AI management reporting and variance commentary

What AI management reporting and variance commentary means

AI management reporting and variance commentary means using a language model to draft the narrative that explains a period's numbers against budget or a prior period, such as why gross margin moved or which cost line drove an overspend, from figures a person has already checked. It matters because a model can write a first draft in minutes from data that would take an analyst hours to describe in words, but the model does not know which explanation is actually true; the controller who signs the pack does.

This guide covers what belongs in a management pack, why a documented set of KPI definitions has to exist before AI commentary is worth drafting, how to structure a request so the model explains variances instead of inventing causes, a worked example of a drafted comment a controller corrects, and where this kind of drafting typically goes wrong.

What a management reporting pack contains

A monthly management pack usually pairs the numbers with the story behind them: a profit and loss compared with budget and the prior period, a small set of KPIs tracked over time, a cash position, and a page or two of commentary that explains the variances a reader would otherwise have to guess at. The commentary is what turns a spreadsheet into something a non-finance manager or a board member can act on, which is exactly why a wrong or vague comment does more damage than a wrong number: a wrong number tends to get checked, a plausible-sounding wrong explanation often does not.

• Profit and loss versus budget and prior period, by month and year to date
• A short KPI set: gross margin percentage, days sales outstanding, operating expense ratio, and whatever else the business actually manages by
• Cash position and a rolling forecast
• Variance commentary: what moved, by how much, and why

KPI definitions have to exist before the AI does

A model asked to comment on gross margin will use whatever definition it infers from the data it is given, and different systems compute gross margin differently: some net off freight, some include an overhead allocation, some do not. If the finance team has not written down its own definition of every KPI in the pack, a language model cannot reliably match it, and a comment that describes the wrong measure is worse than no comment, because it reads as authoritative.

The fix is a short glossary, not a long policy: for every KPI in the pack, the exact formula, the source of each figure, and the rounding or currency convention, written once and given to the model with every request. The same glossary keeps a human analyst consistent from one month to the next, which is a reason to write it even without AI in the picture. Review the glossary itself whenever a KPI's formula changes, for example when an overhead allocation is added to cost of sales, and version it so a reader can tell which definition applied to which month's pack.

Structuring the request so the model explains, not invents

A useful variance commentary request gives the model the numbers, the KPI glossary, and one rule: every explanation must be traceable to a transaction, an account or a known event, such as a price increase, a new contract or a one-off cost, never a guess dressed up as a fact. Asking simply to explain the variance invites the model to produce something that sounds plausible; asking it to explain using only the transaction detail supplied, and to say so when the data does not show a cause, produces something a reviewer can actually check.

• Give the model the variance figures and the transaction detail behind the largest ones, not just the summary totals
• Give it the KPI glossary so a term such as gross margin means the same thing every month
• Require every claimed cause to point to a specific account, transaction or known event
• Require the model to flag a variance it cannot explain from the data supplied, rather than guessing

Worked example: a budget-versus-actual comment a controller corrects

A retailer's management pack for the month shows revenue of EUR 640,000 against a budget of EUR 600,000, a favourable variance of EUR 40,000, or 6.7% (40,000 divided by 600,000). Cost of sales is EUR 416,000 against a budgeted EUR 372,000, an unfavourable variance of EUR 44,000, so gross profit is 640,000 minus 416,000, EUR 224,000, against a budgeted 600,000 minus 372,000, EUR 228,000, a EUR 4,000 shortfall despite the higher revenue.

The AI drafts: revenue grew 6.7% on strong demand, but gross margin fell because cost of sales rose faster than revenue, driven by higher input costs. The controller checks the transaction detail behind the EUR 44,000 cost of sales variance and finds that EUR 30,000 of it is a one-off write-down of slow-moving stock, unrelated to the month's trading, and only EUR 14,000 relates to higher purchase prices on goods actually sold in the period.

