Take a board pack due out on Thursday. On Wednesday evening the CFO reads the variance commentary and stops at a line saying marketing spend was over budget because of a trade show. The figure is right. The trade show was last quarter, and the overspend came from an agency retainer that started early. The wrong sentence reads exactly as confidently as the correct ones around it.
That is the central risk in AI variance analysis. The arithmetic is rarely the problem. The explanation is, and a board that finds one invented cause will start to doubt the rest of the pack.
What AI variance analysis does reliably
Most of the work in variance commentary is mechanical, and software does it well. It calculates every movement against budget, forecast and prior year. It ranks them against the thresholds you set. It pulls the journal lines, vendors and customers behind each material movement, and it notices patterns a tired analyst skips, such as a cost line that has crept up for five months without crossing the threshold in any single one.
Finance teams have been moving this way for a while. Gartner's release of 18 November 2025, based on a May–June 2025 poll of 183 CFOs and senior finance leaders, listed error and anomaly detection (34%) among the leading finance AI use cases, after knowledge management (49%) and AP process automation (37%). It also found "91% of respondents reporting low or moderate impact initially", which argues for putting the controls in place before the volume.
Figures from the ledger, words from the model
The first rule is architectural. The model should never produce a number. Figures come from the ledger or the planning system by query, and the model receives them and writes about them. Every amount and percentage in the commentary should trace back to a source line someone can open.
This also makes checking practical. If the draft says revenue was 4% below budget, an automated tie-out can compare that figure with the variance table before a person reads a word. Any number in the text that does not match the table is a defect, and the draft goes back. Our earlier piece on the month-end close covers the same principle for close commentary.
Checking the explanations
Errors in AI-drafted commentary tend to fall into a few types. The model infers a cause from an account name. It presents a timing difference as performance. It describes one large movement and misses the offsetting one that makes the net figure small. Or it carries over last month's explanation because the pattern looks the same.
A simple discipline is to label each explanation by the evidence behind it.
| Explanation | Evidence | Who confirms | Fit for the board? |
|---|---|---|---|
| Timing, such as an invoice booked late or an accrual reversing | Journal entries and their dates | Accountant | Yes, once confirmed |
| Volume, price or mix | Sales and operational data | FP&A with the business owner | Yes, with the data shown |
| One-off item | Contract, invoice or approval | Budget owner | Yes, once confirmed |
| Cause inferred from the account name or past months | None beyond the pattern | Nobody yet | No: confirm it or remove it |
The last row is the one to watch. An unconfirmed explanation is a question for the budget owner, not a sentence for the board. For how to turn confirmed movements into a story an audience can act on, see From Data to Decisions: Finding the Story in the Numbers.
Materiality thresholds, set in advance
AI will write about whatever it is pointed at. If every movement gets a paragraph, the board reads none of them. Agree thresholds before the month starts, combining a percentage and an amount so that small lines do not dominate, and set them per section where the business needs it. Document them, and let the audit committee or board see what they are.
Below the threshold, use AI for what it does well: trends. Three small adverse movements in the same direction can matter more than one large favorable one, and a person scanning a long variance report may not spot them.
A reviewer, by name
Assurance is where many teams are behind. When KPMG surveyed 1,013 senior finance leaders in March 2026, only 42 percent of organizations were "fully assurance-ready" when it came to AI-enabled finance processes.
Commentary that goes to a board needs one person who has reviewed it and answers for it. That person checks that the figures tie, that every explanation has its evidence, and that the overall message matches what management believes. Record the sign-off with the date and the version reviewed, so the trail survives a question from the audit committee months later.
What AI variance analysis will not tell you
No variance report says what to do. It explains the past month. What it means for hiring, pricing or the forecast is a management judgment, and the commentary should state it as one.
Nor will it tell you whether the budget was realistic. A large variance against a budget nobody believed says more about the budget than the business. AI will explain the gap faithfully either way, so someone has to say which kind of gap it is.
Frequently asked questions
What is AI variance analysis?
AI variance analysis uses software to calculate movements against budget, forecast or prior periods, identify the transactions behind them and draft plain-language commentary for a finance professional to review and approve.
Can AI write board variance commentary?
AI can write a good first draft, provided the figures come from the ledger and a named reviewer confirms each explanation before the commentary goes to the board. It should not be the final author.
How do you stop AI from inventing explanations?
Require evidence for every explanation, such as journal entries, operational data or a confirmation from the budget owner, and treat any cause inferred only from an account name or past pattern as a question rather than a finding.
What materiality threshold should variance commentary use?
Use a threshold that combines a percentage and an amount, agreed before the period starts and documented, so that commentary focuses on movements that matter and small lines with large percentage swings do not crowd them out.
Who should review AI-generated variance commentary?
A named finance professional who understands the business, usually the FP&A lead or controller, with the CFO accountable for anything that goes to the board. The review should be recorded.
Where CFOLogic fits
Navigator, our software, reads QuickBooks Online, Xero or CFOLogic's hosted ledger read-only and explains the numbers in plain English, with each figure traced to the transactions behind it; the live demo runs on fictional data. In that design the ledger supplies every figure and the model only writes the explanation; for larger groups we design and build the same pattern over the ERP they already use, and a named person signs off anything bound for the board.