Let AI prepare intercompany exceptions, not close them

Agentic close changes the speed of investigation, not accounting authority

Intercompany reconciliation is difficult because two legal entities can describe one economic event with different records, dates, currencies, accounts, and evidence. AI can organize that disagreement. It cannot make the disagreement disappear.

Genpact announced general availability of its Record-to-Report Suite on September 1, 2026. The product description includes intercompany agents that detect mismatches, classify root causes, orchestrate resolution, automate accruals, and reconcile balances before close. The launch also publishes expected performance outcomes. Those figures are vendor-provided, explicitly described as indicative, and dependent on client scope and process maturity. They are not independent audit evidence and should not set a team's control design.

A current accounting discussion gives the less polished version of the problem: one entity emails a PDF while the other books the transaction manually a month later, so neither side has a consistent structured record to match. The practitioner suggests a mirrored purchase-order and sales-order model or a shared clearing process. That is one public example, not a market survey, but it exposes the control issue: exception automation is downstream of transaction design and population completeness.

Current Oracle documentation compares provider and receiver intercompany receivable and payable entries, supports common-currency reporting, and permits drill-down from a period summary to sources and journal lines. NetSuite tells teams to identify unpaired records and amount or currency mismatches before period-end elimination. SAP separates access to run matching, assign items, change reason codes, process reconciliation cases, upload records, and delete matching documents. These are system-specific implementations of one general principle: finding a likely pair, explaining a break, correcting a ledger, and releasing a close result are different authorities.

PCAOB amendments addressing technology-assisted analysis are effective for audits of financial statements for fiscal years beginning on or after December 15, 2025. Their focus is sufficient appropriate evidence and the reliability of electronic information. Accounting teams should apply the same discipline before an AI-prepared exception becomes part of management's close evidence: preserve where data came from, test completeness and accuracy, investigate identified items, and retain the exact work reviewed.

Use six states so “matched” never means “done”

Binary matched/unmatched labels hide the most important work. A system can find two records with the same amount while the partner, period, currency basis, or source event is wrong. Conversely, a valid economic pair may have different functional-currency values because each entity used a different approved rate date.

Preserve the transition history. If an item moves from “explained” back to “paired, unmatched” after a new extract, that is evidence of changed data, not workflow failure to hide. The record should show who changed the state, which evidence hash was reviewed, and why.

“Unknown” remains a valid root cause. AI systems are optimized to complete patterns, so they often turn a missing counterparty into a plausible timing explanation. Require exact source anchors for every verified fact and keep hypotheses visibly separate.

Separate population, pairing, judgment, execution, and release

The AI belongs mainly in the comparison and exception planes. It can standardize references, suggest candidate pairs, summarize two-sided evidence, detect recurring patterns, and draft a treatment packet. Deterministic code should recompute amounts and currency conversions. Policy services should apply materiality and routing thresholds. The accounting system should enforce posting rights and period locks.

exception_id: IC-2026-08-USDE-00418
provider: US01
receiver: DE04
provider_record: JRN-88421
receiver_record: AP-55019
pairing_state: paired_unmatched
transaction_currency: EUR
provider_amount: 125000.00
receiver_amount: 125000.00
provider_period: 2026-08
receiver_period: 2026-09
root_cause: proposed_timing
evidence_status: missing_receiver_service_date
materiality_route: controller_review
proposed_treatment: hold_pending_source_evidence
ai_authority: prepare_only
release_state: blocked

Do not store only the normalized record. Keep original provider and receiver values, the rule version that transformed them, and the resulting comparison. A change from accounting period 08 to ISO month 2026-08 is benign when documented. A sign inversion or entity remap can change meaning and must be reviewable.

Run the review from complete population to retested release

1. Freeze both populations

Record group, legal entities, ledgers, accounts, period, cutoff, currencies, report parameters, extraction time, row count, and debit/credit control totals. Reconcile each extract to an independent ledger report. A perfect match inside an incomplete export is still a failed control.

2. Pair with controlled keys

Use approved transaction IDs and intercompany references first. Composite matching may use entity pair, amount, currency, document type, and a policy-defined date window. Preserve one-to-many, many-to-one, and ambiguous states. A nearest-neighbor score is a search aid, not match evidence.

3. Compare the dimensions that change accounting meaning

Test transaction amount, transaction currency, functional amount, rate, rate date, posting period, source date, account, partner, sign, tax, and posting status. Oracle notes that accounted amounts can differ when each side uses a different conversion rate. Normalize to a common comparison currency only as an additional view; never discard the original currency facts.

4. Trace to source on both sides

Drill from the summary to journal, subledger, invoice, order, shipment, service, agreement, or settlement evidence. Ask why the provider recognized the event and why the receiver did or did not. One entity's approved journal is not proof of the counterparty's obligation.

5. Classify the exception without deciding it

Assign a controlled class and show supporting fields. Timing, amount, currency, duplicate, entity, partner, account, document, tax, transfer-pricing, posting, and elimination exceptions require different owners and actions. If support is insufficient, use unknown and identify the missing evidence.

6. Propose a treatment and route authority

A treatment may be wait for timing, correct the source, reverse and repost, fix mapping, accrue, resolve a dispute, review elimination, or escalate. State affected entities, accounts, periods, currencies, amounts, prerequisites, policy basis, reversal plan, and downstream effects. The AI does not produce an executable posting file.

