September 9, 2026
In Excel, an accountant can keep digging until a number makes sense. Agentic analysis moves more of that work into queries, joins, code, and reasoning. A $3,813.58 disagreement showed us why we need way to inspect easily and quickly.
Isotopes AI7 min read
With Excel, everyone checks the formula first and the numbers second.
A senior accountant can get a spreadsheet, check the formula, follow the pivots and explain the number.
With aidnn, an analysis can involve SQL queries, joins, Python, business rules, and AI reasoning before a number comes back. That can shorten the time to an answer, reduce analyst effort, and make the work practical to repeat across larger datasets. Accuracy still depends on the data and the business meaning behind the request, and the reviewer should not have to read Python to understand how the answer was reached.
Thursday morning, revenue accounting
An accountant was checking our output against a workpaper, line by line. One figure would not tie.
Transaction fees and processing fees matched to the second decimal place. The discrepancy column did not.
The $3,813.58 difference was large enough to stop the review. The surrounding figures tied, so the accountant needed to understand this one before signing off.
This was the kind of question they could usually work through in a spreadsheet by filtering rows or adding another pivot. Here, the business interpretation behind the number was not available in the same way.
Both totals came from the same 3,144 rows. We went back to the request:
Add a discrepancy column where the processor fee doesn't match the calculated processing fee, and what the difference is.
The ambiguity was in the word discrepancy. When the difference was negative, should it offset a positive amount, as an adjustment would, or should both directions count toward the total size of the error?
The same 3,144 rows produce a different total when discrepancy is interpreted differently.
The accountant's workpaper used the first meaning, and aidnn used the second. Both were reasonable readings of the request, which never said how negative differences should be treated.
An engineer opened the session's debug view and found the interpretation: the analysis had treated discrepancies as errors and summed absolute values. Once that was visible, the $3,813.58 difference made sense.
We found the reason in about twenty minutes. The harder part was getting everyone back together, comparing the result with the Excel workpaper, and explaining why the two methods produced different numbers.
The answer was already in the analysis. The reviewer just did not have an easy way to see it while checking the number. The accountant could not answer the question from the review itself, so engineers on both sides had to reconstruct the business meaning behind the number.
That experience changed what we expect from verification. During review, the person checking the number should be able to see which rows were used, what changed along the way, which business definition shaped the result, and where a correction would change the answer.
A green check helps, but the reviewer still needs the evidence trail behind it.
Re-running the analysis would have produced $25,888.32 again using the same interpretation of discrepancy. The accountant would still have the same question: should those differences be treated as errors or adjustments?
Verification needs another source of evidence when the business meaning is in question. That may be a source system, a documented rule, a memory supplied by the user, learned context, or the person who owns the definition.
That context also has to be preserved while the analysis is being built. Once the work is reduced to a final number or chart, many of the decisions behind it are harder to recover.
The reconciliation is one example. The same review problem comes up in ARR, MRR, forecast changes, margin analysis, spend reviews, and other finance work where several sources, joins, business rules, and calculations sit behind the number on the screen.
aidnn can do that work directly against the data. As the analysis runs, it keeps an evidence trail of the plan, the purpose of each step, the code that ran, intermediate tables, source references, memories and learned context, expert guidance, and corrections made along the way.
In a workbook, a reviewer can move through sheets, formulas, and pivots to understand how a number was built. In aidnn, that evidence is preserved so the same kind of review is possible without asking someone to read Python or reconstruct the analysis.
The example below uses ARR. It builds a dashboard from a warehouse across sixteen fiscal quarters and seven product lines, with account-level drill-down. The capture is read-only, but you can scroll through the analysis, open steps, expand processing details, and inspect available sources.
Verification uses that evidence trail to review the analysis alongside the result. It can show when the calculation checks out but the business meaning still needs a decision from the person who owns it.
If the evidence is weak, verification can mark Needs human review and point to what needs attention. More detail remains available when someone wants to inspect further.
Back to the accountant. If the analysis had shown that discrepancies were treated as errors and summed by absolute value, the disagreement could have been resolved during review. The accountant could say, Net them. These are adjustments. The definition could be corrected while the work was still in front of everyone.
The calculation followed a reasonable interpretation of discrepancy. The accountant was using a different one.
Verification should put that interpretation in front of the person who owns the definition.
Once the definition is corrected, that correction should not disappear with the session. It becomes context aidnn can use the next time the same question comes up. NeoCortex carries that learning forward, helping aidnn improve with each interaction while keeping the reasoning and source of that learning available for review.
The reviewer can inspect the sources, assumptions, memories, learned context, and checks behind the result. When a number is questioned, they can trace it without rebuilding the analysis from scratch.