InfoSecured tools
AI Assurance Evidence Review Kit
Build a clear record of what your AI review found—and what happens next.
Connect claims, risks, controls, evidence, tests and decisions in one Excel workbook. Use it to review an AI system, assess a supplier, investigate a gap or prepare an assurance discussion.
Put it to work
Start with the decision your team needs to make.
The kit is designed for AI governance, security, technology risk, audit, model risk and vendor risk teams. Choose questions relevant to the system and the consequence of getting the decision wrong.
- Review a system or change
- Define the proposed use and release. Map material failures to safeguards, inspect evidence and record the conditions for proceeding.
- Assess a supplier
- Request specific artifacts, trace their evidence roots and identify claims that still need local testing.
- Test human intervention
- Document what a reviewer saw and decided, then verify whether the hold, correction or escalation changed the outcome.
- Prepare an assurance discussion
- Bring open findings, overdue actions, failed tests and the decision basis into one reviewable record.
Inside version 2.2.1
Blank working sheets. Guidance where you need it.
Enter your own organizational information from the first review. The worked example has its own sheet, separate from your records and live summary.
- Set the scope
- Start Here · Review Brief
Purpose, system boundary, versions, criteria and decision rights. - Define the review
- Claims · Risks & Controls
Assertions, acceptance criteria, library prompt links, reviewer names and dates, and accountable safeguards. - Examine what happened
- Evidence · Tests · Human Oversight
Artifacts, scope fit, evidence roots, observed results and intervention timestamps with verified effects. - Act and decide
- Findings · Decisions · Review Summary
Correction, verified closure and decision records displaying linked Finding, Evidence and Test IDs. - Use the references
- Field Guide · Review Library · Worked Example · Sources & Method
Field-by-field examples with required and conditional inputs, 24 adaptable prompts and a connected review example.
| Field | What to record |
|---|---|
| Review purpose | State the decision this review will support. |
| Version and configuration | Record model, prompt, retrieval index, tools and relevant settings. |
| Human authority | Identify hold, override, escalation, restart and final decision powers. |
| Earlier assessments reused | Identify the assessment, scope and evidence roots; specify fresh checks. |
| Reassessment triggers | List changes or incidents that require a new decision. |
The same Claim ID connects your controls, evidence, tests, findings and decisions.
Built for working in Excel: editable cells, dropdowns, filters, frozen identifiers, worksheet links and formula-driven completion, reference and timing checks. Each working register includes 100 prepared rows. Decisions includes checked finding links and Evidence/Test references, with their role or relevance to the decision. Linked IDs appear beside claim-level and overall decisions.
Review prompts for current AI workflows
Scope and assessment reuse; data quality and performance; prompt injection; retrieval permissions; information disclosure; output handling; tool authorization; approval, cancellation and recovery; supplier changes; evidence provenance; reviewer judgment; logs; fallback; resource limits; incident response; retention.
Each prompt connects a claim, possible failure, control objective, suggested test and evidence request. Adapt the selection to predictive models, LLMs, retrieval systems and agents.
Follow one review
The reviewer selected Hold. Did the action stop?
In the workbook’s fictional service-desk example, a retrieval assistant proposes an access grant. A reviewer approves the recipient, resource, permission and expiry before execution. The review examines whether cancellation also reaches queued actions and retries.
- Claim
A hold prevents a grant before execution.
The acceptance criterion includes cancellation while queued and after a connector failure.
- Evidence
The report covers only part of the workflow.
Pre-dispatch tests passed. Retry behavior was excluded. A supplier memo repeats the same underlying results.
- Test
The retry still grants access.
The reviewer’s hold is recorded, but its downstream effect fails. The finding identifies the missing state check.
- Decision
Restrict execution until the correction is verified.
The owner selects draft-only use, assigns the fix and requires a retest covering the original failure.
Keep the intention and the effect in the same record.
Human Oversight distinguishes the chosen action from its verified outcome. Evidence roots show when two reports depend on one test. Findings and Decisions carry the unresolved question through correction and reassessment.
The example also transfers the approach to AML alert prioritization: define review coverage, inspect sampling and label delays, and record the decision conditions.
Your first review
Choose one system and one consequential claim.
Start with a manageable review boundary. Expand the record as new claims, evidence and findings become relevant.
Download version 2.2.1Excel .xlsx · 183 KB · Released July 31, 2026
- Set the context. Open Start Here, then complete Review Brief with the use, version, decision owner and applicable criteria.
- Make the claim testable. Open Field Guide and Worked Example before your first review. Choose a library prompt, define acceptance criteria and map the safeguard.
