7 Ways to Use AI When the Numbers Don't Match

Using AI when the numbers don't match is the unglamorous skill nobody put in a course. Two reports, two totals, and a quiet dread that one of them is going to a meeting. It happens to the bookkeeper whose bank feed and books are forty dollars apart, the shop owner whose ad platform claims twice the orders the store did, and anyone who has ever had two spreadsheets open side by side, scrolling.

The reason this deserves an edition of its own is that reconciliation has quietly become most people's real job. In Intuit's 2026 Future of Finance report, a survey of 2,000 finance leaders fielded in May 2026, 51% of the finance week goes to manual work like reconciliation, exports, error fixing, and report stitching, and 70% say their critical data sits scattered across systems with no single source of truth. That second number causes the first. Once your information lives in five places, somebody has to be the human join between them.

Most listicles answer this with "use AI to analyze your data," which is not an answer. Every idea below is about the specific shape of this problem: two sources that disagree, and the skill of finding exactly where and why. Each one ends with something you can start this afternoon.

1. Get Both Definitions in Writing Before You Look at a Single Number

The most common mismatch is not an error. It is two systems measuring different things and labeling the results with the same word. "Conversions" in one tool means a sale counted on the day someone clicked an ad. In another it means an event counted on the day it happened. Same word, different day, permanently different totals, and no amount of staring at the spreadsheet reveals it, because the explanation is not in the data. It is in the documentation.

A cutoff makes it worse. A sale captured at 11:40pm Sunday belongs to Sunday in one system and Monday in the other if their time zones differ by a few hours. The Google Ads Help page on data discrepancies lists five separate mechanisms like this, including reporting date, counting method, and how long a click stays eligible. The gap you are chasing may be fully documented and completely expected.

Start today: find each system's own help page for the metric in question, paste both passages into a chat window, and ask: "List every way these two definitions differ, and for each one tell me which direction it pushes the number." You get a short table, and that table usually names your gap. Reading vendor documentation needs no AI at all. The translation of it into a plain comparison is where the model earns its place.

2. Split the Gap Into Three Buckets: Only Here, Only There, Both But Different

Priya keeps the books for a two-van plumbing company. Her bank feed and her accounting software were $212 apart, and she spent two evenings comparing running totals, which is the one approach guaranteed not to work. A total is a single number with every cause folded into it. You cannot see inside it.

Rows are visible. Ask for a classification instead of a comparison: which records appear only in source A, which only in source B, and which in both but with different values. Those three buckets have three unrelated causes. Only in A usually means something was never imported. Only in B usually means a duplicate or a manual entry. In both but different is the interesting pile, because that is where a real discrepancy lives rather than a missing row. Priya's $212 was two refunds sitting in the bucket that existed only in the bank, which is a thirty second conversation once you can see them.

The catch is matching. If the two lists share no identifier, say so up front and name what to match on, usually a combination like date plus amount plus a name fragment, then make it show the rule it used.

Start today: export both sources as CSV, paste them into Claude or ChatGPT, and ask: "Match these on invoice number. Return three lists: only in file one, only in file two, and in both but with a different amount. State the matching rule you used and how many rows you read from each file."

3. Measure the Shape of the Gap, Not Its Size

A $4,000 gap made of one $4,000 transaction and a $4,000 gap made of nine hundred differences of four dollars each are not the same kind of problem. The first is a mistake: something got entered twice, or dropped. The second is a rule: a fee, a currency conversion, a rounding behavior, a tax treatment applied on one side and not the other.

Almost nobody asks this, because the total is what got flagged and the total is what people chase. But the distribution tells you who should fix it and how long it will take. One big difference is an afternoon of detective work on a single record. Nine hundred small ones is a configuration setting, and once you find it all nine hundred resolve at once.

Start today: with both files already pasted in, ask: "How many records differ, what is the smallest, largest and median difference, and show me the five largest. Is the gap driven by a few large differences or many small ones?" If the answer is many small ones at a similar percentage, you are looking at a fee or a rate, and the fix is a setting rather than a correction.

4. Put the Reconciliation on a Standing Diff That Only Reports What's New

Everything above is worth doing once. The problem is that it is never once. The feeds keep arriving, and next month you start from a blank page and rediscover the same six known differences before you reach anything new. That rediscovery is the real cost, and it is the part you can hand off.

The move is to keep the result. A standing reconciliation remembers last run's divergences and reports only what changed: new mismatches, mismatches that resolved themselves, and mismatches that got bigger. The known differences stop appearing. What lands in your inbox is three lines instead of a spreadsheet, which means you read it. The Loop walks the full version through in reconciling invoices against a spreadsheet, and the general shape runs as an Excel and spreadsheet workflow that pulls both exports, matches them, and writes only the differences.

Start today: do the match once by hand, then save that output as a dated file and call it the baseline. Next time, paste the old reconciliation alongside the new one and ask only for the differences between the two. Comparing this month's mismatch list against last month's is the whole mechanic, and once it works by hand it is a workflow you can describe in a sentence.

