One of the most useful things AI can build is not an answer but a decision instrument: a small spreadsheet or application where a person changes assumptions and sees when the conclusion reverses. Instead of asking “Which supplier should I choose?” the user receives a model of costs, benefits, risks and constraints that shows exactly which unknowns matter.[3][4]
The important shift is from recommendation to inspectability. A language model can make a choice sound inevitable. A simulator makes the choice conditional: Option A wins if volume exceeds this level, review time stays below that level and reliability meets the threshold. Change one assumption and the result updates in front of the user.
Start with a decision, not a dashboard
Write the decision in one sentence, list the genuine alternatives and define the outcome that matters. If the objective is “lowest licence price,” the model will optimise something different from “lowest cost per accepted task.” If the objective is not explicit, AI will quietly invent one.
Then separate inputs into three classes: observed facts, controllable choices and uncertain assumptions. Give every material input a source, date, unit, owner and plausible range. The model should never place a confident slider around a number that has no evidential basis.
The minimum data model
| Measure | Value | Context |
|---|---|---|
| Options | The mutually exclusive or combinable choices | Define what is included in each option so costs are comparable. |
| Outcome metric | The value the decision is trying to improve | Examples: total accepted-task cost, time to delivery, reliability or weighted public value. |
| Observed inputs | Prices, volumes, rates and performance already measured | Record source, period and uncertainty. |
| Assumptions | Future volume, adoption, error rates or benefits | Use ranges and scenarios instead of pretending these are facts. |
| Constraints | Budget, law, capacity, time or safety limits | An option that violates a hard constraint should not win because of a score. |
| Acceptance tests | Known examples and boundary cases | Check selected known outputs and boundary behaviour; passing examples cannot prove every possible input correct. |
A worked example: two AI systems
Suppose System A costs £18,000 a year plus £0.06 per completed task. System B costs £4,000 plus £0.16 per task. Ignoring every other difference, the annual cost is equal at 140,000 tasks: the £14,000 fixed-cost gap divided by the £0.10 variable-cost advantage. Below that volume, B is cheaper; above it, A is cheaper.
Now include human review. If A needs three minutes per task and B needs one and a half, at £30 per hour review adds £1.50 for A and £0.75 for B. Total variable cost becomes £1.56 versus £0.91 per task. For N annual tasks, A therefore costs £14,000 + £0.65 × N more than B. Under these hypothetical assumptions, there is no non-negative volume at which A becomes cheaper: B has both the lower fixed cost and the lower total variable cost.
What changes the illustrative choice
| Assumption | System A | System B | Decision effect |
|---|---|---|---|
| Fixed annual cost | £18,000 | £4,000 | Favours B at low volume |
| Variable model cost | £0.06/task | £0.16/task | Favours A as volume rises |
| Human review | 3 minutes/task | 1.5 minutes/task | At £30/hour, adds £1.50 versus £0.75 per task and dominates the model-price gap |
| Model-only switching volume | 140,000 tasks | 140,000 tasks | Valid only when review, failures and tool charges are excluded |
| Reliability threshold | Must be measured | Must be measured | A cheaper result is irrelevant if it cannot meet the acceptance criterion |
Ask for sensitivity and switching values
HM Treasury’s Green Book recommends sensitivity analysis to test how key assumptions affect results and switching values to identify the point at which an option stops being preferred or value for money. These ideas are useful far beyond public appraisal. A simulator should reveal the few variables capable of reversing the conclusion and the range within which the choice remains robust.[1]
For the example above, the first model-only switching value is task volume. Once review time is included, the decisive variable may become the difference in minutes of human correction. Other decisions may be controlled by churn, energy price, failure rate, discount rate or the probability of a costly adverse outcome. AI can solve the formulas and generate a chart, but the user should understand the logic in ordinary language.
Add non-monetary criteria without hiding the judgement
Some choices involve privacy, resilience, strategic control or accessibility that cannot be reduced credibly to a single price. Multi-criteria decision analysis can make those factors explicit, but weights remain value judgements. Show the score under several reasonable weight sets and identify when the preferred option changes. Do not let the AI choose ethical or strategic priorities and then present them as mathematics.[2]
A high-standard build prompt
- Restate the decision, alternatives, objective and hard constraints before building anything.
- Create a data dictionary with units, periods, sources, uncertainty ranges and owners for each input.
- Write every formula in plain language and in the spreadsheet or code; do not hide calculations in narrative text.
- Include base, upside and downside scenarios plus one-variable and two-variable sensitivity tests.
- Calculate switching values for the assumptions most capable of changing the preferred option.
- Create test cases for zero values, missing inputs, extreme values and known hand-calculated examples.
- Display a conclusion only with the assumptions that make it true and a list of missing evidence.
Verify before using it
Recalculate several rows independently. Change one input at a time and confirm the expected direction of the result. Test boundary values immediately above and below every switching point. Inspect the generated code or spreadsheet formulas rather than trusting a chart. If the simulator uses uploaded business data, check that rows, units and time periods were parsed correctly before interpreting the result.[3][4]
The finished tool can be small: a spreadsheet with sliders, a local web app or an interactive notebook. Its value comes from making uncertainty explorable. Instead of AI deciding for the user, AI helps the user see the structure of the decision, discover the assumption that matters and identify the next piece of evidence worth buying or collecting.
Sources & further reading
- The Green Book (2026) HM Treasury · 1 Jan 2026 · Accessed 9 Oct 2026
- Multi-criteria decision analysis: supplementary guidance UK Government · Accessed 9 Oct 2026
- Analyzing data with ChatGPT OpenAI Academy · 10 Apr 2026 · Accessed 9 Oct 2026
- Data analysis with ChatGPT OpenAI Help Center · Accessed 9 Oct 2026