10 AI prompts to build a reliable quarterly rolling forecast in 2026
10 AI prompts to build a reliable quarterly rolling forecast: drivers, scenarios, variance, exec narrative. Copy-paste ready. CFO playbook 2026.
Arthur G.
A quarterly rolling forecast is a financial projection that always looks 4 quarters ahead, refreshed every quarter from actuals and explicit business drivers, not a static annual budget. This article delivers 10 copy-paste-ready prompts to build it with a private LLM (Claude, ChatGPT, Gemini): driver selection, statistical baseline, optimistic/central/pessimistic scenarios, plan-vs-actual variance, quarterly re-forecast, 13-week cash, stress test, Comex narrative, SaaS MRR/ARR model, quality audit. Each prompt includes persona, expected data, deliverables and quantified guardrails. Target audience: CFOs and FP&A leads at B2B scale-ups.
Why move to AI-driven rolling forecasting
The static annual budget is dying: signed off in March of a year that started in January, it is often obsolete by H2. The quarterly rolling forecast (always 4 forward quarters) fixes this but traditionally suffers two flaws: production cost (3-4 weeks at mid-market) and uneven hypothesis quality. LLMs fix both:
- Generate a baseline from actuals + drivers + seasonality, no blank-sheet restart.
- Document every hypothesis: what changes, why, confidence range.
- Run scenarios instantly (3 levers × 3 values = 27 scenarios in 10 minutes).
Three structuring principles:
- Driver-based: no forecast without explicit drivers (volumes, prices, churn, sales ramp).
- Rolling horizon: 4 forward quarters, never a year-end stub.
- Traceability: every cell ties to a documented hypothesis.
How to start in 5 steps (before using the prompts)
- Scope the perimeter - grain (entity, segment, offer) and depth (full P&L or top lines).
- Clean 24-month actuals - strip one-offs, harmonise the chart of accounts, mark scope breaks (acquisitions, divestments).
- Pick a privacy-grade LLM - Claude (Anthropic) or ChatGPT Enterprise for strategic data. Never consumer versions.
- Define drivers with prompt 1, validate with the business, freeze the list.
- Chain prompts 2 → 10 in order, versioning every output (date, author, model).
Table: which prompt for which persona?
| Prompt | Main persona | Cadence |
|---|---|---|
| Prompt 1 — Drivers | FP&A Lead | Annual + quarterly review |
| Prompt 2 — Baseline | Controller | Monthly |
| Prompt 3 — Scenarios | CFO | Quarterly |
| Prompt 4 — Variance | FP&A | Monthly |
| Prompt 5 — Re-forecast | FP&A Lead | Quarterly |
| Prompt 6 — 13-week cash | Treasurer | Weekly |
| Prompt 7 — Stress test | CFO | Quarterly / pre-board |
| Prompt 8 — Comex narrative | CFO | Quarterly |
| Prompt 9 — SaaS MRR/ARR | SaaS FP&A | Monthly |
| Prompt 10 — Quality audit | CFO / senior controller | Before every publication |
Prompt 1 - Pick the business drivers
Use case: extract the right drivers to steer the forecast.
Persona: FP&A Lead.
You are FP&A Lead at a B2B SaaS scale-up. From the company description and 24-month P&L, propose 8 to 12 business drivers for a rolling forecast. Answer in English.
Expected data:
1. 24-month P&L (revenue by line, COGS, OPEX).
2. Revenue model (subscription, transaction, services).
3. Average sales cycle and team structure.
4. Observed seasonality.
Deliverables:
1. Driver table: name, formula, data source, frequency, sensitivity (€ impact at ±10%).
2. Top 3 drivers explaining 80% of revenue.
3. Top 3 drivers explaining 80% of COGS.
4. 3 OPEX drivers (hiring, variable, fixed).
5. Risk of instability and fallback per driver.
Guardrails:
- Monthly-observable drivers preferred.
- Max 12 drivers.
- Composite drivers must expose sub-components.
Data: [PASTE HERE]
Prompt 2 - Baseline forecast from actuals
Use case: generate a reproducible monthly statistical projection.
Persona: controller.
Level: production-grade.
# ROLE
You are a senior controller at a B2B [SECTOR: SaaS / e-commerce / professional services / industrial] scale-up. You prioritize methodological rigor over exhaustiveness.
# OBJECTIVE
Produce a monthly baseline forecast over 12 months (M+1 to M+12) from 24 months of actuals and the defined drivers. The baseline must be reproducible: another controller, with the same inputs, must reach the same result.
