Finance AI forecast baseline change control

Faster forecasting makes baseline control more important

OpenAI CFO Sarah Friar wrote on August 10 that every forecast should carry a clear explanation and every change to an approved baseline should require Finance authorization. That is the most useful control in the current AI-finance discussion because it separates preparation speed from decision authority.

AI systems can gather source material, update formulas, build scenarios, and draft a leadership narrative. OpenAI's current finance guidance promotes protected assumptions and traceable outputs. Its Model ML case study also shows why those controls matter. In the reported Excel benchmark, a model passed the workbook structure contract in every test, but only half of the workbooks had every key output right. A valid, editable file is not the same as a correct forecast.

FP&A practitioners make the same distinction from the other side. An August 5 FP&A Trends article says an agent can generate a baseline in minutes while FP&A applies business judgment. Its 2026 board paper is more direct: even if an agent produces the forecast, FP&A owns the call to the board. Current community discussions about AI-generated dashboards and presentations also emphasize that the valuable role shifts toward architecture, explanation, and knowing what the numbers mean.

The control problem begins when an AI-generated proposal quietly becomes the comparison basis. If one business unit updates volume, another changes price, and a model refreshes FX, the workbook may look internally consistent while leaders compare it with a baseline that no longer exists. Finance then cannot explain which assumption changed, who authorized it, which decisions moved, or whether the same version reached treasury, hiring, and the board deck.

Separate four states that spreadsheets often blur

A “base case” scenario is not automatically the approved baseline. A proposal can contain a base scenario while the prior approved forecast remains the official comparison. Keep those labels distinct until the release gate is complete. This is especially important in planning systems where a user can copy versions, change a working scenario, or promote it with one action.

Do not let the AI choose the state. State transitions belong in the planning system or a controlled change register. The AI may recommend that a proposal is ready for review and list missing evidence. It must not set an approved flag, overwrite the official version, or present a draft as management guidance.

Freeze more than the file

A workbook hash proves which bytes were reviewed. It does not capture a live planning model, external data connection, calculation setting, or source cutoff by itself. The baseline receipt should identify the forecast version, period and horizon, entity scope, reporting currency, book, dimensions, source cutoff, actuals status, assumption register, formula or rule version, mappings, materiality, and approvers.

baseline_id: FY27-Q1-FCST-03
status: approved_read_only
approved_at: 2026-08-18T16:30:00Z
scope:
  entities: [US, UK, DE]
  currency: USD
  horizon: 2026-09_to_2027-12
actuals_cutoff: 2026-07-close-v4
assumptions_digest: sha256:8d1...
formula_version: fp-model-12.4
source_register: sources-2026-08-18
approval_receipt: fin-auth-9917
downstream: [cash-plan-07, hiring-plan-09, board-pack-draft-03]

Record what was still preliminary. If the accounting period remains open, label late accruals, revenue adjustments, intercompany eliminations, and manual estimates. Oracle's UPL-licensed NetSuite finance-analyst skill uses this same operating stance: fix period, subsidiary, currency, book, and comparison basis; label open-period results preliminary; quantify drivers; and require qualified validation. The public skill was reviewed as a reference, not copied into this pack.

Baseline ownership is not an administrative detail. A named FP&A owner should decide which version is authoritative for comparison. The controller owns accounting basis and period maturity. Treasury owns cash and covenant interpretations. Business owners validate operational drivers. The system publisher controls promotion. One person may fill several roles in a smaller company, but the decisions should remain explicit.

Review driver changes before reviewing the finished forecast

A polished full forecast is difficult to challenge. Start with a driver-level register that explains what changed and why before examining the final totals. Each row should show the baseline value, proposed value, unit, effective period, absolute and percentage delta, approved source, rationale, owner, confidence, linked outputs, and status.

“The model predicts” is not an evidence source. The model may transform approved evidence into an estimate, but the row still needs an owner, method, version, uncertainty, and decision about use. External benchmarks can inform a scenario. They do not authorize a company assumption without a Finance and business owner.

Challenge timing before direction. A higher pipeline may shift revenue rather than increase the full-year total. A delayed hire can reduce near-term expense while reducing capacity later. A price change can be offset by mix or churn. The AI should identify these dependencies, not silently net them into one favorable variance.

Worked example: a growth proposal that changes four decisions

Suppose an AI-assisted refresh proposes raising third-quarter new-logo volume from 1,000 to 1,120 based on qualified pipeline and a new channel agreement. At the baseline selling price of $24,000, the gross revenue effect appears to be $2.88 million. That headline is not ready for approval.

Finance first separates price and mix. The new channel carries a lower expected selling price of $22,500, reducing the gross effect to $2.70 million. Historical implementation timing shows only 70% of signed customers begin recognizing revenue in the quarter, reducing recognized revenue to $1.89 million. Support and hosting costs add $0.41 million. The proposal therefore changes gross profit by about $1.48 million before considering churn, sales commissions, working capital, and tax.

