Scenario Modeling and Sensitivity Analysis
Deliberate Academy Editorial Team
Reviewed for accuracy and professional relevance
You're 5 lessons in — don't lose your progress.
Sign up free to save where you are and earn a verified certificate when you pass.
- Apply AI to scale scenario generation and sensitivity table construction well beyond what manual spreadsheet modeling allows
- Structure a sensitivity analysis by identifying which variables to flex, by what range, and why — using AI as a structuring aid rather than an assumption source
- Explain the overconfidence failure mode where AI presents a point estimate from a scenario model as more certain than the underlying assumptions support
- Assess when a wide scenario range communicates more useful information to decision-makers than a single AI-generated best estimate
Scenario modeling and sensitivity analysis exist because the future is uncertain and a single forecast number hides that uncertainty from the people who have to make decisions based on it. AI's biggest contribution here is scale: generating and computing many more scenario variants than manual spreadsheet work makes practical. That scale is genuinely valuable — but it does not change the fact that a scenario model is only as good as the assumptions feeding it, and AI's fluency in presenting scenario output can make weak assumptions look more certain than they are.
Scaling Scenario Generation with AI
Manually building even three or four scenarios — a base case, an upside, and a downside — for a complex model is time-consuming, because each scenario requires flexing multiple interdependent assumptions and recalculating the downstream effects consistently. AI-assisted planning platforms change the economics of this significantly. Anaplan's scenario modeling capabilities, including its PlanIQ predictive forecasting layer, and Pigment's scenario comparison tools are both built to compute dozens of scenario variants from a defined set of assumption ranges in the time it previously took to build three by hand.
This matters most for questions with genuine, high-consequence uncertainty — a cost inflation range that could move margin by several points, a hiring plan that depends on an uncertain fundraising timeline, a market entry decision with a wide range of plausible adoption curves. Where manual constraints previously forced a team down to three scenarios out of necessity, AI-assisted scenario generation allows the team to actually explore the range of plausible outcomes.
Use AI to expand the number of scenarios you can compute, but keep the number of scenarios you actually present to decision-makers small — three to five, clearly labeled and clearly differentiated. A stakeholder presented with twenty scenario variants will not extract more insight than one presented with four well-chosen ones; the value of AI's scale advantage is in exploring the space thoroughly before you decide, not in delivering the entire space to your audience.
Structuring Sensitivity Tables
A sensitivity table shows how a single output — margin, cash balance, headcount cost — responds to a range of values for one or two key input variables. AI can help structure the table itself: given the model's key variables and a description of which ones carry the most uncertainty, AI can suggest which variables belong in a sensitivity analysis and a reasonable range to flex them across, based on the volatility described.
The variable selection and range itself is a genuine judgment call that should stay with the analyst and the business stakeholders who understand the actual uncertainty involved. AI's suggestion is useful as a starting point precisely because building a sensitivity table from scratch involves some decision paralysis about which variables matter most — but the final range should reflect what the business genuinely believes is plausible, not a generic percentage AI has defaulted to. This matters even more when the variable being flexed doubles as a financial-statement estimate rather than a purely internal planning input — a cost-inflation assumption behind an impairment test, for example. IAS 1 requires disclosure of the key assumptions and sources of estimation uncertainty behind material estimates, and SEC guidance expects US public companies to discuss critical accounting estimates in their MD&A, so a sensitivity range that is genuinely business-judged rather than an unreviewed AI default is not only better FP&A practice — it is the standard financial reporting already expects.
An FP&A analyst asks an AI tool to suggest which variables should be included in a sensitivity analysis for a new product launch model and what range to flex them across. The AI suggests flexing unit price by plus or minus 10% as a default range. What should the analyst do with this suggestion?
Select one answer.
The Overconfidence Failure Mode: Point Estimates Presented as Certain
The specific risk in AI-assisted scenario modeling is not that AI generates bad scenarios — it is that AI's fluent, well-formatted output can make a single point estimate read as more certain than the underlying assumptions justify. A model that generates a "most likely" scenario alongside upside and downside cases can produce prose around the most likely case that reads with unwarranted confidence, especially when the assumptions behind it have not been stress-tested as rigorously as the number's polish suggests.
