Worked case: monthly finance review

The controller defines:

• Purpose: prepare an internal draft of questions and verified variance commentary for April.
• Sources: approved April management report, general-ledger summary, budget version B3, cash reconciliation, and migration cost-center detail.
• Scope: Northstar Services consolidated management view, April, USD, displayed in thousands.
• AI role: map sources, extract specified values, organize questions, and draft from verified findings.
• Prohibited role: no journal entries, payment actions, accounting-policy conclusions, forecasts, external communication, or recommendations.
• Owners: analyst prepares; controller reviews calculations and explanations; business owners confirm causes.

The source files stay in an approved environment. Access and retention follow company policy.

The analyst gives the model approved excerpts with headers and asks:

Create a source map. For each document, return title, version, owner,
period, entity scope, currency, units, actual/budget/forecast status, and
included sections. Mark missing or conflicting metadata. Do not extract
performance conclusions.

The model flags that the migration detail says “April preliminary” while the management report says “April closed.” This is a useful finding, but not proof of a mismatch. The analyst asks the controller and learns that the cost-center detail was refreshed before close and is stale. The approved final detail is retrieved.

Without this orientation pass, the review might have mixed versions.

The analyst extracts only named rows with source locations. Then a spreadsheet recomputes:

• revenue versus budget: $2.40m − $2.50m = −$0.10m;
• revenue budget variance: −$0.10m ÷ $2.50m = −4.0%;
• revenue versus March: $2.40m − $2.20m = +$0.20m;
• sequential revenue change: $0.20m ÷ $2.20m ≈ 9.1%;
• operating expense versus budget: $1.32m − $1.20m = +$0.12m;
• operating expense budget variance: $0.12m ÷ $1.20m = +10.0%;
• cash movement: $0.61m − $0.90m = −$0.29m.

The spreadsheet stores exact source cells and formulas. Whether an expense variance is labeled favorable or unfavorable follows the approved reporting convention; the analyst does not ask AI to infer the sign convention.

These calculations show changes, not causes.

The analyst supplies verified calculations and asks the model to generate questions, not explanations:

Using only the verified variance table, list investigation questions.
Separate revenue, operating expense, and cash. For each question, name the
source or owner likely to answer it. Do not infer causes or call a movement
good, bad, material, expected, or unusual.

Useful questions include:

• Did revenue volume, pricing, mix, timing, or classification differ from budget?
• Is the March comparison affected by working days or seasonality?
• How much of the expense variance belongs to migration contractors?
• Were any migration costs budgeted elsewhere?
• Which operating, investing, or financing cash movements explain the $290,000 decrease?
• Did receivables, payables, or deferred revenue change materially under the team’s defined threshold?

The questions direct investigation without converting correlation into certainty.

The final cost-center detail shows migration contractor expense of $150,000 versus a budget of $50,000, a $100,000 variance. The migration owner confirms the work was accelerated into April and provides the approved project schedule. The remaining $20,000 operating-expense variance is spread across several categories and remains unexplained pending review.

The cash reconciliation shows:

• positive operating cash movement before working capital: $180,000;
• receivables increase: negative $260,000;
• annual insurance payment: negative $90,000;
• equipment purchase: negative $120,000;
• other net movements: zero.

These sum to negative $290,000. The analyst independently verifies classification and reconciliation in the approved workbook. AI does not decide whether the insurance payment should be normalized or whether the equipment purchase was appropriate.

Evidence is labeled:

• calculated: revenue and expense variances;
• reported and supported: accelerated migration work;
• calculated from reconciliation: cash bridge;
• open: cause of the $100,000 revenue shortfall to budget;
• open: explanation for remaining $20,000 expense variance.

The analyst prompts:

Draft internal review commentary from the verified findings below.
Preserve currency, units, periods, signs, and labels. Separate verified
observations, supported explanations, and open questions. Do not introduce
recommendations, forecasts, materiality judgments, or causes not supplied.
End with an OPEN ITEMS list.

A suitable draft says:

“April revenue was $2.40 million, $100,000 or 4.0% below budget, and $200,000 or approximately 9.1% above March. The cause of the budget variance remains under review. Operating expense was $1.32 million, $120,000 or 10.0% above budget. Verified migration contractor costs account for $100,000 of that variance; the migration owner reports that scheduled work was accelerated into April. The remaining $20,000 requires review. Cash decreased $290,000 to $610,000. The verified bridge includes operating cash before working capital of $180,000, a $260,000 receivables increase, a $90,000 annual insurance payment, and a $120,000 equipment purchase.”

The commentary avoids saying cash performance “weakened” or the migration “caused overspending.” Those interpretations require criteria and decisions beyond the evidence.

The analyst creates a claim ledger. The controller checks source versions, formulas, signs, labels, the cash bridge, and open items. The migration owner confirms the narrow factual explanation. Because revenue causation remains unresolved, the draft says so.

The controller approves the exact artifact for the internal review meeting. Any number or explanation change invalidates approval. The AI has no send permission and no access to accounting or payment actions.