AI does not forecast - you do, out loud
By now you can brief an AI like a spec (b2) and run a month-end close copilot (b3). Forecasting feels like the natural next ask: "here is our history, predict next quarter." Resist it. A chat AI typing a revenue number is not forecasting - it is guessing plausibly, and it cannot show you the method. Tonight we split the work honestly: the numbers get computed with real methods in a spreadsheet you control, and AI does the thinking-and-writing around them - finding the drivers, drafting the assumptions, stress-testing the scenarios, and turning the results into a narrative your CFO can read.
AI's real role in forecasting 6 min live
The single most important idea tonight: AI is not a black-box predictor. The temptation is to paste a history and ask for "the forecast". What comes back looks confident and is untraceable - you cannot check the method, so you cannot sign it. Instead, keep the statistical math in the spreadsheet and point AI at the four things it is genuinely good at.
LiveWhy free-text "forecasts" are extrapolation, not forecasting3 min▶
When you paste a history into a chat window and ask for next quarter, the model is pattern-matching text - it eyeballs the trend and continues it. That is extrapolation dressed as analysis. It has no seasonality handling, no confidence interval, no method you can name to your audit committee. Worse, it will sound certain.
- Do the numbers with real methods. Trend, moving average, regression, driver math - in the spreadsheet, where formulas are visible and checkable.
- Use AI for structure and story. Which drivers matter, how to phrase an assumption, what a bear case should stress, how to narrate the result.
- The line to remember. If AI hands you a number you cannot trace to a formula, it is a guess. Guesses do not go in a board pack.
Self-studyWhat "good at" actually means here2 min read▶
AI earns its place at four points in the pipeline - and only those four. Everywhere else, your sheet and your judgement lead.
- Identify drivers - read a history plus business context and name the handful of things that actually move the line.
- Articulate assumptions - turn a fuzzy "growth slows a bit" into a labelled, numeric, auditable statement.
- Generate and stress scenarios - build base/bull/bear and then argue against each one to find where it breaks.
- Write the narrative - convert your computed results into prose a CFO reads in thirty seconds.
Notice what is missing: computing the forecast. That stays with you. GROUND it, make it SHOW WORK, VERIFY the math - the same method from b1, now applied to the future instead of the past.
Driver-based forecasting with AI 6 min live
Driver-based forecasting means you build the number up from the things that cause it - units times price, headcount times cost, traffic times conversion - not by drawing a line through last year and extending it. AI is a strong partner for finding drivers and drafting assumptions; the sheet still owns every calculation.
LiveWho does what - the split, task by task3 min▶
| Task | AI assists with... | Human + sheet own... |
|---|---|---|
| Surface drivers | Reads history + business context, proposes the drivers that move the line | Confirm they are real, drop spurious ones, decide the driver tree |
| Draft assumptions | Turns vague expectations into labelled, numeric, testable statements | Set the actual values from evidence; own every number |
| Run sensitivity / what-if | Suggests which assumptions to flex and by how much | Compute the table with formulas; read which driver dominates |
| Rolling re-forecast | Flags what changed vs last cycle, drafts the "why" note | Re-run the model on new actuals; verify the deltas |
| Narrative | Writes the story from your verified results | Check it matches the numbers; sign it |
Rolling re-forecast is where the AI-assist compounds: every month the model is the same, only the actuals change, and AI drafts the "here is what moved and why" note in seconds so you spend your time judging, not typing.
Self-studyFrom history to a driver tree2 min read▶
A driver tree makes your forecast arguable. Instead of "revenue grows 8%", you say "revenue = active accounts x average revenue per account", and now each half has its own drivers and its own assumptions - each one visible, each one challengeable.
The move that unlocks it: ask AI to decompose, not to predict.
You now have a structure to fill with real assumptions and compute in the sheet - exactly the shape we build in Demo 1.
Scenario analysis 3 min live
One forecast is a guess about the future; three scenarios is a conversation about it. Base, bull, and bear turn "the number" into a range with reasons, and that is what boards actually want. AI helps you articulate each case and then stress it - but every assumption goes on the record.
LiveBase, bull, bear - and stress every one2 min▶
- Base. Your best single view - the assumptions you would defend as most likely. Everything else flexes from here.
- Bull. What has to go right, stated as specific assumption changes, not vibes: "conversion +2pts, churn -1pt".
- Bear. What could go wrong, equally specific. The bear case is a risk tool, not pessimism theatre.
