To run scenario analysis with AI, define one decision question and horizon, build an evidence-linked assumption register, and combine a few critical uncertainties into two to four internally coherent scenarios. Calculate impacts in a controlled model, not in the language model. Record leading indicators and review triggers so people can challenge assumptions without treating any scenario as a prediction.

Scenario analysis explores plausible conditions; it does not assign an automatic forecast or choose an option. A single rolling budget belongs in a forecast workflow, while value comparison belongs in cost-benefit analysis. The general How to Use AI guide provides the broader governance boundary.

Key Takeaways

Start with a decision question, horizon, scope, and owner.

Distinguish facts, assumptions, uncertainties, and decisions.

Use two to four scenarios that share consistent definitions and units.

Test dependencies and boundary conditions, not just dramatic narratives.

Calculate quantitative impacts in a deterministic model outside AI.

Do not invent probabilities or present scenarios as forecasts.

Track leading indicators, triggers, owners, and a review schedule.

Which decision should the scenario analysis inform?

Write the decision the analysis should inform without asking the model to make it. Include the owner, horizon, scope, excluded decisions, constraints, and what evidence would be useful.

Field	Record
Analysis ID/version	Stable identifier and revision
Decision question	Specific choice or preparedness question
Decision owner	Authorized person or group
Horizon	Start, end, and review cadence
Scope	Entities, markets, products, and processes included
Exclusions	Questions this exercise will not answer
Common measures	Units, currency, period, and definitions
Constraints	Policy, capacity, safety, or contractual boundaries

OECD describes strategic foresight as a structured way to explore possible futures, challenge assumptions, and make policy more prepared for uncertainty.［1］ It is not a method for claiming certainty. Keep the question broad enough to reveal alternatives but narrow enough that teams can define variables and indicators.

What belongs in a scenario assumption register?

Separate verified facts from working assumptions. An assumption register should show what is believed, why, where it applies, what it depends on, who owns it, and when it must be reviewed.

Field	Purpose
Assumption ID	Stable reference
Variable/baseline	What changes and the current planning value
Range/unit	Plausible bounds and dimension
Evidence	Original source and check date
Dependency	Other variables that constrain it
Boundary condition	Conditions outside which it should not be used
Uncertainty	Why the value or relationship may change
Owner	Person responsible for validation
Leading indicator	Observable signal
Review trigger	Threshold or event requiring reassessment

Use synthetic or de-identified examples. Do not paste confidential customer, account, financial, or approval data into an unapproved model. For source handling, follow safe AI research practices.

How do you identify critical uncertainties?

Not every variable deserves its own scenario axis. First list factors that materially affect the decision, then assess uncertainty and controllability. Separate predetermined elements, trends, critical uncertainties, and decisions under management control.

Prioritize uncertainties that are both consequential and genuinely uncertain within the horizon. Avoid choosing axes only because they create catchy labels. Record why each factor was selected and what was excluded.

Factor	Impact	Uncertainty	Dependency	Role
Synthetic demand timing	High	High	Capacity and lead time	Scenario driver
Signed capacity floor	High	Low	Contract	Boundary condition
Team response choice	High	Controllable	Governance	Decision, not scenario fact

The OECD Foresight Toolkit presents practical approaches for scanning, scenarios, stress-testing, and linking foresight to decisions.［2］ Use such methods as structure, while adapting evidence and governance to the actual organization.

Define consistent variable ranges

For every scenario driver, record a baseline, low and high bounds or categorical states, unit, evidence, dependency, and owner. A range is not a probability distribution. It is a controlled input boundary for exploration.

Reject combinations that violate physics, contracts, accounting identities, capacities, or other established constraints. If two variables are linked, document the relationship instead of varying them independently.

Example using synthetic values:

Variable	State 1	State 2	Unit	Dependency
Demand timing	4	8	weeks to peak	Must fit planning horizon
Supply lead time	3	7	weeks	Cannot be below signed minimum
Available capacity	90	110	units/week	Constrained by approved staffing

These numbers only demonstrate the table. They are not probabilities, forecasts, or recommendations.

