AI Variance Analysis and Management Commentary

By Jane Ha · Updated Oct 11, 2026

AI variance analysis is the use of software to compare actual results with budget or forecast, break each gap into its drivers (price, volume, mix and timing), flag the lines that deserve attention, and draft the written explanation a finance leader sends to management. The arithmetic should come from your ledger and a deterministic calculation. The language model's job is to organize the evidence and write the first draft. A named analyst owns the cause, the numbers and the final wording.

This guide is for FP&A leads, controllers and CFOs who spend the first week of every month turning a variance report into paragraphs. It sits under the planning hub, AI in FP&A, and focuses on the one task where AI saves the most writing time and creates the most risk: explaining why the numbers moved.

Key Facts

• The SEC's MD&A guidance says the discussion should not be merely a restatement of financial statement information in narrative form. A variance narrative that only repeats the table fails the same test.
• US reporting rules ask registrants to describe the extent to which material revenue changes are attributable to changes in prices, volume or new products, which is the price/volume/mix split in regulatory form.
• In the FinanceBench test of questions about public companies, GPT-4 Turbo with a retrieval system answered incorrectly or refused 81% of a 150-case sample. That was a 2023 test of specific models, and it is the reason every figure in an AI draft needs a source cell.
• A variance explanation has three parts: the size of the gap, the decomposition into drivers, and the business cause. AI is reliable on the first two only when a calculation engine produces them. The third needs a human with operating knowledge.

What Variance Analysis Actually Asks

A budget-versus-actual report tells you that revenue is 7% under plan. It doesn't tell you whether you sold fewer units, charged lower prices, or sold more of the cheap product. Variance analysis answers that. Then management commentary translates the answer into plain language for people who won't open the spreadsheet.

Done manually, this is slow for a predictable reason. A mid-sized company has hundreds of lines by hundreds of cost centers. The analyst filters to the biggest gaps, pulls the transaction detail, asks the business owners what happened, and writes it up. Most of the time goes into pulling and reconciling data and chasing explanations, not into thinking.

AI helps in three places:

1. Triage. Rank and flag the lines that moved the most or moved unusually, so the analyst starts with a short list.
2. Decomposition. Split each flagged variance into price, volume, mix and timing components from the underlying data.
3. Drafting. Turn the decomposition and the owner's notes into a paragraph in your house style.

It doesn't help with the question underneath: why did the customer buy less? That answer lives in a sales manager's head, a ticket, or an email. Treat it as an input to the model, never an output.

Price, Volume and Mix: The Decomposition

Revenue variance splits into three effects. The price effect is the change in realized price on the units actually sold. The volume effect is the change in units at the budgeted average price. The mix effect is the change in average price caused by selling a different blend of products.

Here is a worked example, labelled illustrative. A company budgets 10,000 units at a blended budget price of $100, so budget revenue is $1,000,000. It sells 9,600 units. Product A sells 6,400 units at $80 and Product B sells 3,200 units at $130 (budget mix was 6,000 A and 4,000 B at the same prices), so actual revenue is $928,000.

The story changes with the decomposition. Without it, "revenue is down $72,000 on soft demand" is the easy line. With it, more than 40% of the gap comes from selling less of the higher-priced Product B, which sends the question to the sales team and not to marketing volume.

The same logic applies to cost: rate variances (paid more per hour or per kilogram), efficiency or usage variances (used more hours or material), and volume variances against a flexed budget. If your reporting uses a different decomposition, such as customer-cohort or channel, fit the model to that. The point is a fixed method that always sums back to the total.

Materiality Thresholds and Anomaly Flagging

Nobody wants commentary on every line. Most teams set thresholds, such as flagging a variance only when it exceeds both a percentage of budget and an absolute amount. The two-part test matters. A 40% overrun on a $2,000 line is noise. A 3% miss on a $20 million line is not.

AI adds a second filter on top of the fixed thresholds. Instead of asking "is this over 10%?", an anomaly model asks "is this unusual for this line?" It compares the variance to the line's own history and seasonality, so a December freight spike that happens every December doesn't consume analyst time, while a quiet line that suddenly moves does. Anomaly detection and time series analysis are the underlying techniques.

Use both. Fixed thresholds are predictable and auditable. The model catches what thresholds miss: small, persistent drift and offsetting variances that net to nothing at the total but hide two real problems underneath. Write the rule down, and review it each quarter so it doesn't drift with the business.

