How to prepare rolling forecasts with an AI agent

A rolling forecast keeps a fixed horizon by adding a month every time one closes. The agent does the rolling: it loads actuals, extends the horizon, collects budget holders' updates, applies the approved drivers and produces a versioned forecast with changes explained. FP&A owns every assumption, and the CFO approves the version that leaves finance.

What is a rolling forecast, and who owns it?

Unlike an annual budget, which is set once and compared against, a rolling forecast is updated regularly and always looks the same distance ahead. It is built from drivers (volumes, prices, headcount, rates) and from inputs that budget holders submit, such as planned hires or a new contract. It feeds cash planning, covenant monitoring and board reporting.

The FP&A lead owns the forecast: the model, the drivers, the calendar and the challenge of budget holders' inputs. Budget holders own their own inputs. The CFO approves each version for use. Whether an agent can prepare a budget raises the same split between mechanics and judgement, with higher stakes because budgets set targets.

What does the agent prepare?

1. Agent, depth 1: loads actuals. From the closed ledger for the month just ended, mapped to forecast lines.
2. Agent, depth 2: rolls the horizon. Adds the new final month using the approved driver method, so the horizon stays fixed.
3. Agent, depth 2: requests inputs. Sends each budget holder their lines with last forecast, actual and the questions that matter: "Your actuals ran 40,000 above forecast on contractors. Is this continuing?" Under certification, these standard requests can go at depth 3.
4. Agent, depth 1: collects and checks inputs. Flags inputs that conflict with drivers, such as headcount up but salary cost flat, or that differ sharply from history without a reason.
5. Agent, depth 2: builds the version. Applies drivers and inputs, and produces the forecast with a bridge from the last version: actuals in, inputs changed, driver changes, timing.
6. Agent, depth 2: drafts scenarios. Where FP&A asks, upside and downside versions with the driver changes stated.
7. Person, FP&A lead: reviews and challenges. Accepts or changes inputs and drivers.
8. Person, CFO: approves the version. It is then frozen and labelled.

What does a version bridge look like?

Take an illustrative services business whose last forecast showed full-year operating profit of 4,000,000. The new version shows 3,700,000. The agent's bridge explains the 300,000: actuals for the closed month came in 80,000 below forecast, mostly lower utilisation; the sales lead cut pipeline conversion on two large bids, worth 250,000; a planned rent review was confirmed 30,000 lower than assumed; and the rolled-in new month adds nothing material. FP&A challenges the pipeline cut, accepts it after seeing the bid status, and the CFO approves the version with the bridge attached.

What does a named person approve, and what evidence remains?

FP&A approves assumptions: the drivers, every material input and every change from the last version. The CFO approves the version for use, which means the forecast that goes to the board, lenders or anyone else.

Driver register. Each driver has an owner, a value and a reason. Evidence: register version on the forecast.

Input log. Who submitted what, when, and whether FP&A changed it. Evidence: the log per version.

Version bridge. Every version explains its difference from the previous one. Evidence: the bridge.

Frozen approved versions. Approved versions cannot be edited; new work creates a new version. Evidence: version history with approver.

Approved, frozen versions also serve as evidence beyond finance. When auditors look at going concern or impairment, they ask for the forecasts management relied on, so keep them in a form that feeds the year-end audit request list.

Where does it usually go wrong?

Extrapolation passed off as forecast. The agent extends recent trends and calls it a forecast. Trends are not assumptions; someone must decide that they will continue. The agent should show trend-based lines as "no input received" until a person accepts them.

Inputs overwritten. A budget holder's submission is replaced by a driver calculation in the next run. Inputs and driver lines should be marked so one never silently replaces the other.

Version confusion. Three drafts circulate and the board pack uses the wrong one. Only CFO-approved versions are labelled for use.

Tax and other irregular flows missed. Quarterly tax payments, annual insurance and bonuses distort monthly cash. Sales tax payable is often the largest; base it on the workpapers from preparing VAT return workpapers rather than an average.

Accuracy never measured. Forecasts that are never compared with outcomes never improve.

What should you measure?

• Forecast accuracy by line and horizon, comparing each version with later actuals.
• Inputs received on time from budget holders.
• Inputs changed by FP&A, a sign of challenge quality.
• Days from close to approved forecast.
• Bias: whether forecasts are consistently high or low on particular lines.

Where Pryme fits

In Pryme, loading actuals, rolling the horizon and requesting inputs are separate capabilities certified on your own past forecast cycles. Actuals arrive sealed from the ledger, budget holders' inputs as declared, so FP&A can see which lines rest on a person's word. Approved versions are filed with their bridge. The finance bounds are in AI agents for finance.