To create a budget forecast with AI, first freeze the entity, period, currency, units, accounting sign convention, and verified actuals. Build a time-phased ledger, calculate every formula in a controlled spreadsheet or script, and reconcile totals before giving locked results to AI for a narrative draft. The language model must never be the formula source or overwrite approved numbers.

This guide creates one rolling forecast baseline for an operations team. It does not compare alternatives like a cost-benefit analysis, and it does not run multiple futures like scenario analysis. Use the broader controls in How to Use AI, especially when financial data is confidential.

Key Takeaways

Freeze period, entity, currency, unit, sign, and version.

Import verified actuals and commitments from named systems of record.

Store each driver, formula location, source, status, and owner.

Calculate in a deterministic spreadsheet or script, never in prose.

Check duplicates, missing rows, period bridges, subtotals, totals, and rounding.

Give AI only de-identified inputs and locked outputs for explanation.

Require budget owner and finance review before the forecast is used.

What question should the budget forecast answer?

State what is being forecast and why. “Forecast spending” is too broad. A usable question identifies an entity, accounts, start and end periods, reporting cadence, decision use, comparison baseline, and owner.

Field	Record
Forecast ID/version	Stable ID and revision
Entity/scope	Included business unit and exclusions
Period	Actual cutoff and forecast horizon
Grain	Month, week, account, cost center, or project
Currency/unit	Transaction and reporting currency; whole, thousands, or millions
Sign convention	Expense, income, cash-in, and cash-out signs
Baseline	Approved budget or prior locked forecast
Owner	Budget holder and finance reviewer

The GOV.UK Budget Holder Forecasting Handbook emphasizes clear roles, regular review, evidence, and management of assumptions in public-sector forecasting.［1］ Its governance principles are useful examples, but each organization must follow its own accounting policy and local requirements.

Create a controlled input register

List every source before importing data. Record system, extract date, period coverage, query or report ID, owner, and control total. Preserve the raw export read-only. Work on a versioned copy.

Do not upload real payroll, customer, vendor, bank, account, or unreleased financial data to an unapproved model. Design the workflow with synthetic rows such as ENTITY-A, ACCOUNT-4100, and 100 currency units. See AI privacy risks for minimization and access boundaries.

Before modeling, check:

the extract covers the intended entity and period;
the currency and unit are explicit;
dates use one calendar and time-zone rule;
account mappings are current;
reversals and credits follow the chosen sign convention;
the source control total matches the imported total;
duplicate keys and missing periods are reported, not silently repaired.

If cleanup is needed, use a separate, auditable process such as cleaning spreadsheet data with AI assistance, but keep deterministic validation as the authority.

Create a budget forecast with AI from a time-phased ledger

Use one row per stable key and period. Keep inputs, formulas, outputs, and commentary distinct.

Field	Purpose
Version	Identifies the locked model run
Entity/account/period	Composite row key
Currency/unit/sign	Prevents silent scale or direction changes
Actual	Verified posted amount
Commitment	Approved purchase order, contract, or other committed amount
Driver ID/value	Named assumption and numeric input
Formula location	Cell, named range, query, or script function
Forecast	Deterministic calculated output
Source/owner/status	Evidence and accountability

Never type a forecast number into a narrative and later copy it back into the model. The controlled ledger is the numeric truth; commentary is derived from it.

How do actuals, commitments, and drivers differ?

Actuals have occurred and should reconcile to the source. Commitments are authorized future obligations under the organization's definition. Drivers are assumptions such as headcount, usage, unit rate, timing, renewal, or seasonality. Do not blur the three categories.

Example with synthetic data:

Period	Actual	Commitment	Driver input	Forecast status
M01	100	0	Not applicable	Locked actual
M02	120	0	Not applicable	Locked actual
M03	—	80	Volume 10 × rate 12	Formula output 200

The table illustrates structure, not a real forecast. A deterministic formula would reference the commitment, volume, and rate cells. It should also state whether commitments are additive or already included, preventing double counting.

Define every driver and assumption

Give each driver an owner, source, effective period, unit, range, and review date. Record whether it is approved, provisional, or blocked.

