How to develop and assess scenario reasoning in AI-enabled work

Use scenarios to analyze uncertainty, not decorate forecasts

Begin with a focal decision and time horizon. A useful scenario changes what the decision owner would do now. “AI adoption increases” is too vague. “Within 12 months, review volume doubles while specialist capacity stays flat” creates conditions that can be tested against a plan.

The UK Green Book 2026 describes scenario analysis as examining plausible future outcomes and considering a proposal under those outcomes. It distinguishes scenarios from a single forecast and notes that simple what-if questions may be proportionate for lower-cost, lower-risk proposals. The OECD describes strategic foresight as systematic exploration of plausible futures rather than prediction and uses scenarios to stress-test strategies and develop actionable plans.

Do not build a best, middle, and worst case by changing every variable in the same favorable or unfavorable direction. Those bundles often conceal assumptions. Instead, identify two or three uncertainties that are both consequential and genuinely unsettled, then vary them in ways that create distinct operating environments.

Build a BRANCH scenario set

Use BRANCH to turn uncertainty into an inspectable artifact:

| Field | Question | Evidence to retain | | --- | --- | --- | | Boundary | What decision, system, population, and horizon are in scope? | Focal question and exclusions | | Relevant drivers | Which external and internal forces could materially change conditions? | Current evidence, trend, and uncertainty note | | Alternatives | Which distinct, plausible worlds will be examined? | Scenario names and driver states | | Narrative chain | How do the conditions lead to operational consequences? | Cause-and-effect sequence and assumptions | | Choices | Which actions are robust, contingent, delayed, or abandoned? | Strategy-by-scenario matrix | | Harbingers | Which observable signposts suggest conditions are changing? | Indicator, threshold, owner, and cadence |

Each scenario should be plausible, internally consistent, relevant to the focal decision, and sufficiently different to change at least one implication. Plausible does not mean equally likely. Avoid attaching precise probabilities unless there is a defensible basis. Record evidence for the drivers separately from the imagined combination of conditions.

The narrative chain matters because a scenario name can smuggle in conclusions. “High-growth future” says little about workload distribution, budget timing, data access, or user behavior. A chain makes the reasoning challengeable: if review demand rises, which queue grows, what capacity becomes binding, and which control fails first?

Worked artifact: planning an AI review service

Imagine a synthetic enablement team planning a one-year internal service that reviews high-stakes AI-assisted documents before distribution. This example is original and does not describe a Two Prune customer, forecast, or benchmark.

The focal question is: “Which operating model should the team start now if demand and source quality remain uncertain over the next 12 months?” The team identifies two critical uncertainties: review demand and the proportion of submissions with usable source records. It builds four scenarios:

The BRANCH artifact traces consequences. In Focused flow, a small central review team can handle exceptions. In Evidence drag, reviewers spend more time reconstructing provenance than evaluating claims. In Volume pressure, triage and distributed reviewers become important even though each case is relatively inspectable. In Compound strain, the proposed central queue becomes an operational bottleneck and may encourage teams to bypass review.

The strategy matrix identifies one robust action: define submission requirements and capture baseline demand before adding automation. It identifies a contingent action: train distributed reviewers if a four-week demand average crosses the central team's tested capacity. It identifies another contingent action: introduce a source-record return route if more than a stated share of sampled submissions lack the required evidence fields. Purchasing a larger workflow system is delayed until the team has observed which constraint is binding.

The harbinger record names an owner and cadence for each signpost. The service lead reviews weekly case volume and queue age. The quality lead samples source-record completeness twice a month. A threshold is not presented as universal truth; it is an operating trigger chosen for this pilot and is revisited after six weeks.

The scenario set does not predict Compound strain. Its value is showing that a central-review-only plan is fragile when both drivers move unfavorably, while the robust first actions create evidence and preserve options.

• Focused flow: demand stays moderate and most submissions arrive with usable sources.
• Evidence drag: demand stays moderate but source records are often incomplete.
• Volume pressure: demand rises sharply while source records remain usable.
• Compound strain: demand rises and source records are frequently incomplete.

Practice with scenario mutation

Give a learner a plan, a short evidence pack, and two uncertain drivers. Ask for three or four BRANCH scenarios and a strategy matrix. Then mutate one condition without changing the others. For example, add a new source-access restriction, shorten the time horizon, or make a planned action irreversible.

The learner should update the narrative chain, implications, and contingent actions rather than simply rename the scenario. Ask which action remains robust, which becomes unsafe, and which new signpost matters. This exposes whether the scenarios are doing analytical work.

AI can propose drivers, generate counterintuitive combinations, and challenge a narrative for missing consequences. Require the worker to verify present-day driver evidence and remove speculative chains that cannot be explained. Model novelty is not a substitute for relevance or plausibility.

Assess scenario reasoning from the artifact

Look for observable distinctions:

Retain the focal question, driver evidence, uncertainty notes, scenario cards, strategy matrix, signpost register, and revisions. A vivid narrative alone is weak evidence if it does not change action or reveal fragility.

• decision focus: the scenarios connect to a named choice and horizon;
• driver quality: consequential uncertainty is separated from known facts and minor variation;
• plurality: the set explores multiple plausible futures rather than one forecast with cosmetic ranges;
• internal coherence: each narrative explains how driver states produce implications;
• strategy testing: actions are classified as robust, contingent, delayed, or unsuitable;
• signpost design: indicators are observable, owned, and connected to a decision;
• adaptation: new conditions change implications or actions when they should;
• epistemic discipline: scenarios are not reported as predictions or measured outcomes.

Verify the scenario set

Run a consistency check within each scenario: do the driver states, narrative chain, and operational consequences fit together? Then run a diversity check across scenarios: if the same action is recommended for the same reason every time, the scenarios may not be meaningfully different or the action may simply be robust. State which interpretation applies.

Use a pre-mortem on the preferred plan in each scenario and identify the earliest observable sign of failure. Confirm that signposts can be measured at the proposed cadence and that an owner can act on them. Finally, perform a surprise check by combining one neglected driver state with the most important scenario. If that combination changes the decision, add it or explain why it is outside scope.

Scenario reasoning makes uncertainty actionable without claiming foresight. It does not assign reliable probabilities by itself or prove that a strategy will succeed. In a Two Prune assessment, the work can provide task-bounded evidence about Judgment & Synthesis, Context Framing, and Verification & Risk Control. It remains an assessment of observed work in the simulation, not a forecast of future performance or an employment decision.

Sources

UK Government: The Green Book 2026

• OECD: Strategic Foresight Toolkit for Resilient Public Policy
• OECD: Strategic Foresight

Sources: [1] [2] [3]