How to communicate an analytical recommendation with AI

Write the decision sentence first

Start the memo with a sentence that names the owner, choice, timing, and constraint:

> The operations lead should keep the current routing rule for the September pilot and test the proposed rule on a controlled sample before wider use because the estimated time saving is uncertain and the error cost is asymmetric.

That sentence becomes the test for every chart and paragraph. If a result does not change confidence in the choice, explain why it is included or remove it.

Build an evidence-to-action table

Before drafting prose, place the decision evidence in four columns:

| Finding | Magnitude | Uncertainty or limitation | Decision consequence | | --- | --- | --- | --- | | Proposed routing reduced median handling time in the pilot | 6 minutes per case | 42 cases; one team; two-week window | Worth a controlled test, not a full rollout | | Misrouted cases increased | 3 additional cases | Small count; categories changed mid-pilot | Add manual review for high-cost categories | | Workload varied by day | Peak day was 1.8× the quietest | Scheduling data incomplete for two shifts | Stratify the next sample by day and shift |

The figures above are synthetic. Their purpose is to show how magnitude and limitations travel with the recommendation.

Ask AI to challenge each row: “What would a skeptical decision-maker need to know before acting on this finding? Identify missing denominators, time windows, definitions, and plausible alternative explanations.” Resolve the questions in the analysis or label them as limitations. Do not ask the model to invent the missing denominator.

Make uncertainty operational

Avoid a generic paragraph saying that “more data is needed.” State what uncertainty could reverse the recommendation and how the next action reduces it. In the example, a larger average time saving would not justify rollout if the extra misroutes create material customer harm. The next test therefore measures both speed and error severity, with a predefined stop condition.

The UK Government Analysis Function’s guidance on communicating quality, uncertainty, and change emphasizes giving users enough information to judge the strength and fitness of evidence. That principle is broadly useful, but the exact disclosure standard depends on the organization and decision.

Draft in layers

Use a one-page structure with three reading depths:

Let AI produce a plain-language draft from the completed table, then compare each factual sentence with the analysis artifact. Highlight causal language, universal claims, precise numbers, and adjectives such as “significant” or “proven.” Each must have a defined meaning and source.

• the first paragraph states the recommendation and reason;
• the evidence table shows magnitude, uncertainty, and consequence;
• an appendix provides definitions, method, calculations, and source lineage.

Worked memo excerpt

Recommendation: Continue the current rule while running a four-week controlled test of the proposed routing rule.

Why: The pilot suggests a six-minute reduction in median handling time, but it also produced three additional misroutes. The sample covers one team and 42 cases, so it does not establish how the rule will perform across shifts or categories.

Action: Randomly assign eligible cases within each shift, retain manual review for high-cost categories, and stop the test if severe misroutes exceed the agreed threshold. Report both handling time and error severity.

This excerpt does not hide the result, but it also does not ask the decision-maker to treat an early pilot as settled evidence.

Verification pass

Before sending, check that the recommendation names a real decision; every number has a source, definition, denominator, and period; uncertainty is linked to action; charts use consistent populations and units; alternatives and constraints are visible; and the appendix can reproduce the central calculation. Finally, ask a colleague unfamiliar with the analysis to state the recommended action and the strongest reason not to take it. If either is unclear, revise.

Sources and claim map

All operational data in the worked example is synthetic. This resource does not claim validated performance outcomes.

• UK Government Analysis Function, Communicating quality, uncertainty and change: supports transparent communication of evidence quality and uncertainty so users can judge fitness for use.
• UK Government Analysis Function, Communication resources: provides the broader public-sector analytical communication context.
• HM Treasury, The Aqua Book: used as a quality-assurance reference for analytical models and analysis.

Sources: [1] [2] [3]