Before funding
Compare options pursuing the same purpose, remove those that are not admissible and show which assumptions actually govern allocation.
When several AI ideas compete for the same resources, a weighted average is not enough. Arbitration must compare every candidate with a baseline, separate evidence from assumptions and show what would change the decision.
The score is never the starting point. The question, the option without a new AI system and the constraints are fixed before trade-offs are examined.
set the authority, objective, horizon, resources and baseline
assemble options that genuinely answer the same question
apply non-compensatory conditions before any ranking
connect every value to a function, evidence, date and owner
compare the value of state changes in a decision conference
test disagreements, weights, rank reversals and sign the choice
The five domains prevent blind spots; they prescribe neither five scores nor five fixed weights. Each engagement derives criteria that distinguish the options without counting the same effect twice.
| Domain | Question examined | Double count to remove |
|---|---|---|
| Useful effects | Which attributable change is expected against the baseline? | Do not count the same gain under value, adoption and feasibility |
| Data | Can the required data be used for the target population? | Separate measured quality, rights and the cost of remediation |
| Feasibility | Does the chain hold under target and degraded conditions? | Do not turn an uncontrolled risk into mere technical effort |
| Operating capability | Are the process, oversight and skills practicable? | Distinguish willingness to adopt from demonstrated capability |
| Risks and impacts | Which negative effects remain and who can stop them? | An unacceptable limit becomes a gate, not a low score |
The authority sets its risk tolerance and non-compensatory conditions before seeing the scores. These examples are adapted to the portfolio’s context, sector and legal regime.
| Non-compensatory condition | Effect | Evidence required to reopen |
|---|---|---|
| Prohibited use or residual impact outside tolerance | Stop | Revised qualification and risk brought below the approved threshold |
| Data without sufficient rights, coverage or quality | Defer | Rights, population, quality and lineage established |
| No owner, human oversight or fallback mode | Defer | Mandated owner and tested operating controls |
| Performance not demonstrated under target conditions | Reframe | Test protocol, baseline and reproducible results |
The pivot scenario is recalculated across documented bounds. A recommendation that changes after the first reasonable movement in a score or weight is not presented as a priority.
These values are entirely synthetic. They use an illustrative profile to verify the calculation; a live decision adds value functions, gates, sources, intervals and weight scenarios.
| Use case | Impact | Data | Feasibility | Adoption | Controls | Index | Reading |
|---|---|---|---|---|---|---|---|
| A Sales-call transcription and CRM update | 70 | 85 | 80 | 70 | 65 | 74.25 | Rank 1, still conditional |
| B Legal contract review assistant | 90 | 35 | 70 | 55 | 30 | 60.75 | Rank 2, evidence to strengthen |
| C Generative design for a production line | 55 | 40 | 35 | 50 | 45 | 45.75 | Rank 3, dominated in this scenario |
The committee receives the recommendation and everything required to reproduce, challenge or revise it when the context changes.
Primary sources
Last documentary review: 7 September 2026.
Value functions, swing weighting, decision conferencing, sensitivity and ranking robustness.
Swing weighting, sensitivity analysis, switching values and the warning against simple weighted scoring.
Govern, map, measure and manage the risks of an AI system.
Assessing the impacts of an AI system on individuals, groups and society.
Before you write to us
The mandate fixes one decision question, one baseline and the resources in contention. Raw performances are then translated through explicit value functions before weighting.
Because a total can conceal an unacceptable condition, double-counting or a trade-off the authority does not accept. Gates, scales and model admissibility are checked before calculation.
No. They are a coverage control. Criteria and ranges are constructed for the decision; weights then compare the value of changes of state through swing weighting.
No. Rank is relative to admissible options and the versioned frame. The authority receives possible reversals, uncertainty and limitations, then makes and signs the decision.
Next step
State the candidates, the authority deciding and the resources in contention. We will specify the protocol, evidence and robustness tests required.
Tell us about your situation