Driver ID	Definition	Unit	Source	Applies to	Owner	Status
D-01	Synthetic monthly volume	items/month	Approved planning note	M03–M06	Operations	Provisional
D-02	Synthetic unit rate	currency/item	Signed rate table	M03–M06	Procurement	Verified

GAO's Cost Estimating and Assessment Guide discusses credible estimates, ground rules and assumptions, data, sensitivity, risk, documentation, and updating.［2］ A budget forecast is not identical to a cost estimate, but traceable assumptions and reproducible calculations are valuable controls in both.

Why must calculations stay in a controlled model?

Use a protected spreadsheet, database query, or reviewed script with explicit formulas and version control. AI can help describe a requested formula for a human to implement and test, but it must not be the execution authority.

For example, a controlled model may implement:

forecast[account, period] = committed_amount
                          + approved_volume * approved_unit_rate
                          + approved_fixed_adjustment

The real implementation must define null behavior, rounding, currency conversion, inclusions, and whether each component overlaps. Unit-test a script or use spreadsheet checks with known synthetic cases. Protect formula cells and record who changed them.

Never accept “the total looks right” or a model's arithmetic explanation. Recalculate from source cells using a separate check or equivalent deterministic path.

Check units, currency, signs, and periods

Many material errors are dimension errors rather than difficult mathematics. Run explicit tests:

Units: no row mixes whole currency with thousands or items with cases.
Currency: every conversion has an approved rate source, date, direction, and rounding rule.
Signs: credits, reversals, income, expenses, cash-in, and cash-out follow one documented convention.
Periods: actual cutoff and forecast start do not overlap or leave gaps.
Calendar: fiscal and calendar periods map through an approved table.
Rounding: detail and totals use the stated precision and reconciliation tolerance.

Flag, do not coerce, a row with an unknown dimension. Converting blanks to zero can hide missing inputs.

How do you reconcile detail, subtotals, and totals?

Use a check sheet or test report independent of presentation formatting.

Check	Expected result	Failure action
Source actual total − imported actual total	0 within defined tolerance	Stop and trace import
Duplicate composite keys	0	Stop and resolve ownership
Required account-period rows missing	0	Explain or add controlled row
Detail sum − account subtotal	0	Trace formula range
Account subtotals − entity total	0	Trace mapping/formula
Actual cutoff to forecast start	Continuous	Repair period map
Current version − prior version	Fully explained	Add variance owner

The Green Book discusses appraisal, uncertainty, assumptions, and transparent use of evidence for public decisions.［3］ It is not a universal accounting manual. Use its transparency principles as an example and let the responsible finance function define the applicable forecasting policy.

Check duplicates and missing rows

Define the expected key universe. A duplicate may double count; a missing row may incorrectly look like zero. Compare actual key counts with expected accounts, entities, and periods. Report added, removed, and unmapped keys between versions.

Do not let AI “clean” these exceptions by choosing one row or filling an amount. It may group exception messages after identifiers are de-identified, but the data owner must resolve them in the source or controlled transformation.

Bridge actuals to the forecast

The transition period deserves a specific check. Identify the last closed actual period, any open-period accrual policy, the first forecast period, and treatment of late postings. Ensure commitments are neither omitted nor counted in both actual and forecast.

Create a bridge:

Component	Amount	Source/formula	Owner	Status
Last locked forecast	1,000	Version F-07	Finance	Locked
Actual variance	+20	Deterministic actual comparison	Finance	Reconciled
Timing shift	-10	Approved driver change	Operations	Reviewed
New commitment	+30	Commitment register	Procurement	Verified
Current forecast	1,040	Sum of bridge components	Finance	Reconciled

Numbers are synthetic. The bridge must add exactly to the locked current forecast under the defined rounding rule.

Which outputs should be locked before AI drafts commentary?

Export a de-identified result table containing approved totals, variances, driver labels, and explanations. Mark it read-only and give it a version ID. Ask AI to draft narrative only from those values.