The same assumption affects capacity. Delivering 120 incremental customers may require eight implementation hires. If hiring starts six weeks late, near-term payroll improves but customer activation slips. Cash collection may lag revenue by forty-five days. Treasury sees a different quarter-end cash effect from the P&L. A covenant based on trailing EBITDA may not move in the same period as bookings.

The decision packet should therefore show at least three cases. The downside case uses lower conversion and a longer activation lag. The base proposal uses the evidenced assumptions. The upside case may use faster activation but must not assume both maximum volume and minimum discount unless evidence supports that combination. Leaders can then see whether the hiring, marketing, and liquidity decisions change across the range.

An independent recalculation should reproduce the revenue, margin, cash, and headcount bridges from the registered drivers. If the workbook proposal and the independent result differ beyond tolerance, the release stops. The AI can help locate the difference; it cannot waive it.

Scenarios are controlled assumption sets, not alternative stories

Define a scenario by its changed drivers and rules, not by a label alone. “Downside” should say which assumptions move, by how much, over which period, and why. Keep fixed assumptions visible so a reviewer can detect accidental changes. Do not let the AI vary every input at once; that produces a range but little decision insight.

1. Choose the two to five drivers that can actually change the decision.
2. Set plausible ranges from approved evidence, not a model's confidence language.
3. Keep accounting basis, entity scope, currency, and formula versions constant unless the scenario is explicitly about them.
4. Recalculate every linked output and show where a decision threshold is crossed.
5. Record which owner accepted the range and whether probability weighting is permitted.
6. Preserve the baseline in every comparison; never compare two moving proposals without a fixed reference.

Probability-weighted scenarios can create false precision. Use probabilities only when the organization has an approved method and accountable owner. Otherwise present conditional cases and sensitivity. The central question is often not “Which number is most likely?” but “At what assumption does the hiring, liquidity, or spending decision change?”

Release one exact version with a decision receipt

The release gate should decide APPROVE, APPROVE WITH LIMITS, REWORK, HOLD FOR EVIDENCE, or REJECT. It should identify the exact forecast version and digest, the baseline it replaces, approved assumption changes, remaining exceptions, downstream targets, monitoring thresholds, rollback path, and next review date.

Approval with limits is useful when evidence supports internal scenario planning but not board, lender, or investor use. The receipt can restrict the audience, decision, time window, or entity scope. A provisional forecast based on incomplete actuals may support operational staffing while remaining prohibited for external guidance.

Update downstream artifacts as a controlled set. The planning model, cash forecast, hiring plan, dashboard, operating review, board pack, covenant model, and executive narrative should all reference the same release ID. If one surface cannot be updated, record it as an exception rather than allowing two official forecasts to circulate.

Monitor assumptions after release. Define thresholds for material actual-versus-forecast movement, source failure, delayed hiring, FX change, pipeline conversion, collection timing, covenant headroom, and data-quality exceptions. Crossing a threshold opens a new change request; it does not silently edit the released forecast.

Failure modes that make fast forecasts ungovernable

A 30-day pilot for one forecast lane

1. Select one bounded forecast lane with an existing owner, baseline, and decision calendar.
2. Freeze the baseline receipt, assumptions register, calculation logic, materiality, and downstream surfaces.
3. Run two or three AI-assisted proposals without publishing them.
4. Measure source coverage, unsupported assumptions, recalculation differences, review time, rework, and exceptions.
5. Test at least one late-actual, wrong-sign, stale-source, duplicated-row, and unauthorized-assumption case.
6. Compare the AI-assisted process with the existing forecast process using accepted output, total labor, cycle time, and decision usefulness.
7. Have a different Finance reviewer challenge the driver register and scenario ranges.
8. Issue one mock release receipt and verify every downstream link uses the same version.
9. Decide SCALE, LIMIT, REDESIGN, HOLD, or STOP with named owners and evidence.

Success is not “the AI built a forecast.” Success is that Finance reached an accepted, explainable decision faster without weakening source lineage, calculation integrity, review, or version control. Count the time spent preparing, verifying, correcting, explaining, and reconciling downstream artifacts. Faster generation with more review is not automatically a net gain.

Frequently asked questions

Can AI automatically update an approved forecast?

It can prepare a proposal in a controlled workspace. A named Finance owner should authorize every change to the approved baseline after evidence, driver diffs, recalculations, scenarios, and downstream impacts are reviewed.

What exactly is the forecast baseline?

It is the exact approved forecast version plus its scope, currency, period, source cutoff, assumptions, formulas, mappings, materiality, and authorization receipt. It is the fixed reference for explaining change.

Should Finance let AI generate assumptions?

AI may propose estimates or transform approved evidence. Every material assumption still needs a source or documented method, owner, version, uncertainty, and human decision about use.

How do we verify an AI-updated workbook?

Check population and source control totals, formula and mapping changes, driver-level diffs, material outputs through an independent calculation path, scenario boundaries, and linked cash, headcount, working-capital, tax, and covenant impacts.

Can the same AI draft the forecast narrative?

Yes, as a draft grounded in the approved change register. A Finance owner must verify every material statement, separate fact from judgment, and approve the audience and release version.