What AI Does Well, and Where Narratives Go Wrong

The failure patterns are consistent:

Hallucinated causes. Given a variance and no explanation, a language model will supply a plausible one: "driven by supply chain delays" or "softer consumer demand." It reads well and may be completely wrong. This is the AI hallucination problem showing up in a finance report, and it is the most dangerous because nothing in the text signals it.

Numbers that don't tie. Models are built to generate likely text, not to add. A draft may quote a variance of $1.3 million where the ledger says $1.03 million, or let percentages and dollar figures drift apart. Public evidence on this is thin and dated, but the 81% incorrect-or-refused figure from FinanceBench shows how often early systems got questions about real financial filings wrong. Newer models do better. Don't rely on that.

Right numbers, wrong story. A draft can cite correct figures and still pin the variance on the wrong driver, usually because two things moved together. Price rose and volume fell, and the model credits one for the other.

Omitted offsets. Commentary that mentions the big favorable item and skips the equal unfavorable one gives management a false picture of the quarter.

Tone drift. Models smooth language. "Material weakness in forecasting accuracy" can quietly become "minor timing differences." Wording choices in commentary are judgments.

How to Implement It, Step by Step

1. Lock the data source. Pull actuals and budget from the ledger or the planning system, not from an exported sheet someone edited. Check account mapping first; the issues in finance data readiness for AI show up here as wrong totals.
2. Run the arithmetic outside the language model. Compute variances and the price, volume and mix split with code, SQL or the planning tool. The model receives the results as a table. This is the single most effective control.
3. Set materiality rules. Define the percent and dollar thresholds per statement area, plus a "never skip" list such as revenue, headcount cost and anything covenant-related.
4. Add the anomaly layer. Let a model rank lines by how unusual the movement is, and show the analyst the reason for each flag.
5. Collect cause notes from owners. Send each flagged line to its budget owner with the decomposition and a short form: what happened, one-time or recurring, and expected next quarter. Nothing goes into commentary as a cause without this input or a document reference.
6. Draft with constraints. Instruct the model to use only the supplied table and notes, to cite the source cell or note for every figure, and to write "cause not provided" where none exists. Retrieval from your own records (retrieval-augmented generation) helps, but the draft still needs checking.
7. Run the tie-out. A script, not a person, compares every number in the draft with the source table and fails the draft on any mismatch or any figure with no source.
8. Review as a human-in-the-loop. The analyst reads for cause and tone, the controller or FP&A lead signs off. The human-in-the-loop step has to be a real review, with the right to rewrite or delete.
9. Keep the trail. Store the input table, the prompt, the draft and the approved version. When the board asks why you said something in March, you need the record.

A Review Checklist for Every Draft

• Does each number match the ledger or planning tool, to the dollar?
• Do the price, volume and mix effects sum to the total variance?
• Is every stated cause backed by an owner's note or a source document?
• Are offsetting items disclosed, not just the headline?
• Is the year-over-year or budget comparison the one the reader expects?
• Are one-time items labelled one-time, and recurring ones recurring?
• Would the owner of this line recognize the explanation as true?
• Is any forward-looking statement authored and approved by a person?

Using the Same Method for Forecast and Close Work

The workflow carries over. Forecast-to-actual commentary uses the same decomposition against a different baseline, and a good forecast process keeps the variance history as input for the next forecast. In the close, preliminary flux explanations on balance-sheet accounts follow the same pattern: a computed change, an owner's note, and a drafted sentence. See AI month-end close for how that fits the calendar. For drafting rules more broadly, read generative AI in accounting. If you want an agent to run budget-to-actual checks continuously, an AI budgeting agent is the pattern to start from.

Pitfalls to Avoid

• Letting the model do the math. Deterministic calculation first, language second.
• Accepting a fluent paragraph as an explanation. Fluency isn't evidence. Require a source for every cause.
• Commenting on everything. Ten sentences on a $5,000 variance bury the one that matters.
• Copying last month's narrative. Models and people both recycle phrases. Check that the cause still applies.
• Skipping the owner conversation. The tool saves drafting time. It can't replace asking the person who ran the campaign or signed the contract.
• Using unapproved tools with sensitive data. Pasting a trial balance into a public chatbot is a data leak. Use an approved environment.
• No regression test. After a model or prompt change, rerun last quarter's pack and compare. Treat prompts like code.

Where to Go Next

Pick one report, usually the monthly revenue or opex pack, and run the process in parallel for a cycle. Keep the manual commentary as the official one, produce an AI-assisted draft beside it, and compare tie-out failures, invented causes and time saved. Then expand to the next area. Start from the AI in FP&A hub if you're mapping the whole planning workflow.