AI Agent for Planning, Budgeting & Forecasting
A forecast is only useful if you can see what it assumes. This agent builds plans from named drivers with their ranges stated, re-runs the whole model when one of them moves, and attributes every change in the outcome to the assumption that caused it.
What is an AI FP&A agent?
An AI FP&A agent is a governed software worker for financial planning and analysis. It builds budgets and forecasts from explicitly declared driver assumptions, runs scenarios as alternative assumption sets against a single model, performs sensitivity analysis, and attributes every change between forecast versions to the drivers and actuals responsible.
What it does
What it is not
A spreadsheet nobody can re-run
Planning models accumulate hard-coded values, overrides and adjustments that made sense at the time, until the assumptions that produce the numbers exist only in the memory of whoever built it. Asking what happens if one changes becomes a rebuild rather than a recalculation.
Assumptions are buried in cells
A growth rate is typed into a formula rather than declared, so nobody can list what the plan actually assumes.
Scenarios are copies
Each scenario is a duplicated workbook that diverges from the base, and reconciling them later is impossible.
Reforecasts lose the reasoning
The new forecast differs from the old one and nothing records which assumption changed to produce the difference.
The model outlives its author
The person who built it moved on, and the remaining team maintains a structure they do not fully understand.
Plans you can interrogate, not just open
Structure
Assumptions Declared, Not Typed In
Every driver has a name and a range.
Each planning input is declared as a named driver with its value, its range and where it came from, so the plan can be described by listing its assumptions rather than by reading the formulas that consume them.
- Every input declared as a named driver
- Ranges and sources recorded per driver
- No hard-coded values inside formulas
- Plan describable as an assumption list
Named and sourced
Scenarios
Change One Thing, See What Moves
Same model, different assumptions.
Scenarios are assumption sets against one model rather than copies of a workbook, so a downside case differs from the base only in the drivers that were changed and the comparison between them is exact.
Assumption sets
Attribution
Why This Forecast Differs
From the last one, by driver.
When a forecast is revised, the change from the previous version is attributed to the specific assumptions that moved and the actuals that landed, so a reforecast comes with an explanation rather than a new number.
By assumption
How the AI FP&A Agent runs a task
- STEP 01
Anchor on actuals
The plan is based on ledger actuals for the periods already closed rather than on a prior forecast, so a rolling reforecast starts from what happened rather than compounding the assumptions of the last version.
Actuals extractionPeriod anchoring - STEP 02
Declare the drivers
Every input that moves the outcome is named, given a value and a plausible range, and attributed to its source — a signed contract, a hiring plan, a historical rate — so the plan is describable as a list of assumptions.
Driver declarationRange setting - STEP 03
Build the model once
Relationships between drivers and outputs are expressed as calculations over the declared inputs, with no hard-coded values, which is what makes a change to one assumption a recalculation rather than a rebuild.
Model constructionCalculation chain - STEP 04
Run the scenarios
Alternative cases are expressed as different assumption sets against the same model, and the comparison between them isolates exactly which drivers differ and what each contributes to the outcome gap.
Scenario setsComparison output - STEP 05
Attribute each revision
When a forecast is updated, the difference from the previous version is split between actuals that landed differently and assumptions that were changed, so a revision arrives with its explanation attached.
Version comparisonChange attribution
Systems the AI FP&A Agent connects to
Inputs
Modelling
Inputs, outputs and runtime
- Ingests
- Ledger actualsDriver assumptions and rangesHeadcount and hiring plansContract and renewal dataPrior forecast versions
- Produces
- Budget or forecast versionFull assumption registerScenario comparisonSensitivity by driverRevision attribution
- Triggered by
- Budget cyclePeriod actuals closingScenario request
- Human oversight
- Management owns assumptions and approval
- Models
- Open-weight LLMs you host — Llama, Qwen or Mistral class
- Typical latency
- Minutes to re-run a full scenario set
- Deployment
- On-premise or sovereign cloud with egress control
- Data residency
- Plans and cases stay inside your network
Where the FP&A Agent pays back
Annual Budget Construction
Build the budget from declared drivers so every line can be explained by the assumption that produced it.
Rolling Reforecast
Replace forecast periods with actuals as they land and attribute the resulting change to its causes.
Scenario Planning
Run upside, downside and stress cases as assumption sets against the same model rather than as separate files.
Sensitivity Analysis
Identify which drivers the outcome is genuinely sensitive to and which barely move it at all.
Headcount Planning
Model hiring plans with start dates and ramp assumptions and show the cost profile each produces.
Board Scenario Packs
Prepare the comparison between cases with the assumption differences stated explicitly for each.