- Stress it. Have AI argue against each scenario - which assumption is doing all the work, what single break flips the story. This is where AI is genuinely useful and where humans get lazy.
- Document everything. Every assumption, its value, and its source, in a table. If it is not written down, it did not happen - and it will not survive audit.
The bull case that hid one heroic assumption. A team presented a confident bull forecast to the board; the whole upside rested on a single line - "win rate jumps from 22% to 40%" - buried in the model. Asking AI to "tell me which one assumption this scenario depends on most" surfaced it in seconds. They kept the scenario but re-labelled it "aggressive - contingent on win-rate step change", and the board conversation got honest. Stress-testing is cheap insurance.
Drivers in, structured assumption set + 3 scenarios out ★ 12 min · everyone builds
We give AI the drivers plus a short supplied history and ask for a structured assumption set and three scenarios - as structure, not as final numbers. Then you plug the numbers into your sheet, compute the forecast, and verify. AI shaped the frame; you own the math.
Open ChatGPT or Claude, plus a blank spreadsheet. The history below is synthetic - safe for any tool (rail one: public or synthetic data only).
Run the prompt below. You get a driver-based assumption set (each assumption labelled with a value and a rationale) and base / bull / bear variants - as a structured table, not a hand-wavy paragraph.
Move it into the sheet. Put drivers and assumptions in cells, then write the formulas yourself: revenue = accounts x ARPA, and so on. The AI's suggested values are inputs you accept or overwrite - the formulas are yours.
Compute and verify. Check each output ties to a formula you can read. If a number cannot be traced, delete it. This is the SHOW WORK -> VERIFY half of the method.
Stress once. Ask AI which single assumption each scenario depends on most, and sanity-check that against your sheet. Adjust labels if a scenario is riskier than it looked.
Sensitivity table in the sheet, then an AI-drafted narrative ★ 8 min · build your own
Now we close the loop: build a small sensitivity table in your spreadsheet, then hand the computed results to AI and ask it to write the scenario narrative. The numbers are yours and verified; AI turns them into the story.
In your sheet, take the base case from Demo 1 and flex two assumptions - say churn and ARPA growth - across a small grid. A 3x3 table of resulting revenue is plenty. Write the formulas yourself.
Read the table: which assumption moves revenue more? That is your dominant driver, and it is the spine of the narrative.
Hand the results to AI using the prompt below - paste your computed base/bull/bear revenue and the dominant driver you found - and ask for a short scenario narrative.
Verify the story matches the math. Read the draft against your table. If it claims a swing your numbers do not show, fix the prose - never the numbers to fit the prose.
Save both the sensitivity template and the narrative prompt into your prompt library from b2. Rolling re-forecast next month reuses both.
The forecast that finally got read. An FP&A lead used to attach a dense scenario grid to the board pack that nobody opened. She kept the grid but had AI draft a 150-word narrative from her verified numbers - base, range, the one driver to watch. The board engaged with the story and asked better questions. The model did not change; the packaging did. AI was the ghostwriter, she was the author and the accountable one.
Try it yourself - this week ◐ 30-45 min total
- Take one line of your own P&L (public or synthetic if you are in a consumer tool) and ask AI for its driver tree - not a forecast. Confirm or prune the drivers yourself.
- Build a base-case assumption set as a labelled table, then create bull and bear by changing only the assumption values. Document each one and its source.
- Build a 3x3 sensitivity table in your sheet flexing your two most uncertain assumptions. Note which one dominates - that is your headline risk.
- Feed your verified results to AI and get a scenario narrative under 180 words. Check every claim against your table before you would ever send it.
- Optional: run the stress prompt - "which one assumption does each scenario depend on most" - and re-label any scenario that turns out riskier than its name suggested.
Official sources covered
This session teaches from public FP&A and scenario-analysis curricula. Certificates, graded assessments, and paid tooling stay with the official programs - noted honestly. This session covers:
Three questions before you go 🎯 ◐ 90 seconds
1 · In driver-based forecasting with AI, who computes the actual forecast number?
Keep the statistical math in the sheet where formulas are visible and checkable. AI identifies drivers, articulates assumptions, stresses scenarios, and writes the narrative - it never becomes the calculator-of-record.
2 · What does "driver-based" forecasting mean?
Driver-based means you decompose the line into its causes and forecast each driver, so every assumption is visible and arguable - not one opaque growth rate.
3 · You have a confident bull scenario. What is the discipline before you present it?
Scenarios need every assumption on the record for audit, and each one stress-tested. Surfacing the single heroic assumption keeps the board conversation honest.