Practical guide

AI-assisted analysis brief template and guide

An analysis brief should align the commissioner and analyst before a method or AI-generated narrative hardens into the plan. This template defines the decision, evidence boundary, analytical approach, AI-assistance rules, assurance, and delivery record. Copy it into the approved working system and adapt the depth to the consequence and complexity of the decision.

By Two Prune

An analysis brief should align the commissioner and analyst before a method or AI-generated narrative hardens into the plan. This template defines the decision, evidence boundary, analytical approach, AI-assistance rules, assurance, and delivery record. Copy it into the approved working system and adapt the depth to the consequence and complexity of the decision.

Source basis and limits

The 2025 AQuA Book describes an analytical lifecycle of engagement and scoping, design, analysis, and delivery. It calls for proportionate assurance, records of data, assumptions, decisions, limitations, and changes, and independence between analyst and assurer. The UK Government Functional Standard for Analysis defines an assurance statement as a description of assurance work and known conditions or limitations. NIST's Generative AI Profile recommends reviewing and verifying sources and citations in generative-AI outputs during measurement and monitoring. These sources support the template's controls; they do not approve a method, an AI tool, or a result for a particular organization. GOV.UK: The AQuA Book GovS 010: Analysis NIST AI 600-1: Generative AI Profile

Sources: [1] [2] [3]

Copyable analysis brief

### 1. Commission and decision

Brief ID and version: Commissioner: Lead analyst: Independent assurer: Approver: Decision owner: Decision date:

Decision question Write one choice the analysis will inform. Name the alternatives and the action that could follow.

Background and intended use Describe the operating context, affected groups, prior decisions, and why the evidence is needed now.

Out of scope List adjacent questions this analysis will not answer.

Consequence and reversibility Describe financial, legal, rights, safety, operational, reputational, and distributional consequences. State what can be reversed.

Success for the analysis Define fitness for this decision: required precision, timeliness, explanatory depth, accessibility, and assurance. Do not define success as reaching a preferred conclusion.

### 2. Evidence boundary

| Source ID | Description | Owner | Period | Population | Access or handling | Known quality issue | Intended use | | --- | --- | --- | --- | --- | --- | --- | --- | | S-01 | | | | | | | |

Record:

Never place sensitive material into an AI tool solely because it appears in the source packet. The organization's approved-tool and information-handling rules remain authoritative.

### 3. Analytical plan

Hypotheses or propositions State the propositions and credible alternatives. Distinguish descriptive, comparative, diagnostic, predictive, and causal questions.

Method Describe the population, sample, variables or coding frame, transformations, comparison, uncertainty treatment, and planned outputs. Explain why the method is proportionate to the decision.

Confounds and rival explanations List plausible causes of an observed pattern other than the preferred interpretation.

Missingness rule State how unavailable, incomplete, inaccessible, or non-applicable evidence will appear. Do not convert absence into a zero or negative finding without support.

Sensitivity and critical exceptions Define which assumptions, thresholds, subgroups, or cases will be varied. Name failures that remain decision-critical even when an average is favorable.

Change control Describe who may amend the question, data, method, or output after work begins and how the previous version is preserved.

### 4. AI-assistance plan

For each proposed use, complete a control row.

| Step | Intended AI task | Approved inputs | Prohibited action | Human verification | Evidence retained | | --- | --- | --- | --- | --- | --- | | Inventory | List supplied sources and metadata | Source register | Add outside evidence | Reconcile every row | Versioned inventory | | Exploration | Suggest patterns or anomalies | De-identified analysis table | Treat a pattern as a finding | Reproduce with approved method | Query and result log | | Drafting | Propose structure or plain-language wording | Approved results and limits | Invent a fact, citation, or causal claim | Reopen results and sources | Accepted and rejected edits | | Challenge | Generate rival explanations and failure cases | Method and draft | Make final assurance decision | Analyst and assurer adjudicate | Challenge disposition |

Add tool name, version where available, configuration, access conditions, date, and responsible person. Keep AI-generated code or transformations under the same review, testing, and version-control expectations as other analytical work.

### 5. Assurance plan

| Risk or claim | Check | Reviewer | Timing | Evidence | Failure disposition | | --- | --- | --- | --- | --- | --- | | Source completeness | Reconcile against source register | Analyst | Before analysis | Signed inventory | Repair | | Transformation | Recompute representative cases | Peer analyst | During analysis | Test results | Repair or hold | | Interpretation | Test rival explanations | Subject specialist | Before delivery | Challenge note | Narrow claim | | AI citation | Open source and locate support | Analyst and assurer | Before delivery | Claim ledger | Remove or repair | | Communication | Compare narrative with results and limits | Assurer | Before sign-off | Assurance statement | Repair |

State the independence required between analyst, assurer, and approver. For urgent work, record which checks could not be completed and the resulting use restriction instead of implying full assurance.

