Step by step: which parts of budget variance reporting are plain rules, where AI drafts commentary and groups drivers, and which explanations and decisions stay with people.
Partly. Pulling actuals, matching them to the budget and calculating variances is rule-based automation, AI is useful for drafting the commentary and grouping the likely drivers, and people still have to explain why the numbers moved and decide what to do about it. The arithmetic was never the hard part of a variance report. The hard part is the week of chasing budget owners for a sentence that says more than "timing". Below you will find each step with an honest label, what has to be in place first, and a small pilot on one cost centre. The aim is not a report nobody reads, but a report where every material variance has a named owner, a reason that holds up and a decision.
What each step needs
01● Standard automation
Pull actuals and the approved budget into one view
This is an extract and a join, not a model. A scheduled job reads the general ledger, maps accounts to the reporting lines in the budget and loads both into one table for the period. The mapping is fixed and versioned, so a cost that moves from one line to another is a deliberate change and not a surprise. Locking the budget version matters as much as the actuals: if the budget file changes quietly between months, every variance after that is quietly wrong. Run the job after the close checklist says the period is final, not on a calendar date.
⛨Works when the chart of accounts maps cleanly to budget lines, the budget has a locked version and the extract runs only after the period is closed.
02● Standard automation
Calculate variances and apply materiality thresholds
Variance in amount and in percentage, month and year to date, against budget and against the latest forecast, is arithmetic. So is the decision about which lines deserve a comment. Set thresholds in advance, for example both a percentage and an absolute amount so that a 40 percent swing on a tiny line does not drown out a small percentage on a large one. Rules are the right tool because the same line must always be treated the same way, and because you must be able to show an auditor or a board member why one variance was explained and another was not.
⛨Works when the thresholds are agreed with the finance lead, written down and changed only on purpose, and when sign conventions for cost and revenue are consistent.
03● AI candidate
Group variances into likely drivers
Here AI helps. A model can read the underlying transactions behind a variance, such as invoices, purchase orders and payroll postings, and propose drivers: one large supplier invoice booked early, a price increase on a recurring contract, an unplanned hire, a one-off legal fee. It is better than a filter at spotting that forty small postings share a cause. Treat its output as a lead, not a finding. Have it list the transactions behind each proposed driver so a person can check them in seconds.
⛨Works when the model sees transaction-level detail with vendor and description fields, must cite the postings it relied on, and a reviewer spot-checks the groupings every month at first.
04● AI candidate
Draft the first version of the commentary
A model can turn a variance, its drivers and last month's commentary into a clear draft: what moved, by how much, and what the data suggests. That saves the finance team from staring at a blank cell for forty lines. The draft must stay factual. It may say that spend on contractors was above budget by a given amount because of three named invoices; it may not say the overspend is temporary unless someone who knows has said so. Mark every sentence that is an inference so the budget owner knows what to confirm or correct.
⛨Works when drafts carry a visible label, inferences are separated from facts, and no draft goes to management without a human edit.
05● Human review
Get explanations and corrections from budget owners
Only the person who runs the cost centre knows whether the overspend is a delayed project, a bad estimate or a decision made last quarter. Send each owner their own material variances with the AI draft attached and ask three questions: is the draft right, is this timing or permanent, and what happens next month. Rules can send the request and chase it on a schedule, but the answer is human. This is also where the report earns its value, because a variance with no owner and no reason is just a number.
⛨Works when each budget line has a named owner, owners get only their own lines, and there is a fixed deadline with an escalation path to the finance lead.
06● Keep human
Decide on reforecasts, corrective actions and approvals
Whether to hold spend, move budget between teams, accept a permanent overrun or revise the full-year forecast is a management decision with consequences for people and priorities. AI can show the effect of each option on the year-end number, but it should not choose. Keep this step with the finance lead and the budget owners, record the decision next to the variance, and make sure a changed forecast goes through the normal approval route instead of being edited quietly.
⛨Works when decision rights are clear, forecast changes need a named approver, and every decision is logged with its reason.
07● Standard automation
Assemble and distribute the report, then log what was changed
Once explanations are in and decisions are made, formatting and sending is automation. A template builds the summary, the table of material variances and the owner comments, exports the pack and sends it to the agreed list on the agreed day. Store the version that was sent, the AI draft and the final text side by side. That log is how you find out over time which AI drafts people accept and which they rewrite, and it is the evidence you need if someone later asks how the report was produced.
⛨Works when the distribution list is controlled, the sent version is stored read-only and edits to AI drafts are captured for review.
A sensible first experiment
Pick one department with a manageable number of cost centres, for example marketing, using the cost centres in your monthly management accounts. For three month-ends, let rules pull actuals and calculate variances, let AI propose drivers and draft commentary in review mode, and let the budget owner correct every draft before it goes anywhere. Measure three things: how long the finance team spends on commentary per month, the share of AI drafts a reviewer accepts without a factual change, and the number of material variances that still lack an owner explanation at the deadline. Expand only when drafts are accepted reliably and the missing explanations go down.
The trap to avoid
The usual mistake is letting the model write a confident reason that nobody checked. A plausible sentence such as "the overspend is due to timing and will reverse next quarter" can end up in a board pack even though no one with knowledge said it. Keep AI drafts labelled, separate facts from inferences, and require the budget owner to confirm every reason for a material variance.
Questions teams ask
Can AI write the variance commentary on its own?
It can write a useful first draft from the transactions behind a variance, but it cannot know whether a cost is timing or permanent. A budget owner should confirm every reason for a material variance before the report is sent.
Is variance analysis just a spreadsheet job?
The calculation is, and it is best done with fixed rules rather than a model. The time goes into explaining the numbers and chasing owners, which is where AI drafting and automated reminders help most.
What data does the AI need to see?
Ideally transaction-level postings with vendor, description and cost centre, so it can point to the invoices behind a variance. Check the vendor's terms, keep personal data such as payroll to a minimum and log what the model read.
How do we know the AI drivers are right?
Make it list the postings behind every proposed driver and have a reviewer sample them each month at first. Track how often drafts are changed and tighten the instructions where the same mistakes keep appearing.
Illustrative workflow guidance by Arcgent. Each business needs its own assessment. No integration or savings claim has been verified for your systems.