Budget season stalls on collection, not calculation. The agent sends each department head a pre-filled template, chases on a fixed timetable, checks each submission against the guidance and the headcount plan, and sends factual errors straight back. Challenges on judgement go to FP&A. The agent never edits a submission: department heads own their numbers and sign them off.
What is the process, and who owns it?
Each year, or each forecast cycle, FP&A asks department heads for their plans: headcount, spending, projects, and the assumptions behind them. The submissions arrive late, in different versions, with formula errors, missing lines and assumptions that ignore the guidance everyone was sent. FP&A spends weeks chasing and fixing before any real discussion of the numbers starts.
The FP&A lead owns the process, the timetable, the template and the guidance. Each department head owns their submission. HR owns the headcount plan and pay assumptions. The finance director and, in the end, the board approve the budget. Whether an agent can go further and prepare a budget itself is considered in can an AI agent prepare a budget. The wider finance position is in AI agents for finance: books you can audit.
How does the agent collect, step by step?
- Person, FP&A lead: issues the guidance and template. Planning assumptions, such as the pay increase and price changes, the deadline and what must be explained. Evidence: the guidance version.
- Agent, depth 1: pre-fills each template. Current-year actuals and forecast, current headcount from HR, recurring contracts. Department heads start from facts, not a blank sheet. Evidence: sources for each pre-filled figure.
- Agent, depth 3: sends templates and reminders to each department head on the timetable FP&A set. Evidence: messages sent and dates.
- Agent, depth 1: logs each submission. Who sent it, when, and which version. Later versions replace earlier ones, and every version is kept. Evidence: the version log.
- Agent, depth 1: checks factual consistency. Arithmetic, lines mapped to the chart of accounts, guidance assumptions applied, headcount matching the HR plan, required explanations present. Evidence: check results per submission.
- Agent, depth 3: returns factual errors to the department head with each error stated. It does not correct them. Evidence: the return message and the error list.
- Agent, depth 2: drafts judgement questions for FP&A. Large growth on a line, a new cost with no explanation, a project with no owner. FP&A decides which to raise. Evidence: the draft and FP&A's choice.
- Person, department head: signs off the final version. Evidence: the sign-off and the version it applies to.
- Agent, depth 2: consolidates signed-off submissions into the draft budget for FP&A to review.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Guidance | Nothing | None | FP&A sets assumptions |
| Pre-fill | Fills actuals, headcount, contracts | 1 | Nothing |
| Send and remind | Sends on the timetable | 3 | Changing the timetable |
| Log | Records versions | 1 | Nothing |
| Check | Tests factual consistency | 1 | Nothing |
| Return errors | Sends factual errors back | 3 | Nothing |
| Judgement | Drafts questions | 2 | FP&A raises them |
| Sign-off | Nothing | None | Department head |
| Consolidate | Builds the draft budget | 2 | FP&A review and approval |
Why does the agent return errors instead of fixing them?
Because the submission is the department head's commitment. If the agent fixes a formula, it might also change what the department head meant. If it maps an unmapped line to the "nearest" account, it decides where the cost sits. Small fixes made silently are how a department head ends up defending a budget they do not recognise.
Returning the error with a clear explanation costs a day at most and keeps ownership where it belongs. It also gives FP&A a record of who needed what help, which is useful when the template is redesigned.
What should department heads be asked to include?
Most of the template is familiar: people, suppliers, travel, projects. One line is newer and often forgotten: the running cost of AI agents a department uses or plans to use, including model usage and certification. It should sit with the department that benefits, not in a central technology line where nobody owns it. How to estimate it is set out in how to budget for running AI agents.
Which controls matter, and what evidence do they leave?
Pre-filled figures from sealed sources. Actuals from the ledger, headcount from HR. Evidence: source on every pre-filled cell.
Version control. Only the latest version counts, and every version is kept. Evidence: the version log.
No edits by the agent. Evidence: submissions unchanged between receipt and return, by version comparison.
Sign-off before consolidation. An unsigned submission is not consolidated. Evidence: the sign-off record.
Guidance version cited. If FP&A changes an assumption mid-cycle, every submission shows which version it used. Evidence: guidance version on each check.
What should you measure?
- Submissions on time, by department.
- Factual errors per submission, and how many rounds each department needed.
- Days from deadline to consolidated draft.
- Judgement questions raised by FP&A from the agent's drafts, and how many changed the numbers.
- Department head satisfaction with the process. Ask them after the cycle; how to gather that well is covered in how to collect feedback on an AI agent's work.
Where does it usually go wrong?
Templates are left open. Department heads add columns, rename lines, and the checks fail on structure rather than content. Lock the structure, leave the inputs open.
Reminders ignore seniority. A director receives the same chasing message as a team lead and takes offence. FP&A should approve the wording for each level.
Headcount and spend disagree. The submission adds three people but no pay cost. The headcount check catches it, but only if HR's plan is up to date before the cycle starts.
Assumptions change late. FP&A revises the pay assumption after half the submissions are in. Re-check all submissions against the new version and tell department heads exactly what changed.
Questions people ask
Can the agent suggest numbers to department heads?
It pre-fills facts and can show trends. Suggesting a budget figure moves it from collecting to preparing, which most finance teams keep with FP&A.
What if a department head never submits?
After the last reminder the agent tells the FP&A lead, who escalates. Some teams roll forward the current-year forecast as a placeholder, labelled clearly as not submitted.
Does this work for mid-year forecasts too?
Yes. The cycle is shorter and the template smaller, but collection, checking and sign-off work the same way.