Choreography: scenarios and sensitivity with AI
The choreography of using AI to build scenarios and test sensitivity in finance: the machine speeds up the model, but the assumption is yours and every number it spits out is suspect until you audit the math.
You open AI to decide whether to launch the expansion now or wait for next quarter, and in minutes you have three result scenarios ready for tomorrow's meeting. The numbers add up, the chart is elegant, everything looks ready to become a decision. The problem shows up when you look closely: the 8 percent monthly growth baked into all three scenarios didn't come from any of your data, and the "customer acquisition cost" line that dropped on its own between one scenario and the next, nobody checked. AI builds the whole tree in one request; the assumption holding up every branch is still yours, and it's what decides whether the meeting starts from a real number or a well-dressed guess.
A manager asked AI for a twelve-month cash flow projection, with an optimistic, base, and pessimistic scenario. AI handed back a complete model in seconds. In the pessimistic scenario, it assumed default rates would rise to 3 percent and that the interest rate would stay stable. Pretty. Except the company was already running 7 percent default in the current month, and AI had invented the 3 percent figure because "it seemed reasonable." The pessimistic scenario was, in truth, optimism in disguise.
You're opening two positions and ask AI to project the cost and time to hire in three scenarios. It hands back a polished spreadsheet: time to fill the role, cost per hire, offer acceptance rate. In the base scenario, it assumed 1 in every 4 finalist candidates accepts the offer. Except your actual track record is 1 in every 8, because your salary band is below market. The acceptance assumption came out of AI's head, not your recruiting funnel. The base scenario was born optimistic, and the hiring timeline you promised the manager is off by half.
You need to decide whether to launch a feature early or wait for next quarter, and you ask AI for three impact scenarios: adoption, retention, incremental revenue. It builds everything fast, with a pretty adoption curve. In the optimistic scenario, it baked in 40 percent user adoption in the first four weeks. Where did that 40 percent come from? AI didn't look at your user base, doesn't have the numbers from your last similar feature that adopted at 9 percent. It guessed a number that "seemed reasonable" for some generic product. Your entire roadmap prioritization might be resting on that guess.
You're closing the quarter's forecast and ask AI to project the pipeline in three closing scenarios. In seconds a clean table comes back: weighted value by stage, conversion rate, expected revenue. In the base scenario, it applied a 30 percent proposal-to-closed conversion rate. Except your team's real conversion rate, in the CRM, is 18 percent. AI invented the 30 percent because it's a market average, not yours. You were about to take that forecast to the sales director as if it were your funnel, when it was actually nobody's funnel.
You're deciding whether to bring a shipping process in-house or keep the logistics operator, and ask AI for three cost and SLA scenarios. It hands back a complete model: cost per order, delivery time, damage rate. In the pessimistic scenario, it assumed 2 percent damage in transit. Pretty, except your current operation runs at 6 percent damage and nobody fed that data in. AI filled the gap with a number that "seemed reasonable" for a healthy operation, not for yours. The scenario that was supposed to show you the worst case is dressing up exactly the problem you have today.
You need to size the risk of a regulatory exposure and ask AI for three financial impact scenarios for a possible sanction. It builds the structure on the spot: probability of a citation, fine amount, remediation cost. In the base scenario, it assumed a 10 percent probability of a citation. Where did that come from? AI doesn't know your history with the regulator, hasn't seen the three notices you've already received this year. It plugged in a number that looks moderate for a generic case. In compliance, this scenario isn't an exercise: it goes into the risk committee's minutes and becomes a decision to provision funds or not. Invented assumption, contaminated decision.
You're deciding between migrating to new infrastructure now or in six months, and ask AI for three cost and availability scenarios. It spits out a well-built model: cloud cost, engineering hours, expected downtime. In the optimistic scenario, it assumed a 99.99 percent SLA on the migration and a cloud cost that drops 30 percent. Where does that 30 percent come from? AI doesn't have your contracts, doesn't know your real traffic volume, hasn't seen your last deploy that took the system down for two hours. It estimated with numbers from an ideal case. The architecture decision is resting on a fantasy downtime figure.
You're deciding whether to rebuild the checkout flow now or keep iterating on the current one, and ask AI for three conversion impact scenarios. It hands back an elegant spreadsheet: completion rate, drop-off by step, revenue gain. In the optimistic scenario, it assumed the new flow reduces drop-off from 60 to 20 percent. Where did that 20 percent come from? AI didn't run a test with your users, hasn't seen your last usability study, doesn't know the real friction in your journey. It grabbed a number that "seems reasonable" for some pretty generic checkout. The entire redesign decision is resting on a gain nobody measured.
