tools, calculates accruals, and creates parked journal entries for finance review while managing approval workflows. A parked entry is the review handoff described in the account, not evidence of an independently approved posting. AWS describes the architecture and review handoff.

The account reports eliminating more than 15 days of manual work per cycle. Treat that as a company-specific, self-reported result across the implementation—not 15 days removed from the close calendar, and not a forecast for your team. The reported figure does not isolate the contribution of delivery-evidence collection. See the reported outcome.

KPMG’s undated analysis describes coordinated agents continuously monitoring transactions, flagging exceptions and preparing draft entries across existing finance systems. The page describes a broader operating model, not a measured customer result. Read KPMG’s description of the agentic close.

1. Define the worksheet and the rules

Choose a bounded category of purchase orders from an existing accrual worksheet. Ask finance to specify the accounting policy, acceptable supporting evidence, escalation rules and approval responsibilities before connecting an agent.

For this proposed pilot, make the inputs explicit: purchase-order and line identifiers, responsible owners, relevant periods, existing delivery records and the approved procedure version. Identify the authoritative source for each field and the person allowed to correct it.

We would begin by mapping those relationships into a small shared business data model. The scope is the worksheet’s preparation process, not a replacement ERP. Leave accounting treatments, materiality thresholds and calculation methods for finance to establish; do not copy them from another company’s deployment.

2. Collect evidence without manufacturing certainty

Give the pilot read-only access to the agreed records. Initially, have it prepare requests for missing confirmations for a person to review before sending. Any later permission to send should be limited to approved recipients, wording and follow-up rules.

Require every collected response to retain its source, receipt time and purchase-order link. Distinguish an owner’s statement from an attached delivery record; let finance decide which evidence is sufficient.

Our proposed exception handling is deliberately explicit. An unknown owner goes to the designated process owner. Conflicting dates go to finance with both sources attached. An unanswered request remains unresolved. A document that cannot be matched confidently stays outside the completed pack. Do not let the agent infer delivery from silence.

3. Prepare the handoff, not the posting

The output should be an annotated worksheet and evidence pack. For each item, show the supporting records, procedure reference, missing information and recommended review route.

Keep the first pilot outside journal creation and posting. Finance should decide whether an accrual is required, determine its treatment and amount, and approve any journal through the existing process. These are our recommended pilot boundaries, not a description of Rivian’s access controls.

The trade-off is intentional: a narrower pilot leaves more work with finance, but asks a more precise question—does evidence collection make preparation easier without increasing review burden? Do not mistake that boundary for a complete close redesign—or require a complete redesign before testing the handoff.

4. Agree the acceptance test before running it

Compare the existing process and the proposed workflow on the same agreed purchase-order set, including incomplete and conflicting cases. Measure preparation time through reviewer acceptance, not just time to an initial draft.

Record reviewer corrections by type and count unresolved exceptions at the agreed cutoff. Separate active preparation time from waiting for replies. Require traceable supporting records and no unauthorised journal writes as acceptance conditions.

Set improvement targets with finance after measuring the baseline; do not borrow the reported savings. We would leave cost and implementation duration unestimated until ERP access, owner records, evidence sources and security constraints have been examined.

Your next step is to select an accrual worksheet and define what a reviewer must receive to accept it. Expand only when the evidence pack saves preparation effort without shifting hidden work into corrections or unresolved exceptions.