### 6. Delivery specification

Audience and format: Required tables, charts, artifact, or appendix: Accessibility requirements: Decision summary: Uncertainty and limitation statement: Supporting documentation: source register, data dictionary, method, code or workbook, tests, decision log, assurance statement, and version history. Approval evidence: Review or sunset date:

The final communication must distinguish source fact, analytical result, interpretation, and recommendation. It should state what the analysis cannot support and the conditions under which it should no longer be used.

  • available quantitative, qualitative, documentary, and expert evidence;
  • selection and exclusion rules fixed before analysis;
  • missing populations, fields, periods, or perspectives;
  • definitions, provenance, transformations, and linkage;
  • privacy, confidentiality, security, retention, and access controls; and
  • which evidence may be supplied to an approved AI system.

Worked brief: customer-support escalation review

A fictional service team asks whether it should change the escalation guide for complex billing cases. The initial request is “use AI to find why agents escalate too much.” The analyst rewrites the decision: should the team test a revised guide for two named billing pathways during the next service cycle?

The evidence boundary includes six months of de-identified case metadata, a stratified sample of case notes, current guide versions, quality-review decisions, and interviews with agents and escalation specialists. It excludes customer sentiment because a comparable measure is unavailable. Missing case notes remain missing; they are not coded as no issue.

The analysis is descriptive and diagnostic, not causal. It compares escalation patterns by pathway, case complexity, guide version, and evidence completeness. Rival explanations include changed case mix, new staff, access to account history, and inconsistent quality-review coding. The team preselects 60 cases for blind review and defines unresolved customer-harm risk as a critical exception.

AI may inventory approved documents, propose a coding-frame draft, and help write plain-language summaries. It may not read direct identifiers, make final coding decisions, infer customer intent, or manufacture a reason for escalation. Every proposed citation and material narrative statement is checked against the source or approved result. The analyst reproduces all calculations; an independent assurer reviews the sample, transformations, exceptions, and communication.

The first analysis finds a difference between pathways, but 18% of relevant case notes are unavailable and newer staff handle a larger share of one pathway. The final brief does not say the guide caused excess escalation. It recommends a small prospective test after improving note access and coding consistency. The assurance statement records the missingness, confounding, checks completed, and intended-use limit.

Brief quality check

[ ] The decision, alternatives, owner, date, consequence, and intended use are explicit.

  • [ ] Scope and exclusions prevent the analysis from expanding silently.
  • [ ] Source provenance, population, period, quality, handling, and missingness are recorded.
  • [ ] The method matches the type of question and treats uncertainty explicitly.
  • [ ] Rival explanations and critical exceptions are named before interpretation.
  • [ ] Every AI use has approved inputs, a prohibited action, human verification, and retained evidence.
  • [ ] Analyst, assurer, approver, and commissioner responsibilities are clear.
  • [ ] Delivery includes limitations, assurance state, supporting documentation, and a review or sunset date.

Adaptation guidance

For a quick, low-consequence analysis, shorten the brief but preserve the decision, source boundary, method, AI controls, core checks, and limitations. For consequential or complex work, add a data-protection assessment, ethics review, specialist assurance, preregistered analysis, model documentation, accessibility testing, or legal review as required by the organization and jurisdiction.

Version the template locally. Record changes to definitions, source rules, method, and assurance rather than overwriting the approved brief.

Two Prune boundary

Two Prune can assess how a participant frames an analytical commission, navigates supplied evidence, uses AI as a bounded work partner, verifies calculations and claims, handles uncertainty, and communicates a decision-ready deliverable. It reports observed assessment evidence. It does not approve an analysis, validate an AI model, certify compliance, replace independent assurance, predict workplace performance, or make employment decisions.

Sources

GOV.UK: The AQuA Book

  • UK Government Functional Standard GovS 010: Analysis
  • NIST AI 600-1: Generative Artificial Intelligence Profile

Sources: [1] [2] [3]

Sources

  1. 1.GOV.UK: The AQuA Book · UK Government
  2. 2.GovS 010: Analysis · UK Government
  3. 3.NIST AI 600-1: Generative AI Profile · National Institute of Standards and Technology