You're deciding whether to enter a new market now or wait two years, and you ask AI for three market sizing and penetration scenarios. It builds an elegant spreadsheet: TAM, SAM, SOM, adoption curve. In the base scenario, it assumed a 12 percent conversion rate in the first two years. Where did that 12 percent come from? AI doesn't have your conversion data in similar markets, it never saw your last launch that converted at 4 percent. It grabbed a number that "seems reasonable" for some generic market. The entire entry decision, the one the board is about to rubber-stamp, is resting on a rate nobody at your company ever validated.
Think with me for a second. Building a financial scenario has always been the tedious part: opening the spreadsheet, duplicating tabs, changing one cell here, redoing a formula there, and in the end you had one scenario just because there was no time for the other two. AI really changes this game. It builds the structure, generates the variations, and even writes the explanation while you get your coffee. The problem is that speed deceives. It hands you ten scenarios in the time of one, but if the assumption is off, or if it invents a number along the way, you didn't speed anything up. You just got it wrong faster, with a pretty chart on top.
The core idea of this lesson. AI is the best scenario builder you've ever had, but it doesn't own anything. In financial scenarios, the rule is double and non-negotiable: the assumption is YOURS, written down and defensible, because it's reading the world, not arithmetic; and every number AI spits out is suspect until you audit the math. The machine explores; you decide what the scenario means and sign off on it.
01The choreography: who does what
A scenario isn't guesswork, it's a dance with defined steps. And in this dance you and AI have different roles that can't mix. When they mix, the scenario collapses.
AI is good at three things, and only these three. It structures the model: builds the spreadsheet's skeleton, organizes the revenue, cost, margin, and cash flow lines. It generates the variations: from an assumption you give it, it calculates the optimistic, the base, and the pessimistic without you duplicating a single tab. And it writes the explanation: turns the dry table into a paragraph your partner understands. That's the manual labor of modeling, and the machine does in minutes what used to take you the afternoon.
What's still yours, and gets more expensive precisely because the easy part vanished, is the assumption. What growth is reasonable for your market right now? Is the rate going up or staying stable? Does the acquisition cost follow the track record or is there a reason for it to change? That's not arithmetic, it's reading the world. It's you looking at your sector, your customers, the macro picture, and making an informed bet. AI doesn't have your context, and when it doesn't, it invents one that "seems reasonable." Hold onto that word, because it's the source of nearly every disaster in this lesson.
02The scenario tree: optimistic, base, pessimistic
A scenario isn't one number, it's three paths from the same point today. Think of a road that splits. The starting point is your current reality, known and audited. From there, three branches emerge.
The base scenario is your most honest bet: what will probably happen if nothing extraordinary occurs. The optimistic is the branch above: things go well, growth arrives, cost behaves. The pessimistic is the branch below: sales slow down, cost rises, something jams. The point of building all three isn't predicting the future, it's drawing the band within which the future will probably fall. If even your pessimistic scenario keeps you standing, you decide with ease. If the base already squeezes you, you decide with care.
AI builds this entire tree in one request. But notice where the danger lives: each branch depends on an assumption you needed to have given on purpose. If you let AI "choose" the optimistic scenario's growth and the pessimistic scenario's default rate, it will fill in with numbers that seem plausible and aren't yours. The tree ends up beautiful, and it lies in all three branches at once.
03Sensitivity: which variable moves the result
Here lies the question that separates whoever's just playing with a spreadsheet from whoever decides with method. Every scenario has several variables: growth, interest rate, acquisition cost, collection term, default rate. But they don't carry the same weight. Some you can move a lot and the result barely budges. Others you nudge two points and the result flips upside down. Sensitivity analysis is finding out which are the latter.
The choreography is simple and AI speeds this part up enormously. You fix everything at the base scenario and ask: "what if the rate goes up 2 points, keeping everything else the same? What if the acquisition cost goes up 20 percent? What if the collection term stretches 15 days?" One at a time. AI recalculates the result in each case and you see, in black and white, which variable is the lever and which is a detail. The lever is where you need a solid assumption and a plan B. The detail you can estimate roughly without losing sleep.
