Agentic FP&A and the continuous forecast

Financial planning has historically been a periodic exercise. The annual budget is a commitment. The rolling forecast is its update. Both are produced at a cadence — quarterly, monthly, in some cases weekly — that reflects the time and judgement available from the planning team. The forecast is, by construction, slightly out of date the moment it is published.

Agentic AI changes the cost of refresh. A planning system connected to live operational data can, in principle, update its forecast continuously, as the underlying drivers move. That capability is real. It is also more easily oversold than understood. The CFO who deploys agentic FP&A without thinking carefully about governance produces a faster way to be confused.

What "agentic" actually means in FP&A

The word "agentic" is overused. In the FP&A context, it means systems that execute multi-step analytical tasks autonomously, rather than systems that answer single questions. The distinction matters because the failure modes differ.

A single-question AI tool answers a query. The user inspects the answer. If it is wrong, the user does not act on it.

An agentic system executes a chain of actions: it pulls data, applies rules, makes inferences, updates a model, produces a recommendation, in some configurations executes a downstream action. The user may not inspect each step. The failure modes are compound. A small error in step one becomes a larger error in step four. The user, several steps removed, may not be able to identify where the chain broke.

For an FP&A function, the agentic pattern is what allows the continuous forecast. The system observes a change in a driver — booking volume, conversion rate, cost of a commodity — and propagates the change through the model without a human assembling the inputs. The propagation is the value. It is also the risk.

The continuous forecast and why it is not the same as a faster one

A continuous forecast is structurally different from a frequent one. A monthly forecast updated weekly is still a periodic forecast; it is simply produced more often. A continuous forecast updates when the underlying data updates. The cadence is set by the data, not by the planning calendar.

The advantages are real:

• Driver movements are captured in the period they occur, not in the period they are reviewed.
• The forecast at any moment reflects the latest available data.
• Variance against the forecast is visible immediately, not at the next forecast cycle.
• Senior management can ask "what does the forecast say now?" and get a current answer.

The challenges are subtler:

• A forecast that changes every time the underlying data changes can become noise. If the board sees a different number each week, the planning function has produced volatility, not insight.
• The system's confidence in its own output is not the same as the user's confidence in it. A driver movement may be a real change, a measurement glitch, or a transient anomaly. The system does not always know the difference.
• The discipline of explaining variance — what changed and why — does not become easier when the forecast updates continuously. It becomes harder, because the baseline keeps moving.

Governance: when should the forecast update reach management

The governance question is not whether the forecast updates. It updates. The question is when an updated forecast crosses the threshold for management attention, board reporting, or external disclosure.

A defensible architecture has multiple confidence layers.

• The model output is continuous and visible to the planning team.
• A management view is produced at a defined cadence — weekly or monthly — with the planning team's review and commentary applied.
• The board view is produced at the board's cadence, with the planning team's interpretation reconciled to the prior period.
• External-facing forecasts — guidance, covenants, investor disclosures — are produced under the discipline that applied before AI: deliberate, reviewed, defensible.

Each layer absorbs some of the underlying volatility. Without those layers, the agentic forecast can produce an organisational pulse that exhausts the people who have to respond to it.

The judgement layer remains human

The agentic system can identify that a driver has moved. It can update the forecast based on the new driver value. What it cannot reliably do is interpret why the driver moved, predict whether the movement is persistent, or judge whether the business should change course in response.

That judgement requires context the model does not have: the upcoming product launch, the supplier conversation last week, the regulatory change pending in a key market. An FP&A function that lets the model speak for itself, without the human layer that interprets, will produce a forecast that is technically current and analytically wrong.

The role of the FP&A analyst, in this architecture, is not to assemble the forecast. It is to challenge it. To ask why the model has changed its view. To check whether the driver movement is genuine. To explain to the business what is happening and what to do about it. The analyst who can do this work is more valuable in the AI-enabled function than in the pre-AI one. The analyst who cannot will struggle to find the seat.

What the CFO must specify before the build

Before any agentic FP&A system is procured or built, the CFO should specify several things explicitly.

• The scope. Which drivers are in scope for automated update? Which are not — because the driver is too noisy, too contested, or too consequential to leave to the model?
• The confidence thresholds. What level of model confidence is required before an updated forecast is presented to which audience? Below the threshold, the forecast is held for human review.
• The audit trail. Every forecast output must be traceable to the inputs that produced it, the model version that ran, and the human review applied. "The system updated it" is not a planning policy.
• The exception process. How does a human override the model when context the model does not have makes the model's forecast wrong? Who can do this? Under what documentation?
• The performance review. The model's forecast accuracy is measured against actuals. The discipline of measuring and reviewing forecast accuracy is the same discipline that should apply to any forecasting function.

None of these specifications is exotic. They are the basic governance of an analytical function, applied to a tool that moves faster than the previous tools. The principles are unchanged. The cadence is not.

This piece sits inside the CFO in AI framework. See also AI governance for the finance function and AI in the financial close. Lorna writes from practice at IMPT. The verified page records what is and isn't published here.