How to develop and assess structure and hierarchy in AI-enabled work

Use established content-structure principles

The Office for National Statistics advises writers to frontload important information, use a logical heading hierarchy, make headings descriptive, and keep paragraphs focused on one subject. UK government accessibility guidance also calls for meaningful subheadings, a logical heading order without skipped levels, properly identified table headers, and link text that explains its destination. Together, these practices give a concrete starting point for AI-enabled reports, briefs, and recommendations.

Build a SIGNAL outline before drafting

SIGNAL is an outlining method for decision-facing work. Each step produces something a reviewer can inspect.

This outline is intentionally independent of any AI-generated draft. A worker can ask an AI assistant to propose groupings or identify missing transitions, but the worker should decide the hierarchy against the audience, purpose, and evidence.

• Stakeholder decision: Write the decision, action, or understanding the artifact must enable in one sentence.
• Information priority: Rank the facts the reader needs as essential, supporting, or reference-only.
• Group related evidence: Put facts that answer the same reader question together, regardless of where they appeared in the source pack.
• Name sections descriptively: Use headings that state their function or conclusion, not labels such as “Background 2.”
• Arrange from main point to detail: Put the requested answer and its limits before chronology or process description.
• Link dependencies: Point clearly to assumptions, appendices, source material, and decisions that depend on other decisions.

Diagnose hierarchy with a heading-only test

After drafting, hide the body and read only the title and headings. A reviewer who sees that outline should be able to predict the document's argument and find the requested decision. If two headings operate at different levels but look equivalent, the hierarchy is misleading. If a heading contains several unrelated ideas, the section probably needs to be divided or narrowed.

Then run a paragraph-role test. Give each paragraph one label: answer, evidence, interpretation, limitation, action, or reference. Consecutive paragraphs with no clear role often indicate a source dump. Repeated answers at several levels suggest that the draft was expanded without a stable outline. An appendix that contains the only explanation for the main recommendation suggests that essential information has been buried.

AI assistants are good at producing fluent transitions, but fluency can conceal a weak order. The tests above judge navigability and dependency directly rather than rewarding surface polish.

Worked example: restructuring a synthetic options memo

Consider a synthetic operations team responding to repeated delays in an internal approval process. The evidence pack contains timestamps, three stakeholder messages, a process map, and two possible changes. No real incident, customer, or Two Prune outcome is represented here.

The first AI-assisted draft follows the order in which files were opened. It starts with a long history, alternates between evidence and recommendations, repeats the delay figure in four sections, and places the main choice in the final paragraph. Every sentence may be readable, but the memo does not help an operations lead choose an action.

Using SIGNAL, the worker creates this outline:

| Level | Heading | Reader function | | --- | --- | --- | | H1 | Choose a two-week approval triage trial | States the decision requested | | H2 | Current delays are concentrated at one handoff | Gives the principal evidence and its limit | | H2 | Two options address different failure points | Frames the comparison | | H3 | Option A assigns a daily triage owner | Explains mechanism, cost, and dependency | | H3 | Option B changes the approval threshold | Explains mechanism, cost, and dependency | | H2 | Trial Option A with defined stop conditions | Presents the recommendation and why it is reversible | | H2 | Review after ten working days | Names measures, owner, and next decision | | Appendix | Timestamp definitions and calculation | Preserves audit detail without interrupting the decision path |

The revised memo opens with the requested choice, keeps comparable option facts at the same level, and moves calculation detail to an appendix while retaining a direct link to it. The heading-only test reveals the complete decision path. The paragraph-role test reveals whether each section contains evidence, interpretation, or action rather than a mixture with no signal.

Assess structure through a constrained transformation

To assess the skill, give a participant an evidence pack with information in an intentionally inconvenient order and request a defined deliverable for a named audience. Require the participant to retain important qualifications while producing a usable hierarchy. Suitable evidence includes the submitted artifact, the outline or notes, opened source records, revisions, and any AI interactions available within the assessment.

Score observable behavior such as:

Do not award credit merely because a document contains many headings, uses a familiar template, or resembles a polished AI response. Those features can exist without a coherent information hierarchy. Also avoid treating one preferred memo shape as universally correct. The appropriate structure changes with the reader and the decision.

• audience alignment: the top level answers the named reader's actual question;
• prioritization: essential information appears before supporting and reference material;
• grouping: sections contain related evidence and avoid repeated fragments;
• level consistency: parallel choices or findings appear at the same heading level;
• descriptive navigation: headings communicate what a section contributes;
• dependency clarity: assumptions, limits, owners, and next steps are connected to the decisions they affect;
• revision quality: the participant notices and repairs buried conclusions or mixed levels.

Verify the artifact at three resolutions

Check the artifact first at page level: can a reader identify the purpose, main answer, and requested action quickly? Next, check it at section level: does each heading answer a distinct question, and do sibling sections operate at comparable levels? Finally, check it at paragraph level: does each paragraph do one recognizable job, with evidence and qualifications placed near the claims they constrain?

Confirm that headings are implemented as real headings, levels are not skipped, tables have identified headers, and links have meaningful text. Ask a fresh reviewer to locate one key finding, one limitation, and the next action using only the visible hierarchy. Record any navigation failure and revise the outline, not just the transitions.

Structure and hierarchy are task-dependent communication behaviors. A strong result in one simulation does not prove universal communication ability. In a Two Prune assessment, the observed artifact and work process may contribute evidence about Judgment & Synthesis and Communication & Delivery. They remain bounded assessment signals, not mature benchmark claims, promises of outcomes, or decisions about employment.

Sources

Office for National Statistics guidance on structuring content

• UK government guidance on publishing accessible documents

Sources: [1] [2]