Scenario Analysis Prompts for Stronger Consulting Decisions
A polished AI response can still hide a weak assumption, a broken formula, or a number with no source. In consulting, that can send a client discussion in the wrong direction.
Well-built scenario analysis prompts help you structure uncertainty before you start debating conclusions. They give ChatGPT or Claude enough context to organize assumptions, expose information gaps, and produce an auditable first-pass analysis.
Use AI to accelerate the work, then test every material calculation and claim in the underlying model.
Scenario Analysis Prompts: What They Should Produce
A financial model is a structured representation of business performance, often built around drivers such as volume, price, costs, investment, and timing. IBM’s overview of financial modeling is a useful reference for that foundation.
A good consulting prompt doesn’t ask AI to predict the future. It asks the model to make uncertainty visible, use supplied inputs only, and show how different conditions affect a decision.
Scenario analysis changes a coherent set of conditions
Scenario analysis changes several connected inputs at once. A downside case for a manufacturer, for example, might combine lower demand, discounting, higher freight costs, and delayed capacity expansion.
The purpose is to answer a business question under a plausible operating environment: Can the client maintain target EBITDA? Does a market entry still clear the investment hurdle? Which contingency plan should leadership prepare?
CFI’s comparison of scenario and sensitivity analysis captures the central distinction. Scenarios change multiple assumptions together, while sensitivity testing isolates the effect of one input, or a limited set of inputs.
Sensitivity analysis tests the pressure points
Sensitivity analysis holds most assumptions constant and changes one driver at a time. A consultant might test gross margin at 28%, 30%, and 32%, while keeping price, volume, headcount, and overhead fixed.
This makes uncertain variables easier to rank. It also identifies thresholds, such as the minimum annual customer retention rate needed to protect a client’s valuation target.
A sensitivity table can look precise even when the base case is weak. Validate the base case before treating any output range as decision-ready.
Build the Input Brief Before Writing a Prompt
AI needs a defined problem, not a vague instruction to “analyze the business.” Start with a short approved brief that separates known inputs from estimates and unresolved questions.
Do not paste confidential client documents, customer lists, unannounced results, pricing schedules, or personally identifiable information into a public AI account. Use anonymized labels, ranges, and approved workspace controls.
Include the inputs that change the answer
For most engagements, the prompt should name the industry, client situation, decision owner, time horizon, units, key variables, constraints, and decision criteria. Also state whether figures are actuals, management estimates, third-party data, or placeholders.
A helpful input record might distinguish between “FY2025 reported revenue,” “management forecast,” and “assumption requiring validation.” That label prevents AI from presenting a planning estimate as a fact.
Use this intake prompt before requesting scenarios:
Act as a consulting analyst supporting a [INDUSTRY] client.
Client context: [ANONYMIZED CLIENT CONTEXT].
Decision to support: [DECISION].
Time horizon: [TIME HORIZON].
Units and currency: [UNITS].
Known facts and approved assumptions: [INPUTS].
Variables that remain uncertain: [VARIABLES].
Decision criteria: [DECISION CRITERIA].First, create an input register with columns for input, value, unit, source status, confidence level, and effect if wrong. Mark missing information as “[fact check needed].” Do not estimate missing values or invent client facts. Then list the five questions that would most improve the analysis.
That first pass often catches a hidden problem: the team has a conclusion in mind but hasn’t defined the measure that determines success.
Create Decision-Focused Scenarios, Not Fictional Stories
Strong scenarios reflect events management can recognize and act on. “Optimistic,” “base,” and “pessimistic” labels aren’t enough unless each case has a business logic.
For instance, a retail expansion case may connect store opening dates, footfall ramp-up, conversion rate, inventory turns, labor costs, and lease commitments. If a scenario changes revenue but leaves working capital untouched, it may understate cash needs.
Use drivers that move together
Ask AI to flag variables that may interact. Price reductions can increase unit demand but reduce gross profit per unit. Delayed hiring may lower expenses but also slow implementation and revenue realization.
A scenario should not assume every negative input arrives at once without a credible mechanism. Likewise, it should not combine conditions that conflict, such as peak demand and unusually large excess capacity, unless the client context supports both.
Use this copy-and-paste prompt:
You are analyzing strategic scenarios for a [INDUSTRY] client. Use only the information provided below.
Client context: [ANONYMIZED CLIENT CONTEXT]
Decision: [DECISION]
Time horizon: [TIME HORIZON]
Units: [UNITS]
Base-case assumptions: [ASSUMPTIONS]
Key variables: [VARIABLES]
Required outputs: [REVENUE, EBITDA, CASH FLOW, NPV, MARKET SHARE, OR OTHER METRICS]
Decision criteria: [DECISION CRITERIA]Build three scenarios: base case, downside case, and upside case. For each scenario, state the operating narrative, changed assumptions, rationale, interactions between variables, calculated outputs, and risks not captured by the model. Use ranges where inputs are uncertain. Label every unsupported input “[fact check needed].” Do not fabricate benchmarks, market data, customer behavior, or citations. Finish with the decision implications, conditions that would reverse the recommendation, and a list of assumptions requiring client validation.
Run Sensitivity Testing Around the Variables That Matter
A sensitivity analysis is most useful after the team has agreed on a credible base case. Otherwise, you may spend hours testing the wrong question with mathematical discipline.
For a SaaS growth plan, retention, average revenue per account, sales capacity, and gross margin may matter more than a small change in office expenses. For an industrial client, utilization, realized price, input costs, and capex timing may dominate the result.
Request a ranked sensitivity table
Ask for a bounded range, clear increments, and one unchanged reference point. Do not ask AI to select ranges without evidence. The range should come from historical variation, management guidance, contracts, market research, or an explicit planning assumption.
