Consider a hypothetical close: one subsidiary reports a $100,000 intercompany receivable for service recharges, while its counterparty reports $96,000 payable in the same transaction currency and period. Before anyone can clear the $4,000 difference, somebody has to find which postings explain it. Sharing a parent apparently doesn't guarantee sharing an invoice reference. Oracle's intercompany reconciliation documentation describes a familiar target for this work: identifying the provider or receiver side of a transaction that hasn't posted. AI has a useful assignment here if it can shorten that investigation while leaving the difference visible.

Start with the matching rules the team already uses. Oracle documents date and amount tolerances, plus one-to-one, one-to-many, many-to-one, and many-to-many matching. Those capabilities provide a baseline for evaluating any AI proposal. Run the approved rules first, record what they clear, and preserve the unresolved population. A model suggesting that two vaguely similar descriptions belong together should have to explain what it adds beyond those rules. Increasing a tolerance until a discrepancy disappears would improve the dashboard while leaving the accountant with exactly the same problem, now harder to find.

For a pilot, choose two entities and one transaction family, such as monthly service recharges. Tie each scoped extract to the corresponding ledger or subledger totals and document any exclusions. Preserve source totals, stable transaction IDs, posting and document dates, currencies, and counterparty mappings. Oracle's reporting guidance shows why the currency basis needs attention: accounted amounts can differ when the two sides use different conversion rates. Keep original and normalized values available together. In a PE-backed group, limit access to the authorized company group; a shared sponsor doesn't make unrelated portfolio companies one permitted dataset.

Give the model a specific job inside that boundary: retrieve the approved invoice and journal records for an unresolved item, propose candidate counterpart entries, and explain the possible cause using links to those records. It should identify competing candidates and missing evidence. BlackLine describes rules-based imbalance categorization, predictive AI guidance, and structured exception workflows across multiple ERPs, so controllers can already examine commercial products that combine these functions. Those are vendor-described capabilities; the product page doesn't establish an AI-specific financial return. The workflow proposed here still needs testing against the company's own unresolved cases and existing tools.

Build the queue around what each entity accountant needs to do next. A suspected missing posting goes to the entity that needs to confirm it, with the counterparty's originating record attached. A timing item needs the relevant posting dates and an owner to follow it through the next period. Mapping and currency exceptions need the original codes or rates beside the proposed explanation. A suspected duplicate needs both source entries. When the evidence doesn't support a conclusion, keep the item open and say what's missing. For the hypothetical $4,000 difference, the deliverable is a supported explanation that accounts for the amount, with someone responsible for resolving it.

The accountants responsible for both entities should confirm the explanation and their respective balances before closing the case. Any correction goes through the company's existing authorization process, after which the team reruns the reconciliation and retains the supporting records and approvals. Give this pilot read access and recommendation authority; keep journal posting and settlement with authorized people. That follows the treasury approval boundary between preparing work and executing it. A reviewer should be able to reject a plausible proposed match without losing the underlying exception or having to reconstruct the investigation from scratch.

Measure the entire investigation, including review and rework. Compare reviewer minutes per case, false matches, aged differences, reopened cases, and total cost with the rules-only baseline. Include integration, licenses, model usage, and maintenance. Time saved becomes capacity; cash savings need an actual change in spending. Transfers between consolidated subsidiaries don't create new consolidated cash or EBITDA. The consolidation rules eliminate intragroup balances, income, expenses, and cash flows. Report any verified reduction in external fees or borrowing costs separately, with its own evidence, so cleared balances don't become an imaginary return.

Begin with historical cases whose outcomes the team can verify, withholding those outcomes and any later records from the model. Include awkward cases and sample accepted proposals as well as rejected ones, because an agreeable reviewer can let a convincing false match through. Then run the bounded workflow alongside the existing process for two closes, with success criteria agreed in advance. Expand to another entity pair only if reviewers spend less time overall, error rates stay within agreed limits, and unresolved balances remain visible. Each close should also identify recurring causes the team can fix upstream, so next quarter's investigation queue has fewer reasons to exist.