AI can triage GRNI exceptions. Accountants must prove the balance.

GRNI is a population and timing control before it is an aging report

Infor's October 2026 release adds a Financial Accounting Agent that analyzes bank accounts, general-ledger activity, and goods-received-not-invoiced reconciliation. The same release lists invoice-match and period-close agents. That is a useful product signal: enterprise vendors expect AI to sit inside close and reconciliation work. It is not independent proof that an agent can certify a GRNI balance.

GRNI, also called GRIR or accrued receipts in some systems, exists because the economic and accounting events do not arrive together. Goods or services are received; the organization recognizes inventory, expense, or another debit and a temporary liability or clearing balance. The supplier invoice later matches the receipt and relieves the temporary balance. Returns, partial receipts, price differences, credits, tax, currency, and timing can break that expected path.

The report is therefore not the control. A GRNI listing can omit interface failures, manual journals, unmatched returns, invoices recorded outside the purchase-order path, or old records stranded by configuration changes. Before any model explains an item, the team must prove that the complete receipt, invoice, return, credit, subledger, interface, manual-adjustment, and general-ledger populations are in scope.

Exact community discussion of Infor's new agent was thin in the focused scan, so this guide makes no adoption, accuracy, ROI, or close-time claim. The workflow is anchored in Infor's first-party release and reconciliation sequence, cross-vendor ERP documentation, internal-control principles, and clearly bounded practitioner examples.

Build four control layers around the agent

source populations
PO | receipt | invoice | return | credit | journal | ledger
                         |
              deterministic control plane
        completeness | matching | rollforward | cutoff
                         |
                exception evidence plane
          source rows | age | owner | contradictions
                         |
                   AI analysis lane
      classify | summarize | draft query | propose route
                         |
                  human decision gate
       correct | post | credit | write off | close
                         |
                     rerun proof

The source layer preserves original system truth. Export each population with entity, period, account, currency, query or report version, extraction time, row count, control total, and file hash. A normalized table may standardize identifiers and signs, but it must retain every original value and source pointer.

The deterministic layer performs arithmetic and joins. It proves completeness against independent totals, calculates each receipt-level state, reconciles opening plus movements to closing detail, ties closing detail to the ledger, and tests cutoff. These checks should run without an LLM so another accountant can reproduce them.

The evidence layer holds the residual exceptions. Each record names the source rows, amount, currency, aging dates, exception type, contradictory evidence, materiality, owner, deadline, and decision authority. An “other” bucket is temporary and must have an owner who improves the rule.

The AI lane works only on the evidence packet. It can suggest that a receipt and credit may relate, summarize communication, draft a supplier question, or propose routing. It must cite records and label uncertainty. It cannot change calculations, create missing evidence, post, or approve.

The human gate decides what the evidence means and what action is authorized. The final layer reruns the deterministic checks after execution. Without the rerun, the team knows an action was approved but not whether it fixed the intended balance.

Use two equations: one for detail movement and one for ledger completeness

The first equation explains the subledger movement inside the frozen scope. Sign conventions vary, so document them. The important point is that each term is a controlled population rather than a model estimate.

opening GRNI detail
+ new receipt accruals
- invoice relief matched to receipts
- return and cancellation relief
- approved supplier-credit relief
+/- controlled reclasses and corrections
= closing GRNI detail

The second equation compares the closing detail with the general ledger using the same entity, account, currency, period, and cutoff. Manual journals and interface balances must be explicit. If the difference does not tie to the approved rule, stop classification and repair the population or mapping.

closing GRNI detail
+ identified interface-in-transit balance
+/- approved ledger-only timing items
= GRNI general-ledger balance

unexplained difference must equal 0
or the entity's documented immaterial tolerance

Do not use tolerance as an automatic write-off. A small net difference can contain large offsetting errors. Report gross unmatched debit and credit values, count, supplier concentration, and aging alongside the net. Investigate unusual manual journals even when the final balance ties.

Cutoff needs event time and accounting time. A receipt at 23:55 may enter the system after the period closes. An invoice dated before cutoff may post later. A return can occur after receipt but before invoice. Record the physical or service-completion date, source-entry timestamp, accounting date, interface time, and period. The policy determines treatment; the AI only organizes the facts.

Classify by next evidence and authority, not by a vague reason code

Age by more than one date. Receipt age shows how long the liability has existed. Expected-invoice age reflects terms. Exception age shows how long the break has been owned. Last-evidence age shows whether investigation is active. These answer different questions.

AI can be useful when descriptions, correspondence, and history are messy. Require the model to quote source rows, list missing evidence, present at least one alternative explanation, and draft the next question. Do not accept “likely duplicate” without showing both records and the deterministic duplicate rule.

Routing must follow authority. AP can obtain an invoice or credit but may not decide an accounting write-off. Receiving can confirm physical return but may not post a ledger correction. The controller can approve treatment but should not create and approve the same journal. Preserve segregation of duties even when the agent prepares the packet.

Worked example: one old balance contains three different problems

A manufacturer has a $184,000 GRNI item aged 118 days against one purchase order. A simple aging rule marks it “write-off candidate.” The controlled workflow starts lower.

The PO line shows equipment and installation. Two receipt events total $184,000: $150,000 for equipment and $34,000 for service completion. The supplier invoice for $150,000 matched and relieved the equipment receipt. The service invoice is absent. A later return record for $10,000 references damaged equipment, but the supplier credit has not arrived. A manual journal moved $34,000 from the GRNI account to a project accrual and is set to reverse next period.