Construct two to four coherent scenarios

Two scenarios may expose a critical contrast; three or four can reveal combinations without creating an unreadable catalog. More scenarios do not automatically improve analysis. Each scenario should use the same decision question, horizon, units, and common facts.

Create a scenario card:

Field	Record
Scenario ID/name	Neutral identifier, not “best” or “worst”
Narrative	Concise explanation of conditions
Variable combination	Assumption IDs and states
Dependencies	Relationships that must hold
Boundary conditions	Limits and exclusions
Quantitative inputs	Values passed to the controlled model
Model location	Workbook version, query, or script run
Impacts	Locked outputs and qualitative consequences
Leading indicators	Observable signals
Review triggers	Thresholds or events
Disconfirming evidence	What would weaken the scenario
Owner	Review responsibility

Use neutral names such as S1-earlier-demand rather than labels that imply desirability. A scenario is a coherent set of conditions, not a story generated around arbitrary numbers.

Use scenario analysis with AI to challenge structure, not decide truth

A bounded prompt can ask the model to find missing dependencies and contradictions:

Review the supplied decision frame, assumption register, and variable states.
Propose 2–4 internally coherent scenario cards using only listed assumptions.
Do not add facts, probabilities, forecasts, costs, benefits, or recommendations.
For every statement cite Assumption IDs. Flag incompatible combinations,
missing dependencies, boundary violations, and unsupported narrative as [VERIFY].
Return questions for human owners; do not choose a scenario or option.

Treat the output as a challenge list. The model may repeat familiar patterns, overstate causal links, or create false symmetry. Assumption owners must decide what is plausible and supported.

How can you test a scenario's internal consistency?

Review each scenario line by line and across scenarios. Ask:

Do all input units and periods match?
Are linked variables moving in a defensible relationship?
Does any state breach capacity, policy, contract, or physical limits?
Does the narrative introduce a cause not in the register?
Are common facts held constant?
Are decisions mistakenly presented as external conditions?
Would the scenario still make sense if its label were removed?

Record every conflict, owner, decision, and resulting version change. Do not repair a contradiction by hiding it in prose.

Calculate impacts outside the language model

Pass the approved scenario input table into a protected spreadsheet, simulation, database query, or reviewed script. Define formulas, null behavior, rounding, currency, time steps, and dependencies. Test with known synthetic cases and preserve the run ID.

AI may draft code or formulas for human review, but the executed, tested model is the calculation source. Independently reconcile inputs and outputs. The language model must not infer missing values, estimate totals in prose, or overwrite locked results.

Use analyzing spreadsheets with AI only as an assistance pattern; deterministic formulas and responsible owners remain authoritative.

Compare impacts without selecting an option

Create a cross-scenario view using the same measures and definitions:

Measure	S1	S2	S3	Source/model	Interpretation owner
Synthetic backlog at horizon	20	45	10	Model run R-12	Operations
Capacity breach week	None	6	None	Model run R-12	Operations
Required review trigger	T-02	T-01	T-03	Trigger register	Decision owner

Describe sensitivities, bottlenecks, reversibility, and preparedness implications. Do not convert the table into a cost-benefit conclusion or recommendation. Decision makers may need other evidence, duties, and criteria beyond the scenario model.

The Green Book discusses uncertainty, optimism bias, scenarios, sensitivity, and transparent appraisal in UK public decision-making.［3］ It is a regional framework, not universal legal or financial advice. Local finance and decision owners must define the applicable method.

Do not assign unsupported probabilities

A scenario can be useful without a probability. If an organization has validated statistical evidence and an approved method, qualified analysts may maintain probabilities in the controlled model. The language model should not invent “30% likely” because a scenario sounds plausible.

Keep these concepts separate:

plausibility: conditions are coherent and not contradicted by known evidence;
probability: a quantified likelihood supported by a defined method and data;
preference: a decision maker's value judgment;
preparedness: an action that improves resilience across conditions.

Never let prose slide from plausibility to probability or from impact to recommendation.