Why does this matter so much? Because focus is a scarce resource. If the interest rate is the variable that decides your cash flow, that's where you spend your energy reading the world, that's the one you need a defensible number for. Spending three hours discussing a variable that moves the result by 1 percent is waste. Sensitivity tells you where to look. But watch out for the same danger as always: when AI recalculates each case, it can flip a sign, add wrong, or cite a calculation it never did. The recalculation is its job; the checking is yours.
04The central danger: AI invents numbers with the look of certainty
Now the point that names the risk, and that you need to carry with you. AI doesn't fail like a broken calculator, which jams and warns you. It fails with confidence. It spits out a round, well-formatted number, inside a confident sentence, and that number simply didn't come from any calculation. It cited a total it never summed. It flipped a sign and the expense became revenue. It assumed a 3 percent default rate because "it seemed reasonable," while yours was at 7. All of it with the exact same look of certainty as the number sitting right next to it that was correct.
This is where the financial frame becomes non-negotiable. In a marketing copy, a wrong number is a slip. In a financial scenario, a wrong number becomes a wrong decision: you hire when you should hold, or you hold when you should hire. The scenario isn't decoration, it feeds a real choice, with real money. That's why the rule is hard and simple: every number AI spits out is suspect until you audit the math. Not "I trust it because it seemed right." You open it, redo the sum on paper or in your own cell, check the sign, confirm the assumption it used is the one you gave. If it doesn't match, the whole scenario is contaminated.
This connects directly to lesson 4.4, on execution governance. There, the principle was "Claude did it" isn't an excuse: it went out under your name, you sign off on it. In a scenario, it's the same thing squared, because the scenario becomes a decision. And it also connects with the map of the Three Moves: before acting on a scenario, ask how reversible the decision is. Hiring two salespeople based on a scenario and discovering the number was invented is expensive to undo. The less reversible the decision, the more expensive your audit becomes, and the firmer your assumption needs to be. AI lets you explore ten scenarios in the time of one. Wonderful. But one wrong number among those ten, taken seriously, is worth all ten put together with the sign flipped.
Do it now
Take a real financial decision on your desk right now (your real task works well): hiring, investing, changing a price, prepaying something. Let's run the full choreography in four steps.
- YOUR ASSUMPTION: before opening AI, write by hand the three central assumptions (e.g.: monthly growth, rate, acquisition cost) and, next to each one, ONE defensible sentence for why that number is reasonable for YOUR world. If you can't defend it, the assumption isn't yours yet.
- TREE: ask AI to build the model with optimistic, base, and pessimistic using EXACTLY your assumptions (don't let it choose any number). Be explicit: "use these numbers, don't invent others."
- SENSITIVITY: fix the base and ask, one at a time: "what if the rate goes up 2 points? What if the cost goes up 20 percent? What if the term stretches 15 days?" Note which variable moves the result the most. That's your lever.
- AUDIT: choose the 3 most important numbers in the scenario and redo the math yourself, on paper or in a cell. Check the sign. Check whether the assumption used was the one you gave. Only after that does the scenario get permission to become a decision.
If you got stuck on step 1, great: that was exactly the expensive part of the work, and now you know it.
Practice
1. In the choreography of building scenarios with AI, what stays YOURS and gets more expensive precisely because the manual labor vanished?
2. You fixed the base scenario and tested: rate +2 points flips the result upside down, while acquisition cost +20 percent barely moves it. What does this sensitivity analysis tell you?
3. AI handed back a beautiful pessimistic scenario, but assumed a 3 percent default rate when yours is already at 7 percent. What's the correct rule for handling this?
Fair enough? The message wraps up simply. AI is the best scenario builder you've ever had: it lets you explore ten futures in the time it used to take to look at one. But a builder isn't an owner. The assumption is reading the world and it stays yours, written down and defensible. And every number the machine spits out is suspect until you redo the math, because in finance a wrong number isn't a slip, it's a wrong decision with real money. Use the speed to explore more. Use your own head to choose the assumption and audit the result. That pairing is what turns AI from a loose risk into a decision-making lever.
For the board
On the assumptionit is a reading of the world, not arithmetic. When the AI lacks your context, it invents one that sounds reasonable.
On sensitivityit shows you where to look. The variable that turns the result upside down deserves the most defensible assumption.
On the riskin finance the scenario becomes a decision. An invented number becomes a wrong choice.
Thanks for the feedback. It helps sharpen the next lesson.