How to develop and assess trade-off analysis in AI-enabled work

Define the exchange before scoring options

A trade-off exists when an option advances one relevant objective, value, or stakeholder interest while weakening another. It is not simply a list of pros and cons. The comparison needs a shared decision, a credible baseline, a set of feasible alternatives, and evidence about the consequences that differ between them.

The UK Green Book 2026 describes appraisal as assessing costs, benefits, and risks across options so decision makers can understand potential impacts and trade-offs. It also says that a single summary metric is not enough for a balanced judgment and that decision makers may need to consider distributional effects, risk, uncertainty, and unmonetisable benefits alongside calculated values. This provides a useful discipline for ordinary workplace choices even when no formal economic appraisal is required.

Start by writing four anchors:

Constraints should be applied before preference scoring. An option that violates a non-negotiable condition is infeasible, not merely less attractive. Keep uncertainty separate as well. A weak estimate is not a bad outcome; it is a reason to lower confidence, seek evidence, or choose a more reversible path.

An AI assistant can help enumerate possible effects, expose omitted stakeholders, or challenge an option. It cannot decide which interests an organization is entitled to trade, invent reliable weights, or convert missing evidence into a neutral score.

• Decision: What choice must be made, by whom, and by when?
• Baseline: What happens if the current approach continues?
• Constraints: Which legal, ethical, operational, budget, or timing limits cannot be traded away?
• Differences: Which consequences genuinely vary across the feasible options?

Build a TRADE comparison board

Use TRADE as a page-specific artifact for the comparison:

| Field | Question | What to record | | --- | --- | --- | | Thresholds | Which conditions must every option meet? | Constraint, owner, evidence, and pass or fail | | Results | What changes relative to the baseline? | Direction, magnitude or range, timing, and source | | Affected parties | Who receives each benefit, burden, or risk? | Stakeholder, distribution, and material exceptions | | Doubt | Which assumptions or estimates could change the comparison? | Confidence, sensitivity, missing evidence, and reversibility | | Exchange | What is gained and surrendered in each viable option? | Plain-language trade-off statement without a verdict |

The results column should use a common comparison basis. If one option is described by annual cost and another by total three-year cost, normalize them or make the mismatch explicit. If a consequence cannot be quantified responsibly, describe its direction, likely significance, affected group, and evidence instead of assigning a decorative number.

For every major difference, write a sentence in this form:

> Compared with the baseline, this option improves X for party A over period T, while worsening or placing at risk Y for party B; confidence is level because evidence reason.

That sentence forces the analyst to expose distribution and uncertainty. It also makes selective framing easier to detect. A board containing only aggregate benefits can hide that one group absorbs most of the cost. A board containing only central estimates can hide that the apparent ranking reverses under a plausible range.

Weights may be useful when decision criteria are legitimate and understood, but they should not manufacture precision. HM Treasury's guidance on multi-criteria decision analysis says the method can support preference ranking where technical trade-offs are complex. The weights, scales, and evidence still require scrutiny. Run the comparison both with and without the weighted score, and identify any small weight change that switches the order.

Worked artifact: changing an internal research workflow

Consider an original synthetic example. A policy team must choose how to review external sources used in monthly briefing notes. The current process relies on each author to check sources. The team is comparing three changes:

The team has four weeks of workflow logs, eight reviewer interviews, and a sample of 30 briefing claims. These materials can inform the comparison, but they do not establish universal error rates or prove how any option will perform at scale.

The TRADE board first applies thresholds. Any option must preserve a named source for consequential claims, avoid copying restricted material into the briefing system, and fit the existing publication deadline. All three can be designed to meet the first two conditions. Interviews indicate that reviewing every claim may not fit the deadline during peak weeks, so that option remains conditionally feasible only if scope or staffing changes.