The net old balance hides separate states. The service receipt may represent a valid unbilled obligation. The return may require a supplier credit or receipt correction. The manual journal creates a ledger-only timing difference and potential double accrual when it reverses. The AI may assemble the timeline and draft questions, but it cannot choose one write-off treatment for the net.

1. AP confirms whether the service invoice exists or is disputed.
2. Procurement obtains the supplier's acknowledgment of the return and credit status.
3. Receiving verifies the return quantity, date, and receipt linkage.
4. Accounting traces the manual journal, approval, project accrual, and reversal behavior.
5. The controller approves separate treatments for each component.
6. The team executes source corrections and approved entries, then reruns detail and ledger tie-outs.

The close packet retains the original item, component evidence, decisions, transaction IDs, and before/after results. If the late service invoice arrives next month, the system reopens the linked item rather than treating it as unrelated AP activity.

Define what the AI may prepare and what only humans may decide

Bind approval to the exact version. A reviewer signs the scope, source hashes, control totals, exception register, decisions, execution receipts, and rerun. If data or rules change after review, the approval no longer covers the new result.

Retain rejected proposals. They show what the AI suggested, what evidence was missing, and why the accountant chose another path. This is especially important when the same pattern recurs and teams are tempted to convert a judgment into an automatic rule.

Never learn a posting rule directly from an approved exception without governance. One entity's policy, materiality, supplier relationship, tax treatment, and cutoff fact may not transfer. Proposed rules need versioning, owner, test fixtures, impact analysis, approval, and monitoring.

Failure modes that create a false clean close

A false clean close is worse than a visible exception because it removes the prompt to investigate. Sample cleared items, especially automated matches, repeated supplier patterns, manual journals, and transactions just below tolerance. Monitor rule drift after ERP upgrades, chart changes, acquisition, new sites, currency changes, and model or prompt updates.

Reopen behavior matters. Late invoices, credits, returns, disputes, or source corrections can change a prior conclusion. Preserve the released reconciliation, create a linked correction, assess downstream inventory, expense, payable, tax, reporting, and forecast impacts, and issue a replacement when required.

Run one close in shadow mode before granting any action authority

1. Select one entity, one GRNI account family, and one reporting period.
2. Freeze the approved manual reconciliation and source snapshots as the baseline.
3. Build deterministic population, rollforward, ledger, and cutoff checks.
4. Let AI classify and draft only after the checks pass.
5. Compare every AI classification with accountant disposition; retain false positives and misses.
6. Measure reviewer time, evidence completeness, owner response, and reopen events.
7. Test missing source, duplicate receipt, return without credit, manual journal, currency mismatch, and late invoice fixtures.
8. Repeat after an ERP/configuration or prompt change.
9. Approve only the target classes and outputs that passed; keep posting denied.

Measure complete, approved outcomes rather than vendor claims: percentage of total population proven, ledger difference, gross and net exceptions, aged value by cause, first-pass deterministic match, AI classification agreement, unsupported-claim rate, owner response time, review minutes, correction rate, and reopened-item rate.

Segment by supplier, site, entity, account, currency, materiality, exception type, and system version. A falling item count can hide concentration in a few large suppliers. Faster classification can create more reviewer work if packets lack evidence. A high agreement rate can be meaningless if accountants copy the model without independent review.

Set promotion criteria in advance. For example: exact population and ledger tie-out on every run, zero unauthorized action attempts, 100% citations on material AI classifications, all adverse fixtures detected, replayable outputs, stable results across two closes, and controller approval. If the workflow misses any criterion, keep it in shadow mode and fix the control.

FAQ

Can AI certify a GRNI reconciliation?

No. AI may classify and organize exceptions, but accountants must prove population completeness, ledger tie-outs, cutoff, corrections, approvals, and rerun results.

What is the minimum control total?

Use complete receipt, invoice, return, credit, subledger, interface, manual-adjustment, and ledger populations for the frozen entity, account, currency, period, and cutoff, with opening and closing reconciliation.

Should old items be written off automatically?

No. Age prioritizes investigation; it does not establish that the obligation disappeared. Resolve goods, services, invoice, return, credit, dispute, policy, and source-record facts first.

Can the AI post journals or close the period?

No. Posting, write-off, source correction, supplier-credit acceptance, policy judgment, and close remain in authorized human and ERP controls.

How should configuration changes be handled?

Version extraction logic, account mappings, tolerances, matching rules, model, prompt, and output schema. Run regression fixtures and obtain approval before applying the new version to close.

Sources and further reading

• Infor 2026.10 Industry AI release — Financial Accounting, invoice-match, bank-reconciliation, and period-close agent claims; accessed Oct 20, 2026.
• Infor: the GRNI reconciliation process — period freeze, reports, analysis, corrections, and rerun sequence.
• Infor LN Reconciliation and Analysis user guide.
• SAP Invoice and Goods Receipt Reconciliation — cross-vendor GR/IR open-item reference.
• SAP invoice reconciliation guide — receipt and invoice-line exception mechanics.
• COSO Internal Control Framework — control environment, activities, information, and monitoring.
• NIST AI Risk Management Framework — AI governance and monitoring reference.
• Dynamics Community: accrued purchases without matching GRNI report item — practitioner case, not authoritative accounting guidance.
• Planergy GRNI reconciliation process — current practitioner explanation of aging and exception pain.
• Temporal Correctness Contract for Streaming Financial Reconciliation — current research on timing and reconciliation correctness.
• Related: account reconciliation workflow.
• Related: procure-to-pay exception workflow.
• Related: journal-entry review workflow.