# INPUT DATA
1. 24-month monthly actuals (revenue, COGS, OPEX by line, EBITDA).
2. List of drivers (cf. prompt 1) with monthly historical values.
3. One-off events to exclude, dated (month) and quantified (€).
4. Currency: [EUR / USD / GBP]. Unit: [€ or k€]. Precision: round to [thousand / hundred].
# CLARIFICATION QUESTIONS (mandatory if data missing)
Before producing the forecast, if any of the following is missing, ask for it:
- Currency and reference unit.
- List of one-offs to exclude.
- Constant or evolving perimeter (M&A or divestments in the 24 months).
# EXPECTED METHOD (chain-of-thought required)
Before the final table, reason in a "Method" section (5-10 lines):
1. Method choice per line: linear trend (OLS), multiplicative seasonality (Holt-Winters style), or driver-based with explicit coefficient. Justify in 1 sentence per line.
2. Outlier identification with rule (e.g. deviation > 2 standard deviations from 12-month mean).
3. Seasonality test: if autocorrelation lag-12 > 0.3, use seasonal model; otherwise pure trend.
4. Growth assumption retained per major block: revenue, variable COGS, fixed COGS, OPEX headcount, OPEX non-headcount.
# DELIVERABLE 1 — Monthly forecast table
Markdown format with EXACT columns:
| Month (MMM-YY) | Last actual (M-1) | Forecast | Low band (P10) | High band (P90) | Method | Main driver |
One row per month over 12 months. For each cell:
- Forecast = central value.
- Band = ±1.28 historical std dev (P10-P90).
- Method = short code ("OLS", "HW", "DRV").
Produce this table for each major line: total revenue, COGS, gross margin, OPEX headcount, OPEX non-headcount, EBITDA.
# DELIVERABLE 2 — 5-bullet synthesis
- Forecast revenue growth vs last 12 months (% YoY).
- 3 most uncertain months (Σ P90-P10 spread widest).
- Vs 12-month rolling average: cumulative gap over 12 months.
- 1 alert if trend > +50% MoM on non-headcount OPEX without justified driver.
- 1 recommendation: forecast publishable / to challenge / to redo.
# GUARDRAILS
- If seasonality < 6 months of data: flag "insufficient data for seasonality, default to linear trend method".
- If > 30% outliers detected: STOP, ask user which business event disrupted history.
- Do not invent drivers absent from input data.
- If data is missing, indicate "[MISSING DATA: specify X]" rather than extrapolating.
- No revenue growth > +20% YoY without justified business driver (sales hiring, product launch, pricing increase).
# DATA
[PASTE HERE: 24-month P&L + drivers + one-offs]
Prompt 3 - Build 3 scenarios
Use case: stress-test hypotheses before Comex.
Persona: CFO.
You are CFO. From the baseline and drivers, build 3 scenarios: optimistic, central, pessimistic. Answer in English.
Expected data: 12-month baseline + driver ranges (min, central, max) + macro assumptions.
Deliverables:
1. Side-by-side: revenue, gross margin, EBITDA, cash burn / build over 12 months.
2. 5 explicit, quantified key hypotheses per scenario.
3. Absolute and % gap vs central.
4. Assigned probabilities (sum = 100%).
5. Trigger KPI at M+1 / M+3 to switch scenario.
Guardrails:
- Pessimistic vs optimistic EBITDA spread = 30-50% max.
- No scenario without runway threshold.
- Distinguish internal (controllable) vs external assumptions.
Data: [PASTE HERE]
Prompt 4 - Plan vs actual variance
Use case: explain monthly variances at Comex with clean mathematical decomposition.
Persona: FP&A.
Level: production-grade.
# ROLE
You are senior FP&A at a B2B scale-up. You produce variance analyses directly usable by the CFO in Comex, without rework.
# OBJECTIVE
Produce a plan-vs-actual variance analysis for [MONTH / QUARTER], with mathematical decomposition of revenue variance (volume × price × mix) and margin variance (price, unit cost, product mix). End goal: help the CFO decide on a re-forecast.
# INPUT DATA
1. Monthly or quarterly plan: revenue, COGS, OPEX by category.
2. Monthly or quarterly actuals.
3. Breakdown by offer / segment / channel (volumes, unit prices, mix).
4. Period analysed: [MONTH YYYY-MM or QUARTER YYYY-QX].
5. Materiality threshold: [% of line, e.g. 2% of monthly revenue].
6. Currency: [EUR / USD / GBP]. Unit: [€ / k€]. Sign convention: favourable variance positive if revenue > plan or cost < plan.