Act as a financial and strategy analyst for a [INDUSTRY] consulting engagement.
Client context: [ANONYMIZED CLIENT CONTEXT]
Base-case model inputs: [ASSUMPTIONS]
Output metric: [OUTPUT METRIC]
Variables to test: [VARIABLES]
Test range and increments for each variable: [RANGES AND INCREMENTS]
Units: [UNITS]
Time horizon: [TIME HORIZON]
Decision criteria: [DECISION CRITERIA]Hold all other inputs constant unless stated otherwise. Create a sensitivity table for each variable and rank variables by impact on [OUTPUT METRIC]. Identify the break-even threshold where possible. State the formula logic before calculating. Flag interactions that single-variable testing does not capture. Do not create data, round away material differences, or present estimates as facts. List all calculations that require independent spreadsheet verification.
Test two-variable interactions when the decision requires it
One-way sensitivity testing is easy to read, but decisions often depend on two linked drivers. A deal team may need to see how purchase price and exit multiple affect IRR. A transformation team may need to test adoption rate and implementation cost together.
A two-way table can reveal a risk that individual tests hide. However, it can also create false confidence if the range selections lack evidence.
Lucid’s discussion of multi-scenario financial analysis offers a practical reminder that driver-based analysis needs structure, not a stack of disconnected what-if cases.
Use a second prompt when interaction matters:
Using the approved base case below, create a two-variable sensitivity analysis.
Industry: [INDUSTRY]
Client context: [ANONYMIZED CLIENT CONTEXT]
Output metric: [OUTPUT METRIC]
Variable A: [VARIABLE A], tested at [VALUES]
Variable B: [VARIABLE B], tested at [VALUES]
Fixed assumptions: [FIXED ASSUMPTIONS]
Units: [UNITS]
Time horizon: [TIME HORIZON]
Decision rule: [DECISION CRITERIA]Return a matrix with Variable A as rows and Variable B as columns. Show the formula used in each cell, identify the cells that fail the decision rule, and explain whether the two variables may be correlated. If correlation cannot be supported by the supplied information, state that limitation clearly. Do not infer missing values.
Verify AI Calculations Before Sharing Results
AI can describe calculation logic well, but it can still misread a unit, apply a percentage incorrectly, or use an inconsistent sign convention. Treat model output as a working paper until a qualified reviewer checks it.
Start with the most decision-relevant outputs. Recalculate NPV, IRR, EBITDA, cash balance, payback period, and break-even points in Excel, Google Sheets, or the client’s approved planning tool. Then trace each output back to its source inputs.
Audit formulas, units, and time periods
Check whether revenue is monthly while payroll is annual. Confirm whether a percentage change is a percentage-point change. Inspect tax assumptions, discount rates, working-capital timing, and terminal-value logic.
AI should show formulas in plain language, but the human reviewer should rebuild material formulas independently. A matching answer matters less than a transparent method.
Use this review prompt after you have exported model outputs:
Review the analysis below as a skeptical consulting-model reviewer.
Model purpose: [DECISION]
Inputs and source status: [INPUT REGISTER]
Assumptions: [ASSUMPTIONS]
Formulas and outputs: [MODEL OUTPUT]
Units and time horizon: [UNITS AND TIME HORIZON]Identify possible formula errors, unit mismatches, double counting, unsupported assumptions, circular logic, false precision, and missing downside risks. Separate confirmed issues from items requiring review. For each issue, state the exact calculation or source check needed. Do not correct a number unless the provided formula and inputs support the correction.
Separate uncertainty from ignorance
Uncertainty means the team knows a variable can vary within a defensible range. Ignorance means the team lacks enough evidence to define the range at all.
Don’t disguise ignorance with a narrow range or extra decimal places. A 7.25% assumption looks rigorous, but it may be less credible than a stated 6% to 9% planning range with a source note.
For probabilistic work, Monte Carlo methods can propagate uncertainty through a model, but they still depend on defensible inputs and distributions. NIST’s overview of measurement uncertainty reinforces the broader principle: the method does not remove uncertainty, it characterizes it.
Turn Model Output Into a Client Recommendation
Clients rarely need every row of a sensitivity table. They need a recommendation, the conditions behind it, and the choices available if conditions deteriorate.
Lead with the decision. Then state the base-case result, the few variables that most affect it, and the threshold that changes the recommendation.
Use a concise executive-readout prompt
Avoid asking AI for a persuasive story. Ask for a balanced readout that separates evidence, interpretation, and recommended action.
Draft a client-ready scenario analysis readout for [CLIENT ROLE].
Decision: [DECISION]
Base-case result: [BASE-CASE OUTPUT]
Scenario results: [SCENARIO OUTPUTS]
Highest-impact sensitivities: [SENSITIVITY RESULTS]
Assumptions requiring validation: [OPEN ITEMS]
Decision criteria: [DECISION CRITERIA]Write: a direct recommendation, the evidence supporting it, the two most important risks, trigger points that require a revised decision, and practical next steps. Distinguish supplied facts from modeled estimates. Use no confidential details, invented benchmarks, or certainty language beyond the evidence.
The strongest client message may be conditional: proceed only if volume reaches a defined threshold, phase spending until a regulatory decision, or renegotiate a price if the downside case breaches cash limits.
Final Thoughts
Useful scenario analysis prompts create a disciplined starting point, not a validated answer. They make assumptions visible, test the drivers that matter, and reveal where a recommendation could fail.
Keep AI inside a controlled workflow: provide approved inputs, mark evidence gaps, rebuild material calculations, and let the decision criteria guide the final client conversation.