How to check spend against approved budgets with an AI agent

The agent compares what has been spent and committed with the budget that was actually approved, line by line, and warns before a line runs out, not after. It drafts a variance note for each budget holder to explain. FP&A approves the pack. Moving budget between lines, approving overspends and changing the budget stay with the people who own them.

What is the process, and who owns it?

Budget monitoring answers a simple question each month: is each cost centre spending what it was approved to spend? The difficulty is in the inputs. Actuals sit in the general ledger, commitments in purchase orders, card spend arrives late, and the budget itself exists in several versions. A check that uses the wrong version, or misses commitments, gives budget holders false comfort.

The FP&A lead owns the process, the budget versions and the monthly pack. Each budget holder owns their cost centre and its variances. The finance director approves budget transfers above a threshold. The wider finance position is in AI agents for finance: books you can audit.

What does the agent prepare, step by step?

1. Agent, depth 1: reads the approved budget version. The one FP&A has marked current, with its phasing by month. Evidence: the version identifier on every figure.
2. Agent, depth 1: reads actuals and commitments. Posted actuals from the general ledger, open purchase orders not yet invoiced, and card spend not yet coded. Evidence: each source with its cut-off time.
3. Agent, depth 1: removes double counting. A purchase order that has been invoiced appears once, as actual. Evidence: the matching between commitments and actuals.
4. Agent, depth 2: computes position per line. Budget to date, actual plus committed to date, remaining for the year, and a simple projection from the run rate. Evidence: the calculation per line.
5. Agent, depth 2: flags lines. Over budget to date, projected to overrun, or underspent enough to suggest a timing or coding problem. Evidence: the flag and its threshold.
6. Agent, depth 2: drafts variance notes for each flagged line, with the transactions behind the variance, for the budget holder to complete.
7. Person, budget holder: explains or acts. Confirms the explanation, corrects a miscoding through finance, or requests a transfer.
8. Person, FP&A lead: approves the pack before it goes to leadership.

What does a named person approve?

Three things. The FP&A lead approves which budget version is current and the thresholds for flags; if the agent could choose the version, it could choose the one that shows no problem. Each budget holder approves the explanation of their own variances, because they know what the spend was for. And the finance director, or whoever holds the authority, approves any transfer between lines.

The agent can also check a new requisition against remaining budget before it is approved, and show the approver the effect. That is advice at depth 1. Refusing or approving the requisition stays with the approver. Approval routes are themselves the subject of policy, and comparing practice to written procedure is covered in checking procedures against policy.

Where does it usually go wrong?

The wrong version. A reforecast replaced the original budget, but the agent reads the original. Every line looks wrong. Fix the version marker before the first run and cite it on every page.

Miscoding looks like good discipline. A cost centre is under budget because its spend was coded elsewhere. Large underspends deserve a variance note as much as overspends.

Card spend lags. Company card transactions arrive weeks after the spend, so the month looks under budget until they land. Include uncoded card spend as a commitment; how card spend is reconciled is covered in can an AI agent reconcile company card spend.

Phasing is ignored. An annual subscription paid in one month makes that month look overspent. Use the phased budget, not an even split.

Projections are read as forecasts. A run-rate projection is arithmetic, not a forecast. Label it plainly, or budget holders will treat it as a target or a promise.

Payroll costs move late. Headcount changes reach the ledger after payroll runs. Where people costs are a large share of budget, take the checked draft payroll into account; the checks are described in checking payroll changes before the payroll run.

Which controls matter, and what evidence do they leave?

Version cited on every figure. Evidence: the version identifier in the pack.

Cut-off stated per source. Evidence: the timestamp for the ledger, orders and card feed.

Commitments matched, not added twice. Evidence: the match log.

Explanations from budget holders, not the agent. The agent's draft lists transactions; the budget holder writes why. Evidence: the holder's name on each note.

Read-only. The agent cannot post, recode or transfer. Evidence: all verdicts at depth 1 or 2.

What should you measure?

• Overruns spotted before they happened, against those found after.
• Variance notes completed on time by budget holders.
• Miscodings found through the underspend flags.
• Time from month-end to pack approved.

Where Pryme fits

In Pryme, the budget agent's sources are stamped with their provenance: the ledger and approved budget are sealed, while a budget holder's explanation is declared and marked as theirs. The agent works at depth 1 and 2 only, so every figure in the pack can be traced, and nothing in the books changes because of it.