# CLARIFICATION QUESTIONS (mandatory if data missing)
- Which decomposition method: Laspeyres (plan base), Paasche (actual base), or chain-linked? Default: Laspeyres if not specified.
- If volumes or unit prices missing by offer: mark "[MISSING DATA]" on the relevant row rather than inventing.
- Confirm sign convention (favourable = positive or negative).
# EXPECTED METHOD (chain-of-thought required)
Before the tables, reason in a "Method" section:
1. Revenue total variance = Actual revenue − Plan revenue. Confirm sign.
2. Laspeyres decomposition:
- Volume variance = (Actual volume − Plan volume) × Plan price
- Price variance = (Actual price − Plan price) × Actual volume
- Mix variance = residual sum after volume and price, by offer.
3. Check that Volume variance + Price variance + Mix variance = Total revenue variance (±0.5% rounding tolerance). If not, recompute.
4. Same approach on gross margin, isolating: price effect, unit cost effect, product mix effect.
# DELIVERABLE 1 — Variance table by P&L line
Markdown format with EXACT columns:
| Line | Plan (k€) | Actual (k€) | Variance (k€) | Variance (%) | Qualification | Material? (Y/N) |
Rows: Total revenue, Variable COGS, Fixed COGS, Gross margin, OPEX headcount, OPEX non-headcount, EBITDA.
Qualification: "Favourable" / "Unfavourable" / "Neutral" (variance < materiality threshold).
# DELIVERABLE 2 — Revenue variance decomposition (volume × price × mix)
Markdown format:
| Effect | Amount (k€) | % of total revenue variance | Comment (1 sentence) |
Rows: Volume effect, Price effect, Mix effect, Total (must equal revenue variance ±0.5%).
# DELIVERABLE 3 — Gross margin variance decomposition
Same format as deliverable 2, with rows: Price effect, Unit cost effect, Product mix effect, Total.
# DELIVERABLE 4 — Top 3 variances to comment at Comex
Format: 3 short paragraphs (40-60 words each), required structure:
- Unfavourable: [line] / [amount] / [probable identifiable cause] / [reversible or structural?]
- Favourable: same structure.
# DELIVERABLE 5 — Re-forecast recommendation
Two sentences maximum:
1. YTD cumulative variance = [X k€, Y%]. Above / below materiality threshold [%].
2. Recommendation: "Immediate re-forecast" / "Re-forecast at quarter-end" / "No re-forecast needed".
# GUARDRAILS
- Do not comment any variance < materiality threshold (mark "Non-material" and move on).
- Volume / price / mix decomposition only if volumes AND unit prices are independently available. Otherwise: "Decomposition not computable, insufficient data".
- Distinguish one-month variance vs structural (3 consecutive months same direction). If 3 consecutive months: flag.
- Do not invent a cause not deducible from data: use "to investigate" rather than a speculative cause.
- Verify sum of effects = total variance ±0.5%. If not, recompute.
# DATA
[PASTE HERE: plan, actuals, breakdown by offer, materiality threshold]
Prompt 5 - Quarterly re-forecast
Use case: update the forecast at quarter-end.
Persona: FP&A Lead.
You are FP&A Lead. Re-forecast the next 4 rolling quarters from YTD actuals and the prior forecast. Answer in English.
Expected data: YTD actuals + prior forecast + business changes (free text) + updated drivers.
Deliverables:
1. Q+1 to Q+4 forecast: revenue, COGS, OPEX, EBITDA, cash.
2. Gap vs prior forecast: value, %, main reason.
3. Revised key hypotheses and expected impact.
4. 200-word Comex memo on the re-forecast.
5. Tracking KPIs for reliability.
Guardrails:
- Don't roll over prior assumptions: justify each.
- > 15% gap on a key line without explanation → flag.
- Maintain P&L / cash consistency.
Data: [PASTE HERE]
Prompt 6 - 13-week cash forecast
Use case: short-term cash visibility and identification of tight weeks.
Persona: treasurer.
Level: production-grade.
# ROLE
You are treasurer at a B2B scale-up. You maintain a weekly direct cash flow over 13 weeks and identify cash tension points before they materialize.
