AI-Powered Reconciliations: Bank, Intercompany & Balance Sheet

By Christian Matumona, Kyootek AI Financial Controller

Reconciliation is the part of the close that nobody enjoys and everybody has to do precisely. The actual matching — confirming that two records agree — is mechanical. The value is in the exceptions: the items that don't match, and figuring out why. AI doesn't reconcile your accounts. It makes finding the exceptions faster, which is where most of the time goes anyway.

This is the deep dive on reconciliation referenced in the Complete Guide to AI for Financial Controllers.

What reconciliation time actually buys you

For a controller managing multiple entities with meaningful transaction volume, reconciliation typically breaks down like this:

The pattern here matters: AI compresses the finding step, not the resolving step. When the numbers match perfectly, AI hasn't done much because there was nothing to find. When there's a discrepancy, AI helps you isolate where the gap is and what type of mismatch it is — timing, amount, or missing item — before you spend investigation time on it.

The exception-flagging prompt

This works for bank reconciliations, intercompany matching, and any two-sided comparison — GL versus subledger, this period versus last period, books versus a third-party statement.

You are assisting with a reconciliation between two data sets.

Data Set A (e.g. GL/book balance):
[PASTE: Reference | Date | Description | Amount]

Data Set B (e.g. bank/subledger/counterparty):
[PASTE: Reference | Date | Description | Amount]

Compare the two sets and produce:
1. Matched items — count and total only, no need to list
2. Items in Set A with no corresponding match in Set B
3. Items in Set B with no corresponding match in Set A
4. Items that appear in both but with a different amount, with the variance shown
5. For each unmatched or mismatched item, categorize as: Timing difference (matches on amount/description but different date), Amount discrepancy, or Unexplained

Sort all flagged items by absolute dollar value, largest first. Do not attempt to resolve or explain the discrepancies — only identify and categorize them.

The instruction to categorize rather than explain is deliberate. AI can recognize that two amounts differ; it cannot know that the difference is a bank fee that hasn't posted yet or a foreign exchange remeasurement timing issue. That explanation is yours, and it's faster to produce once you know exactly which items need it.

Intercompany reconciliation — a specific case worth calling out

Intercompany reconciliation is where AI-assisted exception flagging earns its keep most clearly, because the comparison is inherently two-sided by construction — every intercompany transaction should appear as a mirror image on both entities' books. Feed AI both entities' intercompany ledgers and the same prompt above will surface the classic intercompany problems quickly: one side booked, the other not yet; FX translation differences between functional currencies; or a timing gap where one entity recorded in the current period and the counterparty recorded in the prior one.

For multi-entity environments running this every cycle, this is also one of the clearest places to extend the same logic into a consolidation-level check — see AI in Multi-Entity Consolidation for how exception flagging scales across more than two entities.

What not to do with AI in reconciliation

Don't let it clear or post anything. AI's output is a list of exceptions for your review, not an instruction to act on. Every resolution — writing off a small variance, posting a correcting entry, escalating a larger one — is a decision a qualified professional makes and documents.

Don't paste full transaction detail with identifying information if you can avoid it. Reference numbers, dates, and amounts are usually sufficient for exception detection. Counterparty names, account numbers, and descriptions containing personal or client-confidential information should be stripped or anonymized before they go into an AI tool, consistent with the data handling guardrails in the Month-End Close Playbook.

Don't skip the categorization step. A flat list of "things that don't match" is far less useful than the same list sorted by type and materiality. The categorization is what turns an hour of scanning into ten minutes of triage.

Building this into a repeatable habit

The first time you run this prompt, do it alongside a reconciliation you've already completed manually, so you can check AI's exception list against what you actually found. Most controllers find the lists match closely on bank reconciliations and need one or two rounds of calibration on intercompany, where description fields are less standardized.

Once calibrated, this becomes one more piece of the close where assembly is delegated and judgment stays put — the same shift that defines the AI-Augmented Financial Controller.