How to Use AI for Reconciliation and Spotting Anomalies

Hypotheses, never conclusions

There is one sentence that makes AI safe and useful in reconciliation work, and you should tape it to your monitor:

AI generates hypotheses. You generate conclusions.

Everything in this lesson follows from that. The tool is allowed to say "these two entries look like they might be duplicates". It is never allowed to say "these are duplicates" and it is certainly never allowed to say "the reconciliation is complete". The first is a lead worth two minutes of your time. The second is an assertion about your client's records that only you can make, because only you can be held to it.

Frame it that way and the tool becomes genuinely valuable. It is fast, tireless and pattern-hungry across material that makes human eyes slide off the page after forty minutes. It just has no idea whether it is right.

Four things worth asking it to look for

Possible duplicates. Not just identical rows, which your spreadsheet already finds. The interesting ones are the near-misses: same supplier, same amount, invoice numbers one apart, dates four days apart. That is either a genuine second invoice or the same one entered twice, and only a human looking at the document can say which.

Unusual amounts. Round numbers where round numbers are rare. Amounts that sit just under an approval threshold. A supplier who normally invoices in the hundreds suddenly invoicing in the thousands. None of these are wrong. All of them are worth a glance.

Timing oddities. Entries dated at the very end or very start of a period. Backdated postings. A supplier whose invoices arrive monthly and then arrive twice in one week. Gaps where a regular payment should be.

Description mismatches. The narrative says one thing and the coding says another. A description mentioning a deposit sitting in an expense code. A payment described as a refund with a positive sign.

Here is a list of transactions. I want you to flag things that
LOOK unusual. You are generating hypotheses for me to test.
You are not concluding anything and you must not say
that anything is confirmed, correct, or an error.
Flag and group under these headings:

Possible duplicates, including near-matches on supplier,
amount, reference or date. Say what makes each pair suspicious.
Unusual amounts, given the surrounding pattern.
Timing oddities: period boundaries, backdating, unexpected frequency.
Description and coding mismatches.
For every item, give me: the row reference, what caught your attention
in one sentence, and the specific check I should perform to confirm
or rule it out.
Rank by how much money is at stake, largest first.
Do not flag anything you cannot give me a check for.

That last line is the quality control. If it cannot tell you how to test a flag, the flag is noise.

The reconciliation itself is not the AI's job

Be clear about the split. The reconciliation is arithmetic and matching, and that belongs to your accounting system or your spreadsheet. What AI adds is help with the residue: the unmatched items, the difference you cannot explain, the six entries that refuse to tie out.

I have an unexplained difference on a reconciliation.
The amount is [state it]. The account is [describe it in general terms].
Do not perform any calculation and do not tell me the cause.
List the classic signatures a difference of this shape can have,
such as transpositions, sign reversals, doubled entries,
timing cut-offs and rounding accumulation.
For each signature, tell me the arithmetic test I should run myself
in the spreadsheet to see whether it applies.

Proving it yourself

A flag is worth nothing until you have gone to the source. That means the invoice, the statement, the contract, the remittance advice, the person who authorised it. Not a second opinion from the same tool, and not a longer explanation from it either, because a longer explanation is just more generated text about the same unverified guess.

Keep a short note of what you checked and what you found, including the flags that turned out to be nothing. Your working papers should show the flag, the check and the outcome. If somebody later asks why you looked at that entry, "the pattern matching flagged a near-duplicate and I traced it to the supplier's original invoice" is a professional answer. "The AI thought it was fine" is not an answer at all.

Take the flags you produced and turn them into a checklist for me.
One line per item, in this format:
row reference, what to look at, what would confirm it,
what would rule it out.
Add a column for me to write the outcome.
Do not add any new flags and do not remove any existing ones.
Do not suggest what the outcome will be.

The honest scorecard

Run this properly and you should expect most flags to be nothing. That is not failure, it is how screening works. A smoke alarm that only went off during real fires would be a miracle, not a smoke alarm.

What you are buying is coverage. The tool reads all four hundred rows at hour six of a long day with exactly the same attention it gave row one, which is more than any of us can honestly claim. You bring the judgment, the source documents and the signature. It brings the stamina.