5. Name the One Source That Gets to Be Right

Two sources disagreeing is a tractable problem. Two sources disagreeing with no agreed winner is an argument, and arguments have no end state. Before you look at a single row, decide which source is allowed to be right, and pick the one with the least room to be wrong. Money that left a bank is harder to dispute than money a dashboard thinks left. A settlement file from a payment processor, which Stripe's documentation describes as working like a bank statement, is a stronger anchor than an analytics report that models some of its own numbers. A physical count beats a system quantity every time, and it is the one people trust least because counting feels unsophisticated.

With an anchor, the question flips into a better one. Instead of "why don't these match," you ask "what would have to be true in the second system for it to agree with the first." That produces candidate explanations you can test, each one a thing that either happened or did not. It also fixes the write direction for anything you automate later: corrections flow from the anchor outward, never both ways, because two systems correcting each other is a loop nobody can unwind.

Start today: write one sentence naming your anchor and why, then paste it above your data every time. "The bank export is the source of record because the money moved; the invoicing tool is a record of intent." Then ask what would have to be true in the second source to match it. Ranking two sources by how wrong each can be is a cousin of deciding what to fix first.

6. Find the Third Artifact That Breaks the Tie

Marcus runs receiving at a distributor. A supplier's packing slip said 48 cases, his inventory system said 44, and the two sides spent a week trading emails about whose number was wrong. The answer was in neither document. It was in the gate log, which recorded a second truck the next morning.

When two sources disagree there is nearly always a third artifact that settles it, and it is usually something nobody thinks of as data: a shipping manifest, a door badge log, an email timestamp, a photo of a shelf, the till roll, a help desk ticket. The useful move is not analyzing the two sources harder. It is asking, before you dig in, which third record would make this decidable.

The same logic applies to software. If your store's finance reports and your ad platform disagree, the payout file is the third source, and the Shopify Help Center documents what each finance report counts, which is often the detail that decides it.

Start today: describe the two disagreeing sources in three sentences each, then ask: "What third record, log, or document would prove which of these is correct? Rank the options by how easy each is to get." Then go get the top one. Half the time you already have it.

7. Write Down the Gap You're Allowed to Have

Here is the part that saves the most hours and is the least satisfying to hear: some gaps are correct and permanent. Two systems can work exactly as designed and never agree, because of a processing fee, a timing cutoff, a rounding rule, or a definition that was never going to line up. Chasing those to zero is a treadmill, and people stay on it for years because nobody gave them permission to stop.

So give yourself permission, in writing. A tolerance is two sentences: the gap you expect and why, and the threshold past which it becomes real. "These two will always differ by roughly 2.9% plus thirty cents per transaction, because that is the processing fee the second system does not see. Anything beyond $50 or 4% is worth my time." Next month's reconciliation then takes three minutes, because you are comparing against a documented expectation rather than investigating.

Start today: take the mismatch you already live with, paste both sources in, and ask: "What is the structural reason these two will never match exactly, and what formula describes the expected difference?" Write the two sentences and save them next to the report. From there this part needs no AI at all. An expected gap plus a threshold is a subtraction and an alert, which any spreadsheet or reporting workflow handles with no model involved.

Pick One

Seven is too many for one week. Pick the one that made you wince, probably the mismatch you have been living with for months and quietly rounding off in your head. If you only do one thing, do number seven: writing down the gap you are allowed to have retires a recurring chore permanently and takes about ten minutes.

Sunday we report what changed in AI, Wednesday The Loop builds one workflow all the way through, and Monday this column hands you the ideas in between. If your mismatch turns out to be monthly rather than one-off, that is a workflow, and the agent workflow directory has the shape of it.

Frequently Asked Questions

• Why do two systems report different totals when neither one is broken?+
  Most of the time because they are not measuring the same thing. The two systems can disagree on what counts as one event, on which day an event belongs to, on the time zone the day starts in, and on how long after a click or a visit something still counts. None of that is a bug. It means the two numbers were never supposed to be identical, and your job is to work out the size of the gap you should expect rather than to drive it to zero.
• Can I just paste both spreadsheets into a chat window?+
  For a few hundred rows, yes, and it works well. Three cautions. Check your own confidentiality rules before pasting customer or payroll data anywhere. Long files get truncated quietly, so ask for the row count back and compare it to what you sent. And always make it state the rule it used to match records, because a wrong matching rule produces a confident, tidy, completely wrong answer. Past a few thousand rows, move the job out of the chat window and run it as an automated spreadsheet workflow instead.
• What if the gap is small? Is it still worth chasing?+
  Small and stable is usually a documented tolerance, not a problem. Small and growing is a problem wearing a disguise. The useful move is to write down the expected gap and the threshold that makes it worth your attention, then only look when the threshold is crossed. Deciding what deserves the next hour is its own skill, and the same ranking logic applies here as anywhere else on your list.
• Do I need AI for this at all?+
  Often no, and it is worth saying so. If both lists share one clean shared identifier, a lookup formula in a spreadsheet finds the mismatched rows in five minutes and costs nothing. AI earns its place in three situations: when the identifier is messy and the match has to be fuzzy, when the valuable output is a plain-language explanation of why two numbers differ rather than the list of differences, and when the job repeats often enough that you want it running without you.