AI-assisted consulting analysis: trace each recommendation through its assumptions

Define the decision before building the story

Write the decision, decision owner, timing, alternatives and constraints in plain language. “Improve performance” is not a decision. “Choose whether to pilot one of two intake processes in a single regional team next quarter” is testable and bounded. Separate the client's stated objective from the consultant's proposed mechanism. AI may organize supplied evidence, but it should not manufacture agreement among stakeholders or convert an aspiration into a measured baseline.

The UK Government's AQuA Book describes proportionate analytical quality assurance, verification, validation and explicit treatment of uncertainty. HM Treasury's Magenta Book distinguishes implementation, outcomes and impact and emphasizes theory of change, assumptions and alternative explanations. These are public-sector analytical guides, not mandatory rules for a private consulting engagement. They are useful here as primary examples of tracing a decision through evidence and uncertainty.

Sources: [1] [2]

Build a recommendation-assumption map

Imagine a fictional service organization considering a new request-intake form. The pack contains four interview notes, a queue spreadsheet and a previous pilot summary. The AI draft recommends rolling out the form nationally because “the pilot cut turnaround time by 30 percent.” The spreadsheet shows that the pilot team's median turnaround fell from ten to seven days, but the periods contain different request mixes. One interview says the team also added two temporary reviewers. The pilot summary records the form launch but does not isolate which change caused the difference.

Create a map with one row per recommendation link: proposed action, intended mechanism, supporting observation, assumption, alternative explanation, decision sensitivity, verification step and owner. The observed median change belongs in one cell with its exact periods and denominator. “The form caused the change” belongs in the assumption column, alongside staffing and case-mix alternatives. The map might still support a limited pilot with measurement, but it does not support a causal national-rollout claim.

Ask AI to draft a short option comparison only after the map exists. Require at least three options: keep the current process, run a bounded pilot, or expand immediately. Record what would have to be true for each option to make sense. If the model writes that the bounded pilot is “low risk,” replace the label with named operational risks and mitigations. If it adds a financial benefit absent from the pack, delete it and log the unsupported claim.

Challenge the recommendation with a reversal test

Run a reversal test: what single new fact would change the preferred option? If evidence that temporary reviewers handled most of the improvement would reverse the recommendation, staffing is a critical unresolved assumption. Add a synthetic follow-up showing that urgent cases rose in the comparison period. Does the analysis explain how the mix affects the median, or does it preserve the original headline? Add an interview from a region with a different approval sequence. Does the proposal identify a localization question or assume one workflow everywhere?

The final slide should pair the recommendation with its evidence limit and next learning step. A defensible statement could be: “The observed pilot period had a lower median turnaround, but the supplied evidence does not isolate the form from staffing and case-mix changes. Run a time-bounded regional pilot with agreed measures before deciding on wider adoption.” The map remains behind the slide so a reviewer can inspect every link.

Keep analysis and assessment boundaries explicit

Real consulting work requires current client data, confidentiality controls, domain expertise and accountable client decisions. This synthetic exercise does not forecast returns, prove causality or establish that a pilot is appropriate for any organization. Two Prune could observe whether a participant separates observation from mechanism, surfaces alternatives and calibrates a recommendation. Its report would describe task-specific assessment evidence for human interpretation, not make a business or employment decision.