# OBJECTIVE
Produce a weekly direct cash flow forecast over 13 rolling weeks, with automatic identification of critical weeks and 3 sensitivity scenarios. The table must be directly usable in the weekly cash management meeting.
# INPUT DATA
1. Starting cash balance (W0).
2. Customer aging report with age buckets (Not due / 1-30 d / 31-60 d / 61-90 d / > 90 d) and collection probability per bucket.
3. Revenue forecast for upcoming weeks (from prompt 2).
4. Active customer payment plans, dated.
5. AP aging: amount and due date per invoice.
6. Payroll calendar: dates and amounts including social charges.
7. Tax calendar: VAT, corporate tax, local taxes, others.
8. Recurring uninvoiced OPEX (rent, subscriptions) with expected dates.
9. Available credit lines: drawable amount, conditions, cost.
10. Cash safety threshold: [€ amount, e.g. 1 month of OPEX].
11. Currency: [EUR / USD / GBP]. Unit: [€ / k€]. Convention: inflows positive, outflows positive (absolute value).
# CLARIFICATION QUESTIONS (mandatory if data missing)
- Collection probability per age bucket not provided: ask for them. Conservative default if never used: Not due 95%, 1-30 d 85%, 31-60 d 60%, 61-90 d 35%, > 90 d 15%.
- Start date of week 1: Monday of which date?
- Credit lines: amount drawable immediately vs subject to conditions.
# EXPECTED METHOD (chain-of-thought required)
Before the table, reason in a "Method" section:
1. Customer inflow allocation per week: each invoice in aging × collection probability, placed on expected collection week (due date + average bucket lateness).
2. Forecast revenue allocation: apply average DSO to lag forecasted collections.
3. Supplier outflows: contractual due date, no late optimisation.
4. Verification: end-of-week balance N = start-of-week N + inflows N − outflows N. Start week N+1 = end week N. If rounding gap, flag.
# DELIVERABLE 1 — 13-week table (central scenario)
Markdown format with EXACT columns:
| Week | Dates | Start balance | Customer inflows | Other inflows | Total inflows | Supplier outflows | Payroll | Tax | Rec. OPEX | Total outflows | End balance | Margin vs threshold |
One row per week over 13 weeks, date format W01 = Monday-Friday start-end.
Last column: End balance − safety threshold. Negative = alert.
# DELIVERABLE 2 — Critical weeks
List of weeks where (End balance − threshold) < 0 or < 20% of threshold. For each critical week:
- Week N° and dates.
- Shortfall in k€.
- 3 recommended actions ranked by implementation lead time (fast / medium / slow):
- Accelerate customer X collection (estimated impact k€).
- Draw available credit line (estimated impact).
- Negotiate outflow delay (which one).
# DELIVERABLE 3 — 3 sensitivity scenarios
Comparative table:
| Scenario | Assumption | Week 13 balance (k€) | # critical weeks | Runway (weeks) |
Scenarios:
- Central: baseline assumptions above.
- Customer delay +10 days: apply +10 d on all customer inflows.
- Major contract delayed: remove largest customer inflow of the period (to identify).
# DELIVERABLE 4 — CFO synthesis (100 words max)
Four points:
- Runway in weeks (central scenario).
- Trough and date.
- Headroom = gap between trough and threshold.
- Top recommendation: action to launch this week.
# GUARDRAILS
- End balance week N = start balance N + inflows N − outflows N. Auto-verify each row.
- No variable outflow shown < 10% of weekly average (likely missing, ask confirmation).
- Do not invent customer payments: if an invoice > 90 d is marked dispute in data, collection probability = 0% by default.
- If week 13 balance < 0 in central scenario: mark "URGENT - financing required" at top of deliverable.
- Always state sign convention at top of table (inflows +, outflows +).
# DATA
[PASTE HERE: W0 cash, customer aging, revenue forecast, AP aging, payroll / tax calendars, recurring OPEX, credit lines, cash threshold]
Prompt 7 - Sensitivity and stress test
Use case: prep for the board meeting.
Persona: CFO.
You are CFO. Run a stress test of the central forecast over 4 quarters. Answer in English.
Expected data: central forecast, key drivers, critical thresholds (min runway, covenants).
Deliverables:
1. Sensitivity table: EBITDA and cash impact at ±10% and ±20% on each top-5 driver.
2. 3 stress scenarios:
- Demand: revenue −20% over 2 quarters.
- Payment: DSO +15 days.
