Reports that hold up
- reports
- evidence hygiene
Reports fail in review, not in drafting — and the failures repeat: the reader’s decision is missing, claims arrive unsourced, numbers do not recompute, confidence outruns evidence. AI accelerates drafting, which means it accelerates all four unless the structure forces them out.
The report skeleton
Five sections, in this order, every time:
- Purpose and decision — what the reader should do after reading. If this line is missing, the report is a diary.
- Findings as claims — one-sentence claims, each with its evidence attached. Not narrative — claims.
- Numbers with working — every load-bearing figure, its unit, and the calculation that produced it.
- Risks and caveats — what could make the findings wrong, and what you do not yet know. This is the section AI omits by default.
- Next steps — owners and dates. A recommendation without an owner is a wish.
Evidence hygiene with AI
Weak prompt
Write a report on our Q3 performance, make it impressive and complete
Stronger prompt
Draft the skeleton only: (1) purpose and decision, (2) five findings as one-sentence claims — mark any needing a source with [SOURCE?] — (3) risks section with at least three caveats, (4) next-steps table. Use only the figures I list below; do not add numbers.
Why the stronger one works
- “Impressive and complete” invites fluent filling; the skeleton request invites structure with visible gaps.
- The [SOURCE?] flags turn missing evidence into something you can see and chase — instead of prose that reads finished.
- “Do not add numbers” removes the single biggest failure mode of AI-assisted reports in one clause.
A bad example
Ask: “Write a competitive analysis from these notes.”
You get a confident narrative with smooth generalities, undated competitor claims and a tidy conclusion — and no way, on the surface, to see which sentences are load-bearing facts and which are the model being agreeable.
A better example
Ask: the skeleton prompt above, then fill the evidence slots yourself from your notes, recomputing figures as you go and leaving the [SOURCE?] flags until each one is honestly resolved.
The document now has the two properties review looks for: every claim traceable, every gap visible.
Why it works
A review-proof report is engineered, not written. Decision at the top, claims that are checkable one by one, numbers that show their working, caveats that were written before someone else finds them. AI remains genuinely useful inside this frame — as an outliner, a rewriter, a consistency checker — because the structure it operates within is yours. The caveat entry is the spirit of section four: say what you do not know before the review says it for you.
Practice
Structure the status report
Project Atlas is two weeks behind its integration milestone, and the steering committee meets Friday. You have raw notes: mixed progress, one slip, a budget that is still fine, and an unconfirmed rumour that a key vendor may deprioritise you. You want an AI assistant to help — but a fluent draft that smooths over the slip would be worse than useless.
Draft the report skeleton you would ask the AI to produce: purpose and decision, findings as sourced claims, numbers with units, risks with caveats, and next steps with owners and dates. This is the text you would paste as your structure brief.
A model answer (self-assessment — not machine-graded)
Purpose and decision: The committee is asked to approve keeping the October integration date with reduced scope (not moving the date). Findings (claims, one per line): (1) Integration milestone slipped by two weeks due to vendor API changes [SOURCE? — link the change log]; (2) Reduced-scope option keeps the October date with three of five planned features [SOURCE? — scope sheet draft]; (3) Budget remains within the approved envelope, at 61% spent against a 70% planned burn as of 20 September 2026 (actuals from the finance tracker; planned burn from the approved plan — growth rate recomputed, working attached). Risks and caveats: (a) unconfirmed report that the vendor may deprioritise our integration — not yet verifiable, mitigation: written commitment requested by 25 September; (b) scope reduction removes the bulk-export feature customers asked for — churn impact unknown; (c) two-week slip compressed QA to four days. Next steps: written vendor commitment — owner: PM, due 25 September; scope decision record — owner: programme lead, due 30 September; QA plan for the compressed window — owner: tech lead, due 28 September.
Hint
The committee's job is a decision — the skeleton should say which one before anything else.
The vendor rumour is unconfirmed: that is exactly what the risks section exists for. Name it as unconfirmed rather than omitting or asserting it.
Why this is the answer
The skeleton disciplines the whole engagement: the decision is named first, findings are discrete claims you can source or flag one by one, the arithmetic is annotated for recomputation, the awkward risks are named while they are still cheap, and every next step has a name and a date. An AI assistant can produce this structure in seconds from your notes — and because the structure makes gaps visible, you can check it in minutes. Compare that with a fluent narrative draft, where the slip would be buried in paragraph three and the vendor rumour would either vanish or harden into a fact.
Transfer
- Weekly updates: purpose line first, claims with links, risks section even when it feels unnecessary.
- Proposals: the decision and the ask belong on page one, not page nine.
- Any AI-heavy draft: run the five-section skeleton over it — the gaps surface immediately.
- Numbers: recompute before they enter the skeleton; checking figures gets its own pass.
Key takeaways
- Reports are engineered: decision first, claims with evidence, numbers with working, caveats before review finds them.
- Draft the skeleton with AI, fill evidence yourself, keep [SOURCE?] flags until truly resolved.
- The risks section AI omits by default is the one reviewers respect most.
Next
Next: Support replies that de-escalate — the hardest small document in any business, where tone and promises both carry risk. The rubric entry covers checking work like a reviewer, and unsupported claims covers what the skeleton is designed to expose.