Why keep an audit trail for AI-assisted decisions?

When AI influences decisions that affect people — hiring, credit, eligibility, content moderation — you need to show what happened and why. An audit trail records the model, the inputs, the output, who reviewed it, and any override, which is essential for accountability, dispute resolution, and frameworks like the EU AI Act and GDPR's automated-decision rules.

What fields should an AI decision record contain?

At minimum: a decision ID and timestamp, the decision type, the AI system and model version, a summary of the inputs, the AI's output and any confidence indicator, the human reviewer, the final decision, and the rationale for any override. The template assembles these and adds framework-specific fields where relevant.

How does this relate to the EU AI Act?

The EU AI Act expects record-keeping and human oversight for high-risk AI systems. Capturing model version, inputs, outputs, and the human reviewer's role helps demonstrate that a person was meaningfully involved and that decisions can be reconstructed and explained after the fact.

Should I log the raw input or a summary?

Log enough to reconstruct the decision while minimizing sensitive data. Often an input summary plus a reference to where the full input is stored is the right balance — avoid duplicating regulated personal data into yet another store unless required.

Does this tool send my data anywhere?

No. The template is generated and filled entirely in your browser. Nothing you enter is uploaded; the JSON export is created locally.

What is the AI Decision Audit Trail Template?

Create a structured audit trail template for documenting AI-assisted decisions — capturing model version, input summary, output, confidence indicators, human reviewer, and override rationale for regulatory accountability. It runs free in your browser on Gera Tools, with nothing uploaded.

AI Decision Audit Trail Template

Name: AI Decision Audit Trail Template
Creator: Gera Tools
License: https://creativecommons.org/licenses/by/4.0/

Get one useful tool a week

Like this tool? Enter your email and we'll send you one genuinely useful Gera tool a week — plus a link to come back to this one. No spam, one-click unsubscribe any time.

When AI helps make a decision that affects someone, “the model said so” is not an acceptable record. This tool generates a structured audit-trail template that captures exactly what to log — model version, inputs, output, confidence, the human reviewer, and any override — so AI-assisted decisions can be reconstructed and defended.

How it works

You pick the decision type, the AI system, and the regulatory framework you operate under. The tool builds an audit record with the appropriate fields, including framework-specific additions (for example human-oversight and explainability fields for the EU AI Act, or automated-decision fields for GDPR). You fill in the specifics for a real decision and export the record as text or JSON for your audit store.

Why each field matters

The model version lets you tie a decision to a specific behavior if a model is later found faulty. The input summary and output allow reconstruction. The confidence indicator flags decisions that warranted closer review. The human reviewer and override rationale demonstrate meaningful human oversight — the difference between an accountable process and a black box.

What the record looks like in practice

A complete audit record for a hiring-screen decision might include:

Decision ID and timestamp — unique ID plus the exact time in UTC
Decision type — “Initial candidate screening — shortlist / reject”
AI system and version — the name of the screening tool and the model version it ran on
Input summary — “CV for [anonymised reference] applying to [role], assessed against [criteria]”
AI output — “Recommended: shortlist; primary factors: [skills match, experience band]”
Confidence indicator — “High confidence (above 85%) / Low confidence (below 60%)”
Human reviewer — name and role of the person who reviewed the recommendation
Final decision — what the human decided (accept, reject, override)
Override rationale — if overridden, why (for example: “Skills match understated due to non-standard CV format; human review found relevant experience”)

This record answers the questions a regulator, an auditor, or a rejected candidate would ask: what did the AI say, who checked it, and why was the final call made?

Framework-specific additions

EU AI Act (high-risk AI): adds human-oversight attestation (confirming a person genuinely reviewed and could override), explainability notes, and system-level logging of the monitoring plan. GDPR Article 22 (automated decisions): adds a reference to the legal basis for automated processing and the right-to-contest mechanism provided. EEOC / NYC Local Law 144 (employment): adds the bias-audit reference and candidate-notification record.

Tips and notes

Log enough to reconstruct the decision, but minimize duplicating sensitive data — a summary plus a reference to where the full input lives is often the right balance.
Standardize one record shape across your organization so logs are queryable and comparable.
Pair this with the Zero-Trust AI Access Policy so the systems producing these decisions are themselves governed.
Review override rates periodically: a high override rate signals the AI’s recommendations are not trusted, which is itself evidence the system needs retraining or replacement.