Watch for AI-generated scenario narratives that describe a single point estimate using confident, definitive language — "revenue will reach," "margin is expected to be" — without disclosing the assumption range behind it. A scenario output is a conditional statement: given these specific assumptions, this is the result. If the narrative strips out the conditionality and presents the number as a prediction, it has overstated the certainty the underlying model actually supports, regardless of how well-written the sentence is.
Widening a cost-inflation scenario range before a resourcing decision
Context
An FP&A manager at a mid-market logistics company was asked to model the impact of fuel and labor cost inflation on next year's margin ahead of a fleet expansion decision. Using Pigment's scenario comparison tools, the team generated fifteen scenario variants across combinations of fuel price movement and labor cost inflation in under a day, a scope that would have taken roughly a week to model manually with the same rigor.
Action
Reviewing the AI-generated scenario summary, the manager noticed the narrative described the 'most likely' scenario's margin impact in confident, singular language — stating the expected margin compression as if it were a settled figure — despite the underlying model showing a wide plausible range depending on how fuel prices moved over the following two quarters. The manager rewrote the summary to present the full plausible range (a 1.8 to 3.4 percentage point margin impact) rather than a single point figure, with the assumptions behind each end of the range stated explicitly.
Outcome
The CFO used the wider range in the fleet expansion discussion with the board, explicitly noting that the decision needed to be resilient across the full range rather than optimized for the single point estimate the AI narrative had originally implied. The expansion plan was adjusted to include a contingency trigger tied to fuel price movement rather than committing to a fixed capital plan based on the midpoint estimate — a decision the FP&A manager attributed directly to having caught and corrected the overconfident framing before it reached the board.
What is the overconfidence failure mode this lesson describes in AI-assisted scenario modeling?
Select one answer.
Exercise
Your Task
Take a current planning question in your organization that involves genuine uncertainty — a cost input, a demand assumption, a hiring or investment decision. Use AI to help you structure a sensitivity table: identify two or three key variables, and draft a plausible range for each based on your own judgment of the uncertainty involved (not a generic default). Then ask AI to draft a short narrative summarizing the scenario output, and review that narrative specifically for language that presents a single point estimate as more certain than the range actually supports. Rewrite any sentence that fails this check.
Success looks like
- The sensitivity table variables and ranges reflect your own judgment of genuine business uncertainty, not an unreviewed AI default
- You have identified at least one sentence in the AI-drafted narrative that overstated certainty, or can explain why the draft passed the check cleanly
- The final narrative preserves the conditional nature of the scenario output — stating what result follows from which assumptions, rather than presenting a single number as settled fact
Watch out for
- Accepting an AI-suggested variable range without adjusting it to reflect the actual uncertainty of your specific business context
- Reading the AI-drafted narrative for grammar and clarity but not specifically for overstated certainty, which is a different kind of review
Hint
Read the narrative summary out loud and listen for any sentence that could be followed by 'according to whom, and under what assumptions?' — if that question has an obvious, uncomfortable answer, the sentence has overstated its certainty.
- AI-assisted scenario modeling tools like Anaplan and Pigment allow FP&A teams to compute far more scenario variants than manual spreadsheet work makes practical, which is valuable for exploring genuinely uncertain planning questions thoroughly.
- Expand the number of scenarios you compute using AI, but keep the number presented to decision-makers small and clearly differentiated — scale in exploration should not become scale in the final presentation.
- AI can suggest which variables to include in a sensitivity table and a starting range, but the actual range should reflect genuine business judgment about plausible uncertainty, not an unreviewed generic default.
- The overconfidence failure mode is the central risk: AI-generated scenario narratives can present a single point estimate in confident, definitive language that strips out the conditionality the underlying assumptions require.
- Review every AI-drafted scenario narrative specifically for language that overstates certainty — a well-written sentence is not evidence that the estimate behind it is more solid than the range actually supports.