AI FP&A Agent vs chatbots and SaaS copilots
Planning tools are usually judged on how quickly they produce a number, when the property that determines whether a plan survives contact with reality is whether anyone can say what it assumed.
| Generic chatbot | SaaS copilot | VDF AI | |
|---|---|---|---|
| Assumptions | Implicit | In cells | Declared with ranges and sources |
| Scenarios | Regenerated | Copied workbooks | Assumption sets on one model |
| Re-running | Starts over | Manual rebuild | Recalculation from drivers |
| Revision explained | Not attempted | Not attempted | Attributed to driver changes |
| Base data | What you paste | Uploaded file | Ledger actuals directly |
| Approves the plan | Not applicable | Not applicable | Management approves |
| Where the downside case sits | Vendor service | Vendor tenancy | Inside your own network |
Governance and controls
A plan becomes a commitment once it is approved, so what matters for control is that the assumptions behind an approved version are recorded and that the version itself cannot quietly change afterwards.
Assumption register per version
Every plan states what it assumed
Versions immutable
An approved plan is not edited in place
Actuals read-only
The agent cannot post to the ledger
Approval by management
Only people approve a budget
Ranges not point estimates
Uncertainty is carried in the driver
Scenario access controlled
Downside cases restricted by role
Evidence it leaves behind
What changes after rollout
Who runs the AI FP&A Agent
Head of FP&A
Answers a scenario question in the meeting rather than the following week, because a changed assumption recalculates the model instead of requiring a new workbook to be built and reconciled.
Finance analyst
Maintains a model whose logic is legible without archaeology, and produces reforecasts that arrive with an explanation of what moved rather than a number requiring justification afterwards.
Chief executive
Sees which assumptions the plan is genuinely sensitive to, which focuses management attention on the three or four drivers that determine the outcome rather than on the whole model.
Questions about the AI FP&A Agent
What is an AI FP&A agent?
It is an agent for financial planning that builds budgets and forecasts from named driver assumptions, treats scenarios as assumption sets against one model, and attributes every change between forecast versions to the driver or actual that caused it.
How is an AI FP&A agent different from a generic chatbot?
A chatbot can produce a plausible-looking projection. This agent builds from your actuals and declared drivers, and every figure it produces can be traced to the assumption that generated it.
Can an AI FP&A agent run on-premise on planning and actuals data?
Yes. A forecast reveals hiring intentions, renewal expectations, pricing plans and downside cases well before any of it is public, which is why planning stays inside your perimeter.
What does an AI FP&A agent produce, and in what format?
A plan or forecast with its full assumption list, scenario comparisons against one model, sensitivity results by driver, and an attribution of each change from the prior version.
Where does an AI FP&A agent fit in a governed AI programme?
It builds the model; people own the assumptions. Setting targets, approving a budget and committing to a plan are management decisions, and posted actuals come from the ledger.
Does it replace our planning software?
Not necessarily. Where you run a dedicated planning platform, this agent works alongside it — building and interrogating driver models, running scenarios and explaining revisions — while the platform remains the system of record for approved versions. Where planning currently happens in spreadsheets, the practical gain is largest, because the assumption register and the re-runnability are what spreadsheets structurally lack.
How does it decide what the driver assumptions should be?
It proposes them from historic data, signed contracts and stated plans, and the business owns them. A renewal rate derived from three years of actuals is a defensible starting point; whether next year will resemble those years is a management judgement. The agent makes the proposal and its basis explicit so that judgement is made deliberately rather than inherited from a spreadsheet.
Can it approve a budget or commit to a target?
No. It produces versions with their assumptions recorded, and approval is a management act with organisational consequences. Once a version is approved it becomes immutable, and later changes create a new version with its own attribution — which is what allows a board to compare what was approved against what is now forecast and see exactly what moved in between.
How does it handle forecast accuracy over time?
By comparing each prior forecast against the actuals that eventually landed and reporting which drivers were systematically wrong. That is more useful than an overall accuracy figure: a forecast consistently missing on one assumption is a fixable problem, whereas an aggregate error rate tells you only that forecasting is difficult, which everyone already knows.
How does this differ from the financial analyst agent?
One looks backward and one looks forward, and they meet at the actuals. The financial analyst explains what the ledger says happened and why, decomposed to postings. This agent takes those actuals as a base and projects forward from declared assumptions. Running both means a reforecast can distinguish between a driver assumption being wrong and the base period being misunderstood.
Interrogate the plan, not just the number
See the AI FP&A Agent build a driver-based forecast and re-run a scenario.