Delivery completeness in AI-enabled work: close the decision, not just the document

Define completion before drafting

Translate the request into acceptance conditions before using AI to draft. Record the intended reader, decision or action, required evidence, required format, deadline, constraints, approver, and downstream handoff. Separate mandatory conditions from optional improvements.

The U.S. Government Accountability Office's 2024 Government Auditing Standards define a complete performance-audit report as one containing sufficient appropriate evidence for its objectives, without omitting significant relevant information. The standards also call for stating what was and was not done and describing important data limitations or access constraints. A normal workplace deliverable is not a government audit, but these elements provide a useful completeness test: evidence, scope, and limitations must match the stated objective.

Do not ask an AI assistant only to “make this complete.” Give it the acceptance conditions and ask it to identify missing items without inventing them. Missing evidence should become an explicit gap or follow-up task, not plausible filler.

Sources: [1]

Use the CLOSE acceptance matrix

Build a five-column matrix before release:

An exit test should be specific. “Reviewed” is not enough. A stronger test might be: the finance owner can reproduce the option totals from the linked workbook; the approver can identify the requested decision and deadline; or every mandatory requirement has a source, a gap label, or an assigned follow-up.

NIST's AI Risk Management Framework Core emphasizes documenting context, knowledge limits, intended use, human oversight, roles, and responsibilities. Apply that principle narrowly to AI-assisted work: record which parts were generated or transformed with AI where relevant to the workflow, who checked the result, and which unresolved issues still require human judgment. This does not turn a routine document into an AI-risk assessment. It prevents the handoff from hiding ownership or limits.

• Criteria: What must the deliverable answer, contain, or enable?
• Links and evidence: Which source, calculation, record, or observation supports each material point?
• Owner and action: Who decides, approves, executes, checks, or receives the next handoff?
• Scope and limits: What period, population, assumptions, exclusions, uncertainty, or access constraint applies?
• Exit test: What observable check shows the artifact is ready for this stage?

Sources: [2]

Worked artifact: a synthetic supplier-options memo

A fictional operations team asks for a recommendation among three document-processing suppliers. The first AI-assisted memo contains a fluent comparison and chooses Supplier B. It is not complete. The prompt omitted the security questionnaire, implementation capacity, contract notice period, currency date, and who can approve the spend.

The analyst builds this CLOSE matrix:

| Criteria | Links and evidence | Owner and action | Scope and limits | Exit test | | --- | --- | --- | --- | --- | | Compare three-year cost | Pricing sheets and calculation workbook | Finance checks totals | Exchange rate dated 15 September; usage forecast is an assumption | Finance reproduces all totals | | Describe security status | Completed questionnaires and exception log | Security owner accepts or rejects exceptions | Supplier C has not answered two controls | Every open control has an owner and due date | | Test implementation fit | Team-capacity calendar and supplier plan | Operations confirms start window | Two engineers are unavailable in November | Plan fits documented capacity or shows a gap | | Request a decision | One-page options table | Budget owner chooses an option by 30 September | Recommendation remains conditional on security closure | Approver can state choice, conditions, and deadline |

The revised memo recommends Supplier B conditionally, links the cost model, shows the unanswered controls, names the security owner, and includes a fallback if the controls remain open. It also states that supplier performance has not been tested in production. The document becomes complete for the decision even though some evidence remains unresolved, because the gaps and response path are explicit.

This example is synthetic. It does not establish supplier quality, savings, security, or an organizational outcome.

Run a reverse-handoff check

Give the artifact to a reviewer who did not help create it. Without verbal explanation, ask the reviewer to identify the requested decision, evidence, material limitations, responsible owners, deadline, and next handoff. Then trace every answer back to the CLOSE matrix.

Also inspect attachments and links, not only prose. Confirm versions, units, formulas, permissions, file names, and dates. Check that appendices promised in the text exist and that the main document does not contradict them. If a required item is unavailable, label it, explain its consequence, and assign a next action. Do not hide the gap in a footnote or let AI fabricate a substitute.

Assess delivery-completeness behavior

A controlled assessment can provide an assignment, evidence pack, format constraint, and incomplete draft. Observable evidence includes whether the participant extracts acceptance conditions, distinguishes mandatory from optional material, traces claims to sources, exposes gaps, assigns ownership, checks attachments, and tests the handoff from the reader's perspective.

Do not reward length, visual polish, or a high checklist count by themselves. A concise artifact can be complete, while a long artifact can omit the only fact needed for a decision. Report the behavior and work product observed in the scenario, including missing evidence and confidence limitations. Two Prune provides task-bounded assessment evidence for human interpretation and does not make employment decisions.