Audience adaptation in AI-enabled work: preserve the decision, change the delivery

Separate the evidence core from the delivery layer

Create one evidence core before producing audience-specific versions. It should contain the source references, calculations, assumptions, material uncertainties, and approved conclusions. Then define what each audience needs from that core.

The GOV.UK Service Manual recommends learning about users and their needs, basing those needs on research rather than assumptions, and focusing on what a person is trying to accomplish. Use the same discipline for workplace communication: describe the audience's job, decision, knowledge, constraints, and likely questions before choosing format or tone.

Sources: [1]

Use the RANGE adaptation map

For each audience, record five elements:

The map should change emphasis, explanation, and navigation. It must not change a number, source, confidence level, or recommendation merely because one audience prefers a simpler answer. Keep a traceable link from every version to the same evidence core.

GOV.UK guidance on writing for user interfaces recommends clear language, short and direct copy, important words first, language users recognize, accessible link text, and content that works for people who scan. Those principles can improve workplace artifacts too, but the correct structure still depends on the task and audience.

• Role: Who receives the work, and what authority or responsibility do they have?
• Action: What must they understand, decide, approve, challenge, or do next?
• Necessary evidence: Which facts, calculations, examples, and source links are required for that action?
• Gaps and risk: Which uncertainty, limitation, dissent, or missing evidence would be material to this reader?
• Expression: Which structure, vocabulary, depth, channel, and accessibility choices make the work usable?

Sources: [2]

Worked artifact: one synthetic service review, three audiences

A fictional service team reviews a monthly support backlog. The evidence core includes case volume, age bands, repeat-contact rate, two data-quality gaps, a staffing constraint, and source links. The team's first AI-assisted draft sends the same twelve-page report to everyone.

The analyst uses RANGE to produce three linked views:

| Audience | Action | Necessary evidence | Material gap | Expression | | --- | --- | --- | --- | --- | | Executive sponsor | Choose whether to fund a four-week intervention | Trend, impact range, option costs, decision deadline | One category has incomplete tagging | One-page decision note with three options | | Operations lead | Plan the intervention | Queue breakdown, staffing calendar, case examples, dependencies | Supplier schedule is provisional | Working plan with owners, dates, and exception list | | Quality analyst | Verify the diagnosis | Row-level definitions, formulas, exclusions, source versions | Twelve cases lack a stable category | Reproducible workbook and validation log |

The executive note does not remove the tagging gap. It expresses the limitation in decision terms: the estimated affected volume is a range until the twelve cases are reviewed. The operations plan links the same cases as an exception task. The analyst view retains the detailed rows and formula checks.

During review, the executive note says the intervention "will reduce repeat contacts." The evidence core supports only a testable hypothesis. The statement is revised to say that the intervention is intended to test whether clearer routing reduces repeat contacts over four weeks, with a defined comparison and stop rule.

This example is synthetic. It does not establish a service outcome, staffing requirement, or customer result.

Run a cross-version integrity check

Compare every audience version against the evidence core before release. Check that:

Then test usability with a representative reader. Ask them to state the decision, supporting evidence, uncertainty, and next step without verbal coaching. A short document that prompts the wrong action is not successfully adapted.

• numbers, units, dates, and definitions match;
• material caveats appear wherever they could change a decision;
• recommendations use the same confidence level;
• links identify the source rather than saying only "click here";
• the reader can distinguish observation, interpretation, and proposal;
• no version grants authority that the recipient does not have;
• updates propagate to every dependent version.

Assess audience-adaptation behavior

A controlled assessment can provide one evidence pack and three audience briefs. Observable evidence includes whether the participant identifies different decisions, maintains a shared evidence core, selects appropriate detail, preserves caveats, changes structure and language, and catches contradictions across versions.

Do not reward simplification by itself. Strong adaptation may add detail for an accountable reviewer or slow a reader down before a consequential action. Report only the behavior and artifacts observed in the scenario. Two Prune provides task-bounded assessment evidence for human interpretation and does not make employment decisions.