The results section then records differences from the current process:

| Consequence | Review every claim | Evidence table plus sample | Focused review | | --- | --- | --- | --- | | Author effort | Likely lower than the evidence table but dependent on reviewer queries | Higher at drafting time because rationale and uncertainty are recorded | Moderate, provided tagging rules are understood | | Reviewer effort | Highest and sensitive to claim volume | Bounded by sample size, with added time for exceptions | Sensitive to how many claims are tagged | | Inspectability | Reviewer comments show individual checks | Table retains source, claim, uncertainty, and disposition | Strong for tagged claims, limited for untagged ones | | Coverage risk | Fewer unchecked claims if the process is completed | Sampling may miss a weak unreviewed claim | Misclassification can keep consequential claims out of review | | Deadline risk | Material in peak weeks | Lower if the sample remains bounded | Moderate if tagging becomes overinclusive |

The affected-parties section shows that authors carry most of the documentation burden in the evidence-table option, while designated reviewers carry most of the timing burden in the full-review option. The focused option shifts judgment to authors who must tag claims correctly. Those are different exchanges, even if a simple score makes the options look close.

The doubt section identifies the comparison's pivotal unknown: how consistently authors can identify consequential or uncertain claims. The team therefore avoids recommending an option yet. It proposes a two-week tagging exercise using historical, non-sensitive briefings, with an independent reviewer checking disagreements. The exercise is evidence collection for the trade-off board, not proof that the focused approach is best.

The final exchange statements are plain:

No option is declared universally superior. The decision owner can now see which burden moves where and which missing evidence matters.

• a mandatory second-person review for every external claim;
• a structured evidence table completed by the author and sampled weekly;
• a focused second-person review only for claims tagged consequential or uncertain.
• full review buys broader review coverage at the cost of reviewer capacity and deadline resilience;
• the evidence table buys inspectability and bounded review effort at the cost of author effort and residual sampling risk;
• focused review concentrates human attention but depends on reliable tagging and leaves less visibility into ordinary claims.

Practice by changing the decision context

Give a learner three feasible options and a small evidence pack. Ask for a TRADE board without a recommendation. Then change one feature of the decision:

The learner should update only the affected fields, explain whether an option becomes infeasible, and identify whether the exchange changes or merely becomes less certain. This exercise tests whether the comparison is built from evidence and structure rather than a preselected favorite.

AI can produce a first-pass list of consequences or argue from each stakeholder's perspective. The learner should verify sources, remove invented effects, reconcile units, and explain any value judgment. Retain the prompt that generated a useful challenge, but assess the finished comparison and the human's revisions rather than rewarding volume of AI interaction.

• the deadline moves forward;
• the option becomes difficult to reverse;
• a stakeholder previously outside scope bears a new burden;
• the most uncertain estimate widens;
• one constraint becomes mandatory.

Assess observable trade-off analysis

Look for evidence in the board, source map, revisions, and explanation:

A high score should not depend on choosing the assessor's preferred option. The evidence is whether the participant constructed a fair, decision-relevant comparison and controlled unsupported claims.

• feasible choice: the comparison includes a baseline and credible alternatives after hard constraints are applied;
• comparable evidence: costs, benefits, time periods, and populations use consistent bases or disclose mismatches;
• distribution awareness: benefits, burdens, and risks are connected to affected parties rather than hidden in totals;
• uncertainty discipline: ranges, assumptions, confidence, and missing data remain visible;
• exchange clarity: the analyst states what is gained and surrendered without turning the board into advocacy;
• sensitivity: pivotal values or weights are varied to see whether the apparent order changes;
• traceability: material consequences connect to an inspected source or are labelled as judgment;
• revision quality: new evidence changes the relevant cells without rewriting the entire rationale opportunistically.

Verify the board before it informs a decision

Run four checks. First, perform a unit audit: align time periods, populations, currencies, and denominators. Second, reverse every directional statement and ask what evidence would be needed for the opposite to be credible. Third, remove the headline score, if one exists, and check whether a decision owner can still understand the exchange. Fourth, conduct a burden audit by asking who pays, waits, reviews, loses flexibility, or carries residual risk.

Then test sensitivity. Vary one pivotal estimate or weight through a plausible range and record whether the option order changes. If a small change reverses the ranking, report a fragile comparison rather than a winner. If evidence is too weak to distinguish options, the appropriate output may be a bounded information-gathering step.

Trade-off analysis structures a choice; it does not grant authority to make it. In a Two Prune assessment, the TRADE board can provide task-bounded evidence about Judgment & Synthesis, Evidence Navigation, Verification & Risk Control, and Communication & Delivery. It does not prove workplace performance outside the simulation and does not make an employment decision.

Sources

UK Government: The Green Book 2026

• UK Government: Use of multi-criteria decision analysis in options appraisal
• UK Government: The AQuA Book

Sources: [1] [2] [3]