- Cost: variable cost +15%.
3. Per stress: runway impact, covenants, available levers.
4. Alert thresholds for contingency plan.
5. 250-word board summary.
Guardrails:
- No > 2 simultaneous stresses without flag.
- Quantify cash impact, not only EBITDA.
- Separate reversible from irreversible levers.
Data: [PASTE HERE]
Prompt 8 - Comex narrative on the re-forecast
Use case: quarterly framing memo directly usable in Comex.
Persona: CFO.
Level: production-grade.
# ROLE
You are CFO of a B2B scale-up. You write factual Comex memos, no marketing jargon, every figure traceable to source. You write for busy executives who want to understand the situation in 3 minutes.
# OBJECTIVE
Draft the quarterly re-forecast Comex memo: 500 words max, English, factual tone, structure imposed below. This memo must be readable in 3 minutes and enable the Comex to make 3 concrete decisions.
# INPUT DATA
1. YTD actuals (cumulative since fiscal year start): revenue, EBITDA, cash.
2. Prior forecast (quarter N-1) by key line.
3. New forecast (from prompt 5) by key line.
4. Initial annual approved budget.
5. Revised drivers with old / new values and short justification.
6. 3 scenarios (from prompt 3): optimistic, central, pessimistic.
7. Perimeter: [entity / consolidated group / segment X].
8. Currency and unit: [EUR / USD / GBP], [€ / k€ / M€].
# CLARIFICATION QUESTIONS (mandatory if data missing)
- Target Comex date and expected perimeter (group / entity).
- Is there a sensitive topic to treat carefully (layoff plan, refinancing, loss of major customer)? If yes, flag the expected tone.
- Probability assigned to each scenario: has it changed since the previous Comex?
# IMPOSED DELIVERABLE STRUCTURE
The memo must follow EXACTLY this structure, in this order, with these headings:
## Summary (3 sentences max, ~60 words)
Sentence 1: new full-year revenue trajectory vs initial budget (% and k€).
Sentence 2: new full-year EBITDA trajectory vs initial budget.
Sentence 3: cash trajectory, anticipated trough and runway.
## Why this re-forecast (3 main reasons, ~120 words)
One reason = one paragraph of 2-3 sentences with:
- Driver concerned (e.g. churn rate).
- Old value and new (e.g. 1.2%/month → 1.8%/month).
- Quantified impact on annual EBITDA (k€).
List reasons by decreasing impact.
## Comparison table (~80 words commentary + table)
Mandatory markdown table:
| Indicator | Initial budget | Prior forecast | New forecast | Variance vs budget |
Rows: Revenue, Gross margin, EBITDA, End-period cash, Runway.
After the table: 2 sentences to highlight the 2 most important shifts.
## Decisions expected from Comex (3 decisions, ~120 words)
One decision = one paragraph with:
- Precise question to Comex (phrasing: "Comex is invited to decide on...").
- Minimum 2 quantified options (option A vs option B with EBITDA / cash impact each).
- Recommended decision deadline.
## Risks and mitigation (~100 words)
3 risks maximum. For each:
- Description (1 sentence).
- Probability (low / medium / high).
- Impact (k€ if materialized).
- Proposed mitigation (1 sentence).
## Conclusion (1 sentence)
"Comex is invited to validate [main decision] before [date]."
# GUARDRAILS
- No marketing adjectives: banned "significant", "impressive", "remarkable", "strategic" (used more than once), "transformative".
- Each cited figure must be traceable to source: prompt 5 (re-forecast), initial budget, or YTD actuals. If a figure comes from elsewhere, mention in parentheses.
- No sentence over 25 words. If you have one, split it.
- No bullets in the summary or conclusion: pure prose.
- Strict total < 500 words. If you exceed, compress "Risks" first.
- If a re-forecast figure is inconsistent with a provided scenario (e.g. new EBITDA > optimistic), flag at top "[INCONSISTENCY DETECTED: specify X]" rather than producing the memo.
- Never end with "in conclusion" or "to summarize": the conclusion is already titled.
# DATA
[PASTE HERE: YTD actuals, prior forecast, new forecast, initial budget, revised drivers, 3 scenarios]
Prompt 9 - SaaS forecast: MRR, churn, expansion
Use case: decomposed SaaS revenue forecast.
Persona: SaaS FP&A.
You are SaaS FP&A. Build a monthly MRR / ARR forecast over 12 months, decomposed into new, expansion, contraction, churn. Answer in English.