- Gather and challenge the evidence. Record scope and origin, perform the tests and verify intervention effects.
- Assign action and record the decision. Use Findings for correction and closure. Record claim-level conditions in Decisions and the overall disposition, authorized scope and reassessment triggers in Review Summary.
Working with organizational records: store and share the workbook under your organization’s access and retention rules. Link to detailed artifacts in the appropriate evidence repository.
Preparing a review snapshot: set a fixed snapshot date in Review Brief, resolve completion checks and use Review Summary to inspect outstanding work.
Method and development
Evaluate evidence against the claim you need to rely on.
Review scope, integrity and support explicitly. The workbook includes field examples, reviewer attribution, finding links and an overall disposition with its authorized scope. The summary shows where work remains. The decision record preserves the reasoning and the decision owner’s conditions.
The design draws on NIST’s AI risk guidance, OWASP’s AI security guidance, GAO’s accountability framework and research on human–AI performance. InfoSecured’s practical synthesis adds evidence roots, assessment reuse, downstream effects and reassessment. Explore the AI assurance approach.
Sources and scope
- NIST AI RMF 1.0: context, ownership, measurement and ongoing risk management. NIST is updating version 1.0.
- NIST Generative AI Profile: risks and actions relevant to generative AI.
- OWASP LLM Top 10, 2025 and AI Agent Security Cheat Sheet: security failures, authorization and testing.
- GAO AI Accountability Framework: governance, data, performance and monitoring.
- DSIT Introduction to AI Assurance: evaluating and communicating evidence about AI claims.
- Vaccaro, Almaatouq and Malone, 2024: human–AI performance varies across tasks; test the actual workflow.
The workbook contains the full inventory. Bank model-risk guidance supports the statistical-model example within its stated scope; the cited 2026 guidance excludes generative and agentic AI. Fields also connect to developing CEDAR and RISC research, whose status and role are explained in Sources & Method.
Where is the user guidance?
Start Here introduces the review sequence. Field Guide provides 157 field entries with examples and required, conditional, optional or calculated inputs. Worked Example follows a connected review. Sources & Method records definitions, sources and migration instructions.
All four are included in the workbook. Each working sheet links to its relevant Field Guide entries.
How do I add records and keep the links working?
Each working table includes 100 prepared rows. Add rows inside the relevant Excel table, copy the formula cells and validation from an existing row, and check the new row’s completion message. In Decisions, insert inside the table so the link tables below move with it.
Use one stable ID per record. Finding links connects a decision to a Finding ID. Decision support links the decision to an Evidence or Test ID with its relevance and limits. Use the Overall decision ID from Review Summary for links to the overall disposition.
What changed from version 2.2?
Version 2.2.1 clarifies wrong-type reference messages, improves completion counting and highlights outstanding completion issues. Decision link shortcuts appear in the first visible columns, guidance links open the relevant fields, and Field Guide entries align consistently.
The review structure remains fourteen worksheets with blank working registers and a separate worked example.
What changed from version 2.1?
Linked Finding, Evidence and Test IDs now appear beside claim-level decisions and the overall disposition. Each Evidence or Test reference records its relevance and limits. Completion checks flag unknown IDs, wrong record types, duplicate references and claim mismatches.
Field Guide adds clearer examples for shared evidence roots, scope mismatch, test limitations, reviewer judgment, remaining uncertainty and accepted risk. Start Here directs first-time users to field examples and the complete worked review.
What changed from version 2.0?
Field Guide explains every working field with an example and its required, conditional, optional or calculated status. Claims records the source prompt, reviewer and review date. Review Summary holds the overall disposition, basis, authority, scope and review date.
Human Oversight distinguishes the oversight record ID from the overall review ID and records proposal, action and effect-confirmation times. A finding-link table connects each decision to specific findings and checks the references. Summary labels link directly to the relevant sheets.
What changed from version 1.0?
Working records start blank. The example lives on its own sheet. Claims, tests, findings and decisions have distinct records; scope and dependencies are visible; human review includes verification of the effect. The library adds security and operational questions for LLMs, retrieval systems and agents.
Follow the migration instructions in Sources & Method to transfer existing records and reassess their current scope. Version 2.2.1 includes worksheet guidance; the earlier v1.0 PDF describes the historical workbook.
Version and file verification
AI-Assurance-Evidence-Review-Kit-v2.2.1.xlsx
Version 2.2.1 · July 31, 2026 · 186,994 bytes
SHA-2569e20629e2aaad836fd2ff006e8c7498b111501964e91e35414e34df6d3e6cb47