7. Re-extract, reperform, and release

After authorized execution, obtain fresh extracts and repeat the comparison. Reconcile updated matched and unmatched totals to both ledgers. Confirm that a correction did not create a new period, currency, tax, or elimination issue. Bind final approval to the exact exception-set hash and system receipts.

Worked example: same invoice, different period and rate

US01 provides engineering services to DE04 for EUR 125,000. US01 records intercompany revenue and receivable on August 29. DE04 receives the invoice on September 2 and records the expense and payable in September. The provider converts using its August 29 functional-currency rate; the receiver uses its September 2 rate. The consolidation report shows both a period break and a functional-currency difference.

An AI matcher pairs the entries using entity pair, invoice reference, EUR amount, and document type. It proposes “FX timing difference” and recommends an accrual. That answer is plausible but incomplete. The reviewer still needs the agreement, service-acceptance date, invoice, both accounting policies, close calendar, rate sources, period status, and whether DE04 had an obligation at August close.

The evidence confirms that service acceptance occurred August 28 and the invoice was available to a shared mailbox August 29, but an interface failure delayed DE04 intake. Accounting policy requires an August accrual. The receiver controller approves an August accrual with automatic September reversal; an authorized operator posts it. Fresh extracts then align the transaction period, but the functional amounts still differ because each entity applied the approved local rate basis.

IFRS material on intragroup foreign-currency loans illustrates why “eliminate everything to zero” is not a sufficient instruction: intragroup balances and transactions are eliminated in consolidation, while required exchange effects may remain in profit or loss. The exact treatment depends on the fact pattern and applicable framework. The consolidation owner, not the matcher, decides whether the remaining difference is a valid exchange effect, a rate error, or another exception.

Map root cause to an allowed treatment, not a generic fix

SAP's permission model is a useful reminder that changing a reason code, assigning an item, uploading a matching document, and deleting a document are distinct actions. Mirror that separation in workflow roles. A preparer should not gain posting or deletion authority because the AI interface places the buttons together.

Release evidence must survive the next close and the next audit

The exception packet should retain frozen report identifiers and parameters; both original records; source-document anchors; pairing keys and alternatives; normalization rules; rates and dates; root-cause facts and hypotheses; approved treatment; operator receipt; retest results; remaining differences; reviewers; timestamps; and the final hash. A screenshot of a green dashboard is not enough.

PCAOB technology-assisted analysis materials emphasize information reliability and investigating identified items. Management's workflow is not an audit procedure, but the design lesson transfers: prove that the electronic population is complete and accurate for the purpose, then investigate the exceptions the tool surfaces. Do not tune the matcher to suppress inconvenient items after seeing the result.

Use preventive measures where recurrence evidence supports them: shared references, mirrored order flows, common partner master data, interface completeness monitors, rate-policy alignment, and pre-close aging. Keep the AI from changing pairing rules or tolerances based only on historical reviewer acceptance. A reviewer may have accepted an immaterial difference once; that is not authority to redefine policy.

• Require independent population totals before matching begins.
• Require two-sided source evidence for material root-cause conclusions.
• Keep reason code, accounting correction, system execution, and elimination approval separate.
• Reperform calculations outside the language model.
• Lock approval to one exception set, report version, and evidence hash.
• Re-extract after every material correction rather than editing the reconciliation output.
• Track post-close adjustments and reviewer overrides by root cause.

Common failure modes make a fast close less trustworthy

Run a 30-day pilot on one entity pair and one account family

Choose an entity pair with recurring but understood exceptions. Freeze four to twelve months of historical populations and replay the workflow without permitting postings. Compare AI candidate pairs and classifications with signed prior conclusions. Include duplicates, one-to-many settlements, missing references, different periods, different currencies, and intentionally withheld evidence.

Measure population completeness, exact-pair precision, false-match rate, ambiguous rate, root-cause support rate, unknown rate, reviewer correction rate, minutes per exception, aged balance, recurrence, post-close adjustment, and duplicate action count. Segment by exception class; one aggregate accuracy score hides dangerous rare cases.

Move to supervised current-period use only if population tie-outs are complete, no material false match is accepted, all executable actions remain segregated, retesting catches seeded errors, and reviewers can reconstruct each conclusion from source evidence. Set stop conditions for missing reports, schema drift, unsupported currencies, changed policy, unexplained override growth, or any unauthorized system action.

Frequently asked questions

Can AI automatically resolve small intercompany differences?

Only if an approved accounting policy defines the threshold, exception class, evidence, allowed action, segregation of duties, and monitoring. The language model should not invent or change that policy. Even immaterial items may reveal a recurring interface or master-data failure.

Should all intercompany balances eliminate to zero?

Intragroup balances and transactions are generally eliminated in consolidated statements, but specific accounting effects can remain, including certain foreign-exchange effects. Consolidation treatment depends on facts and the applicable framework; the matcher should prepare evidence, not issue the conclusion.

What if one entity has no matching structured record?

Keep the item unpaired, trace the source event and transmission path, and obtain or create an authorized system-of-record entry through normal accounting controls. Do not let the AI synthesize the missing side from the provider's record.

Is this the same as ordinary account reconciliation?

No. Ordinary reconciliation often compares one account to an external or subsidiary record. Intercompany review must align two controlled legal-entity populations, partner dimensions, currencies, periods, and consolidation treatment, with owners on both sides.

Can the workflow support US GAAP as well as IFRS?

Yes as an operating pattern, but accounting conclusions, elimination rules, tax treatment, materiality, and documentation must be configured to the organization's applicable framework and policies. The pack deliberately does not encode one universal accounting answer.