Draft a variance explanation from the supplied LOCKED table.
Copy numbers, units, periods, signs, and driver labels exactly.
Do not calculate, interpolate, forecast, rank, recommend, or fill blanks.
For each sentence cite the row or bridge ID. Mark ambiguity as [VERIFY].
Return commentary only; do not modify the table.

Compare every number in the draft to the locked export. Use fact-checking AI answers as the verification pattern. A fluent explanation can reverse a sign or confuse a timing shift with a permanent change.

Review budget variances and versions

For each material change, record amount, percentage where meaningful, period, driver, evidence, owner, and whether it is timing or permanent. Define materiality through finance policy, not a model-generated threshold.

Compare the current version with the prior locked version and the approved baseline. Preserve both the numerical bridge and narrative changes. A changed formula, mapping, actual cutoff, or rate can explain a variance even when the business driver did not change.

NIST's generative AI profile identifies confabulation and governance risks.［4］ Separating deterministic calculation from language generation, restricting data, and requiring human review reduces those risks but does not certify the forecast.

Approve and archive

The budget holder validates operational assumptions and commitments. Finance validates accounting treatment, model controls, reconciliation, materiality, and presentation. Other specialists review tax, treasury, procurement, or legal assumptions where applicable. They approve one identified version.

Archive the source-register version, controlled model, check report, exception resolution, locked output, narrative draft, human edits, approvals, and review date. Do not archive secrets or unnecessary personal data. Set triggers for new actuals, material commitments, rate changes, errors, or policy changes.

Final checklist

Entity, scope, period, grain, currency, unit, sign, and version are explicit.
Actuals reconcile to a named source and cutoff.
Commitments and drivers are separate, sourced, and owned.
Every formula lives in a controlled spreadsheet, query, or script.
Units, currency, signs, periods, calendar mappings, and rounding were checked.
Duplicate and missing keys were resolved, not silently coerced.
Detail, subtotals, totals, actual-to-forecast bridge, and version bridge reconcile.
AI received only approved, de-identified, locked results for narrative drafting.
Every narrative number maps to the locked output.
Budget owner and finance approved the same archived version.

FAQ

What is the minimum input for a budget forecast?

You need scope, periods, currency, units, sign convention, verified actuals, commitments, drivers, account mappings, sources, owners, and a controlled calculation model.

How should a rolling forecast handle periods?

Close actual periods under the approved policy, advance the forecast horizon consistently, and check that the actual cutoff and first forecast period neither overlap nor leave a gap.

Are commitments the same as forecast expenses?

No. Define commitments under local policy and state whether they are included in, or added to, driver-based forecast amounts to prevent double counting.

How should a driver-based forecast work?

Define each driver's unit, source, owner, approved value, applicable period, and formula mapping. Calculate it in the controlled model and reconcile the result to source rows; do not let AI supply a missing driver.

Can AI convert currencies or calculate totals?

Not as the source of truth. Use approved rate data and deterministic formulas, then reconcile. AI may explain locked results without changing them.

How should negative values be handled?

Document one sign convention and test credits, reversals, income, expenses, cash-in, and cash-out. Never infer meaning from a minus sign alone.

What should happen to duplicate or missing rows?

Flag and resolve them through the data owner and controlled transformation. Do not ask AI to select a duplicate or turn a missing amount into zero.

How do I check the math independently?

Recompute control totals and bridges through a separate deterministic check, verify formula ranges and mappings, and test known synthetic cases.

What is AI allowed to do in this workflow?

It can organize field definitions, surface missing explanations, and draft commentary from locked results. It cannot be the formula authority, approve assumptions, or alter numbers.

Sources

UK Government, “Budget Holder Forecasting Handbook” — https://www.gov.uk/government/publications/budget-holder-forecasting-handbook/budget-holder-forecasting-handbook
U.S. Government Accountability Office, “Cost Estimating and Assessment Guide” — https://www.gao.gov/products/gao-20-195g
UK Government, “The Green Book 2026” — https://www.gov.uk/government/publications/the-green-book-appraisal-and-evaluation-in-central-government/the-green-book-2026
National Institute of Standards and Technology, “Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile” — https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence

Sources checked 31 August 2026.

How to Create a Budget Forecast with AI—and Check the Math