Expected data:
1. Starting MRR by acquisition cohort.
2. Historical rates: new MRR/month, expansion %, contraction %, churn % (logo and $).
3. Weighted pipeline.
4. Sales capacity.
Deliverables:
1. MRR table: starting + new - churn + expansion - contraction = ending.
2. Period-end ARR + YoY growth.
3. Forecast NRR on existing cohorts.
4. Sensitivity: ±5pts churn, ±10% new business.
5. Bottleneck (capacity, conversion, churn).
Guardrails:
- No growth beyond sales capacity.
- Separate logo vs $ churn.
- NRR excludes new business.
Data: [PASTE HERE]
Prompt 10 - Forecast quality audit
Use case: verify robustness before publication.
Persona: senior CFO / controller.
You are senior CFO. Audit the forecast before publication. Answer in English, checklist.
Expected data: full forecast + hypotheses + drivers + YTD variance.
Deliverables:
1. Control table: P&L ↔ cash flow ↔ balance sheet consistency, internal sums, hypothesis traceability.
2. Top 5 quality risks.
3. Undocumented hypotheses to challenge.
4. Sanity tests: gross margin, growth rate, EBITDA % — out of plausible range?
5. Recommendation: ready, rework, redo.
Guardrails:
- Do not validate without 3 scenarios.
- Flag any P&L / cash inconsistency > [THRESHOLD] €.
- Document every red flag.
Best practices for FP&A prompting
- Choose drivers before the model — 70% of quality comes from there.
- Standardise the template — same format for every forecast.
- Version every re-forecast (snapshot, model, author, date).
- Top-down first, bottom-up validation second.
- Wire to EPM (Pigment, Anaplan, Cube, Mosaic, Drivetrain). Isolated prompt = stale data.
Limits and watchpoints
- Linear-trend bias: LLMs over-extrapolate. Always challenge growth assumptions.
- Math hallucinations: cross-check aggregations at least once.
- Confidentiality: forecasts are strategic. Private LLM only (Claude, ChatGPT Enterprise), never consumer versions.
- Optimism bias: commercial drivers are usually overstated. AI mirrors what you feed it.
- No causal modelling: AI doesn't know the Q3 recession killed growth. You explain it.
FAQ
What is a quarterly rolling forecast?
A quarterly rolling forecast is a financial projection that always looks 4 quarters ahead, refreshed every quarter from the latest actuals. Unlike the annual budget, it is updated continuously, making it a pilot tool rather than a frozen budgeting exercise.
Monthly or quarterly cadence?
Monthly for the 13-week cash re-forecast, quarterly for the P&L, annual for the strategic budget. That combination wins in 2026 for B2B scale-ups.
EPM tool required?
Not below 100 FTEs. Above, yes (Pigment, Drivetrain, Cube, Mosaic, Anaplan). LLMs alone don't replace the collaborative modelling layer.
Which AI model should I use?
Claude (Anthropic) excels at P&L / cash consistency and hypothesis documentation. ChatGPT (with Code Interpreter) is strong at variance math. Gemini handles long historical tables well. Test in parallel on a real case before standardising.
What ROI should I measure?
Two actionable metrics: −50% forecast cycle time and −30% plan-vs-actual variance. Per the RGP survey (200 US CFOs, November 2025), only 14% of CFOs report measured AI ROI, and the leading blocker is data quality: 35% cite data trust as the top barrier.
Will AI replace the controller?
No. AI eats data-entry and consolidation time (around 60% of the role today) and shifts work to challenging hypotheses (40% today, up to 80-90% tomorrow). The role evolves, it doesn't disappear.
Should I start with prompts or with data infrastructure?
Infrastructure. The Bain CFO Survey 2026 shows that only 12% of finance functions have deployed ML in FP&A at scale — not for lack of prompts, but because of inconsistent data between ERP, CRM and billing.
How do I avoid math hallucinations?
Three reflexes: (1) ask the LLM to justify each aggregation, (2) manually verify at least 3 key cells of the generated table, (3) run the forecast through prompt 10 (quality audit) before publication.
Conclusion
The quarterly rolling forecast is no longer a luxury reserved for Anaplan-equipped mid-markets. With 10 well-framed prompts, a 50-500 person scale-up can sustain a clean, documented, audit-ready cycle in 5 days instead of 3 weeks. The condition is not the algorithm, it is data infrastructure and hypothesis discipline.