Which scenario indicators and triggers should prompt a review?

A leading indicator is an observable signal related to an assumption. A trigger is a pre-agreed threshold or event that requires review; it is not an automatic decision unless governance explicitly says so.

Indicator ID	Assumption	Measure/source	Cadence	Trigger	Owner	Response
I-01	A-03 lead time	Approved supplier report	Monthly	Two periods above bound	Procurement	Reassess all scenarios

Check that the indicator can be observed soon enough, is sourced consistently, and is not merely the final outcome. Record false-positive and lag risks. Assign owners and escalation paths.

Run a cross-functional challenge

Invite operations, finance, commercial, risk, legal, security, or other relevant functions to challenge assumptions within their authority. Ask each reviewer to identify missing dependencies, boundary violations, disconfirming evidence, and signals they would trust.

Use project planning with AI to organize owners and review dates, but do not turn the plan into an automatic approval process. Record minority views and unresolved issues rather than forcing consensus.

NIST's generative AI profile highlights confabulation, privacy, and governance risks.［4］ Bounded inputs, traceable assumption IDs, external calculation, and human challenge reduce those risks but cannot prove a future state.

Schedule review and archive versions

Review on a fixed cadence and when a trigger fires, evidence changes, boundaries break, or the decision question changes. Retire scenarios that no longer share the current horizon or definitions; do not silently rewrite their history.

Archive the decision frame, assumption register, evidence references, scenario cards, controlled model and run IDs, check report, indicator register, challenge notes, decisions, and approvals. Exclude secrets and unnecessary personal data.

Final checklist

The decision question, owner, horizon, scope, exclusions, and measures are explicit.
Facts, assumptions, uncertainties, and controllable decisions are separate.
Every assumption has evidence, bounds, dependencies, owner, and review date.
Two to four scenarios use consistent units, periods, definitions, and facts.
Each scenario records variables, dependencies, boundaries, impacts, indicators, and triggers.
Internal contradictions and disconfirming evidence were documented.
Quantitative impacts came from a tested controlled model outside AI.
No unsupported probability, prediction, recommendation, or cost-benefit conclusion was added.
Cross-functional owners challenged the scenarios.
Review triggers and version archives are in place.

FAQ

How many scenarios should I create?

Usually two to four are enough to expose meaningful combinations while remaining reviewable. Choose the smallest set that challenges critical assumptions.

Is a scenario the same as a forecast?

No. A forecast is an estimate of an expected path under defined assumptions. A scenario explores a coherent set of possible conditions without claiming it will occur.

How should I choose assumptions?

Start from the decision and evidence, then prioritize factors that are both consequential and uncertain. Record exclusions and dependencies.

Should every scenario have a probability?

No. Do not assign one without validated data, a defined method, and qualified ownership. AI must not invent probabilities.

How do I test internal consistency?

Check units, periods, dependencies, boundaries, common facts, causal statements, and whether controllable decisions were mistaken for external conditions.

What is a useful leading indicator?

It is observable, consistently sourced, timely enough to inform review, and meaningfully related to a registered assumption.

Is a trigger an automatic decision?

Not normally. It is a threshold or event that requires reassessment and escalation under the defined governance process.

When should scenarios be reviewed?

Review on a fixed cadence and whenever indicators cross triggers, evidence changes, a boundary fails, or the decision question or horizon changes.

Sources

OECD, “Strategic Foresight” — https://www.oecd.org/en/about/programmes/strategic-foresight.html
OECD, “Foresight Toolkit for Resilient Public Policy” — https://www.oecd.org/content/dam/oecd/en/publications/reports/2025/01/foresight-toolkit-for-resilient-public-policy_9ad1cd60/bcdd9304-en.pdf
UK Government, “The Green Book 2026” — https://www.gov.uk/government/publications/the-green-book-appraisal-and-evaluation-in-central-government/the-green-book-2026
National Institute of Standards and Technology, “Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile” — https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence

Sources checked 31 August 2026.

How to Run Scenario Analysis with AI—and Test Assumptions