Partly. An AI agent can do the arithmetic and the digging of variance analysis: compute budget and prior-period variances, split them into price, volume and timing, and trace each one to the transactions behind it. What it cannot supply is the reason a number moved. The budget holder explains causes, and a finance business partner challenges them.
Where does the agent's work end and judgement begin?
Variance analysis has two halves. The first is finding what changed: which accounts, which cost centres, which suppliers or customers, and by how much. That is data work, and an agent does it faster and more thoroughly than a person paging through a ledger.
The second half is saying why. "Marketing spend is over budget by 30,000 because of an agency invoice" is a finding. "Because we brought forward the spring campaign to respond to a competitor" is an explanation, and only someone who was in the room knows it. An agent that writes the second sentence from the first is guessing, and guesses written in confident prose are worse than gaps. That argument is made at length in variance commentary needs a person's judgement.
What may the agent prepare?
| Output | Depth | Notes |
|---|---|---|
| Variance tables against budget, forecast and prior period | 1 | Read-only on the ledger and budget |
| Driver split: price, volume, mix, timing | 2 | Shows the method used for each line |
| Transaction drill-down for each material variance | 1 | Lists the largest contributing entries |
| One-off and timing flags | 2 | Accruals reversed, invoices booked in the wrong month |
| Draft commentary | 2 | Marked as findings, with the cause left for the owner |
| Questions for budget holders | 2 | Specific, one per material variance |
The agent's draft should separate what it found from what it infers. "Agency costs rose by 30,000, of which 25,000 is one invoice dated the 28th" is a finding. "Possibly a timing difference" is an inference, labelled as one.
Which step keeps the separation of duties?
The explanation and its challenge are done by different people. The budget holder supplies the cause, because they own the spend. A finance business partner, who does not own the budget, tests it against what the agent found. If the budget holder says the overspend is timing and the agent's drill-down shows no matching underspend expected next month, the partner asks the question.
The agent must not be the one that closes that loop. If it drafts both the cause and the acceptance of the cause, the budget holder's explanation is never tested, and an overspend that should have been challenged becomes a sentence in a pack. The detailed workflow is in how to prepare variance commentary for management accounts.
An illustrative example
Take a logistics company reviewing its monthly results. Fuel costs are 48,000 over budget.
The agent splits the variance. Price accounts for 20,000: the average cost per litre was higher than budgeted. Volume accounts for 31,000: more litres were used. A small favourable mix effect of 3,000 offsets part of it. It then drills into volume and finds that two depots used far more fuel per delivery than the rest.
It drafts commentary stating those findings and sends the operations director one question: why did fuel per delivery rise at those two depots? The director explains that a bridge closure forced longer routes for three weeks. The finance business partner checks the dates against the depots' daily data, finds they match, and accepts it. The pack goes out with the director's explanation and the agent's numbers, each marked as such.
What controls make it safe?
- Read-only access. The agent analyses; it does not post corrections it finds along the way. Misposted items go to a person as proposed reclassifications.
- Materiality set by people. The thresholds for what gets commentary are agreed in advance.
- Findings labelled. Every draft distinguishes evidence from inference.
- Named explainer and challenger. Recorded for each material variance.
- Same method every period. The driver split uses a documented method, so variances are comparable over time.
The same discipline helps upstream, in flux analysis on the trial balance, and outside finance, in complaint root cause analysis, where the agent finds patterns and a person names the cause. The broader controls are in AI agents for finance: books you can audit.
Questions people ask
Can the agent reforecast after finding a variance?
It can draft a revised forecast line showing the effect if the variance continues. Whether to change the forecast is the budget holder's decision, approved by finance.
How does the agent decide which variances are material?
By thresholds you set, usually a value and a proportion of budget together. The agent applies them consistently and lists anything just below the line for information.
Will budget holders trust numbers the agent produced?
They will if every number traces to transactions they can open. Drill-down links in the draft do more for trust than any description of how the agent works.