# AI Decision Log Template and Audit Readiness Checklist Everything to reconstruct a single AI decision about a single person months later. A decision-log schema by decision type, a versioning and change policy, a retention schedule tied to the underlying legal regime, and an audit-readiness checklist. **Who this is for:** The owner of a system that makes or informs decisions about individuals, and the auditor who has to trace one. Source playbook: https://aigovernance.com/playbook/ai-decision-auditability --- ## AI decision log schema _The fields logged at the decision level, not the system level. Adjust the type-specific rows to your use cases._ ### Template | Field | Required | Notes | |---|---|---| | decision_id | Yes | Unique, referenced in the outcome record | | timestamp | Yes | With timezone | | subject_ref | Yes | Pseudonymous ID linking to the individual | | system_id and model_version | Yes | Exact version that produced the output | | inputs | Yes | The feature values or prompt actually used, or a hash plus a retrievable copy | | output | Yes | Score, class, or text | | confidence / score | Yes | Where the model produces one | | explanation artifacts | If available | Top features, rationale string | | human_reviewer | If applicable | Identity of the person who reviewed | | human_decision | If applicable | Accept, override, or modify, plus reason | | final_outcome | Yes | The decision communicated to the subject | ### Worked example | Field | Value | |---|---| | decision_id | dec-2026-0091823 | | timestamp | 2026-09-02T10:14:07+01:00 | | subject_ref | appl-7f3c9 | | system_id / version | resume-screener / 2.4.0 | | inputs | structured feature vector v2.4 (stored); raw CV hash sha256:9c1f... | | output | score 2 of 5 | | confidence | 0.71 | | explanation | top features: years_experience (low), skill_match (medium) | | human_reviewer | recruiter u-4471 | | human_decision | override to "advance"; reason: portfolio not captured by the model | | final_outcome | advanced to phone screen | ### Acceptance criteria - Logging is at the decision level: one record per decision, not per batch or per day. - The exact model version is captured on every record. - Inputs are recoverable, either stored directly or via a hash plus a retained copy. - Where a human reviews, their identity, decision, and reason are logged. --- ## Model versioning and change policy _So a logged model_version always resolves to a known model. Short; defer detail to the model registry kit._ ### Template > One page. - Every change to weights, prompt, inference parameters, or retrieval data creates a new version with a changelog. - The version string in the decision log matches a version record in the registry. - Version records are retained at least as long as the decisions made under them. - A changelog entry states what changed, why, the evaluation delta, and who approved it. ### Worked example **Version log excerpt, resume-screener:** - 2.3.1 (2026-06-10): capped tenure weight after v2.3 bias finding. Approver: R. Nkemelu. - 2.4.0 (2026-08-30): base model VendorLM-3 to VendorLM-4. Adverse-impact ratio 0.85 to 0.88. Approver: R. Nkemelu. Decisions logged with model_version 2.3.1 remain resolvable to this entry for the full retention period. ### Acceptance criteria - Every model_version value that appears in the decision log resolves to a retained version record. - Version records outlive the decisions made under them. --- ## Retention schedule by decision type _Decision records inherit the retention rule of the underlying decision, not a generic log policy._ ### Template | Decision type | Governing regime | Minimum retention | Notes | |---|---|---|---| | Employment (hiring, promotion) | | | Longer if a claim is filed | | Credit / lending | | | From date of adverse action | | Insurance underwriting / pricing | | | | | Benefits / eligibility | | | | | Content moderation affecting an account | | | | | General automated decision under GDPR Art. 22 | GDPR | For the period needed to demonstrate compliance | Plus right-to-explanation support | ### Worked example | Decision type | Regime | Minimum retention | Notes | |---|---|---|---| | Hiring (screening score + outcome) | EU member-state employment law + GDPR | 2 years from decision | Extend to end of any litigation | | Refund / goodwill (support copilot) | Consumer + internal audit | 3 years | Matches finance record retention | Retention configured per decision_type at write time; archival job moves records to cold storage at 1 year, deletes at the limit. ### Acceptance criteria - Each decision type maps to the specific legal regime that sets its retention, not a blanket period. - Retention is enforced by an archival and deletion job, not left manual. - A litigation hold process can suspend deletion for named records. --- ## Audit readiness checklist _Run per system to confirm a single decision can actually be reconstructed._ ### Template | Check | Status | Evidence | |---|---|---| | Pick a random past decision; retrieve its full log record | Y / N | | | Resolve its model_version to a version record and changelog | Y / N | | | Recover the exact inputs used | Y / N | | | Produce the explanation artifacts for that decision | Y / N | | | Show the human review step and its reason, if applicable | Y / N | | | Confirm the record is within its retention period and will be kept | Y / N | | | Time taken to do all of the above | | target: under 1 hour | ### Worked example | Check | Status | Evidence | |---|---|---| | Retrieve full log record | Y | dec-2026-0091823 pulled in 3 min | | Resolve model_version | Y | 2.4.0 to registry entry + changelog | | Recover exact inputs | Y | feature vector stored; raw CV via hash lookup | | Produce explanation | Y | top-features list on the record | | Show human review | Y | recruiter override with reason | | Within retention | Y | 2 yr, archived at 1 yr | | Time taken | 22 min | under the 1-hour target | ### Acceptance criteria - The reconstruction test has been run on a real past decision, not a fresh one. - Every step succeeds within the time target, or the gap has an owner and a fix date. --- ## Governance controls this kit produces evidence for - **ALC-001**: The decision-log schema is the decision-logging standard, applied per decision. - **ALC-002**: Schema plus retention schedule plus the readiness test are the high-risk audit trail. - **ALC-003**: The retention schedule is the log retention policy, tied to each decision's legal regime. - **CHM-001**: The versioning policy keeps every logged model_version resolvable to a retained record. - **ALC-004**: The explanation-artifact fields provide per-decision explainability evidence.