The controller rejects higher input costs as the sole cause and rewrites the line: gross margin fell EUR 4,000 despite 6.7% revenue growth, because a EUR 30,000 stock write-down and EUR 14,000 of higher purchase prices, 30,000 plus 14,000 equals the 44,000 cost of sales variance, more than offset the extra margin from higher sales. The rewritten comment separates a one-off from a trend, which changes what a reader should do about it; the AI's draft did not, because the totals it was given did not distinguish the two without the transaction detail.

What the reviewer checks before the pack goes out

A drafted comment is not ready for a board pack until someone has tested it against the same discipline as any other output a person is accountable for.

• Trace every named cause to a transaction, account or event; reject anything that reads as inference
• Check the sign: a comment calling a variance favourable when the figures show unfavourable is a common and embarrassing error to miss
• Separate one-off items from trend items, because they call for different management action
• Confirm the KPI in the comment was computed using the glossary definition, not a different one the model assumed
• Read it as the intended audience would: a comment correct in every fact can still mislead by emphasis

Where AI reporting commentary typically fails

The most common failure is not a wrong number but a wrong story: a model given only the current and prior period totals will confidently attribute a swing to whatever cause is most commonly associated with that KPI in general, rather than the one that actually happened this month, because totals alone give it no way to tell the difference.

• Mixing a one-off item into a trend explanation, or the reverse
• Commenting on a variance below the pack's own materiality threshold, adding noise instead of signal
• Currency or consolidation mismatches: comparing a branch's local-currency actuals with a group-currency budget without saying so
• A KPI computed differently this month than last, breaking the trend line the commentary describes
• A prior-period comparison silently drawn from a since-restated figure, so the stated variance no longer matches either version of the number

Governance and data protection for management reporting

A management pack is commercially sensitive before it is public, so treat an AI drafting tool with the same access discipline as any other system that touches it: read access to the figures and transaction detail it needs, never write access to the ledger or the ability to publish the pack, and a log of what was sent and what was drafted. Unreleased figures are also inside information for a listed group, so the same insider-trading and disclosure rules that govern who sees a board pack before it is released apply to whoever can request a draft of it.

• As of September 2026, Anthropic states that by default it does not use inputs or outputs from its commercial products, including the API, to train its models, unless the customer chooses to share them
• Restrict who can request a draft covering unreleased figures, the same way access to the figures themselves is restricted
• Keep the KPI glossary and the drafted commentary under version control so a reviewer can see what changed and why
• Never let the pack leave the business, drafted or not, without the sign-off management accounts already require

Why the analyst's role shifts rather than disappears

Drafting a first pass of routine commentary is the part of the job that benefits most from a model, because the pattern of a management pack repeats every month and most of the words describe the same handful of KPIs in a similar structure. Judging whether a drafted cause is real, whether it is worth mentioning at all, and what a reader should conclude from it is the part that does not repeat in a way a model can learn reliably, because it depends on knowing the business, this month, not last month's phrasing.

That shifts the analyst's time from writing every line from a blank page toward reviewing, correcting and deciding what belongs in the pack, which is a different skill, not a smaller one; the worked example above shows a controller catching an error a model had no way to catch on its own, from data it was never given.

Doing this in Skyline Nexus ERP

Skyline Nexus ERP already produces the numbers a variance commentary has to explain. The Fiscal Authority Profit & Loss report compares a period with no comparison, the previous period or the previous year, filtered by project or branch, and Budget vs Actual reports total budget, actual spend and variance by fiscal year, account and cost centre, with a status of over budget, warning or on track and a trend chart. Cost Center Analysis breaks revenue and expense down the same way for departments that need it.

Drafting the narrative commentary itself with AI, from those figures and a business's own KPI glossary, is on the Skyline Nexus ERP roadmap and being rolled out. Today, the in-app assistant can already answer how-do-I questions about running the Budget vs Actual and Profit & Loss reports, citing the help page it used, and the reports themselves give a reviewer, or an external drafting tool, the same reconciled numbers to start from.