AI Audit Trail: What to Record and How to Start

An AI audit trail is a time-ordered, tamper-evident record of what an AI system was asked, what it used, what it produced and who approved the result. It is built so someone can rebuild a single decision months later. For most Australian small and medium businesses, the hard part is not the logging format. It is that nobody can say which AI tools staff use today, so there is nothing yet to log.

This guide covers what a trail contains, how it differs from ordinary logs, what Australian guidance expects, and a practical order of work.

CalibreforumMap AI Use Before You AuditCalibre Forum’s AI readiness assessment maps how employees use AI tools and provides a scored report with a 90-day action plan.Explore the AI readiness assessment

TL;DR

An audit trail links each AI output to the request, model version, data used and the person or account that authorized it. Raw system logs alone do not do this.
Australia’s Voluntary AI Safety Standard asks organizations to keep records and an AI inventory. It is voluntary, but it shows where expectations are heading.
You cannot audit tools you do not know about. Build the inventory first, then decide what to record for each tool.
An audit tool gathers evidence. It does not replace governance or an independent reviewer.
Our anonymous staff survey and 90-day plan give you the usage map that a trail depends on. We do not provide logging software.

What an AI audit trail actually is
What a good trail records
Audit trail, audit log and audit: three different things
What Australian guidance expects
Why most small businesses stall before they start
A sensible order of work
For agents and automated workflows
Our perspective
Where the readiness assessment fits
FAQ
Sources

What an AI audit trail actually is

A useful definition comes from Identra’s glossary: a chronological, tamper-evident record covering the prompt, the context retrieved, the model version, tool calls, policy decisions and the final output or action. Unlike debug logs, which are often sampled or short-lived, a trail is designed for long retention and for forensic reconstruction.

The question it answers is narrow and practical. Not “what does this model usually do?” but “what did this one do, on whose behalf, and who approved it?”

A simple example: a staff member uses an AI assistant to draft a client letter that quotes a figure. Months later the figure is disputed. A real trail lets you see the prompt, the document the assistant pulled the figure from, the model version, who reviewed the draft and when it was sent. Without it, you have a sent letter and a shrug.

What a good trail records

The IETF draft on agent audit trails proposes a JSON-based record with mandatory fields for agent identity, the authorizing principal, action classification and outcome. You do not need to adopt a draft standard to learn from it. The fields translate to plain questions:

Question	What to capture
Who asked?	The requesting user, kept separate from the agent or service account that did the work
What was asked?	The prompt or instruction, with sensitive details redacted or hashed
What did it use?	Documents or data retrieved as context
Which model?	Name and version at the time
What did it do?	Tool calls, actions taken, outputs produced
Who approved?	The human or policy that authorized the action
When?	Timestamps from the source system and from when the event was collected

Two details from the same draft are worth copying. First, record both the source timestamp and the collection time, because late event delivery and clock skew can scramble the order of an investigation. Second, keep the requesting user and the executing agent as distinct identities, or accountability gets muddy.

Tamper evidence

Identra describes entries that are hash-sealed on write: each one includes a fingerprint of the one before, so an edit anywhere breaks the chain. For a ten-person firm using a chat assistant, this is probably more than you need today. For anything that takes actions or touches client decisions, it is the property that makes a record believable.

Linked audit records reveal a broken chain

Evidence by design

Wolters Kluwer makes a point about treating governance decisions as evidence: model approvals, waivers and policy settings belong in the record next to the technical logs. That turns accountability from a scramble during an incident into a routine. In practice, it means your “we approved this tool for drafting, not for client data” decision gets written down, dated and kept.

Audit trail, audit log and audit: three different things

People search for these terms interchangeably, and the mix-up causes bad purchases.

Audit log: a raw technical record from one system. Useful, but often short-lived and incomplete.
Audit trail: logs joined into a story about one decision, with identity, authority and outcome.
AI audit: a review, ideally by someone independent, of whether your use of AI is governed, evidenced and safe.

An analysis of AI audit tools makes the key point clearly: no platform makes an audit independent just by producing a report. Tools are the evidence-gathering layer. Governance, evidence quality and reviewer competence decide whether the audit means anything.

What Australian guidance expects

There is no single Australian law titled “AI audit trail”. But two pieces of official guidance point straight at record-keeping.

The Voluntary AI Safety Standard

The Department of Industry, Science and Resources sets out ten guardrails. Guardrail 9 covers keeping records so third parties can assess compliance, including an AI inventory and documentation of accountability, purpose, testing and risk assessments. The standard’s introduction describes the guardrails as voluntary and applying to organizations across the AI supply chain. Voluntary means nobody is forcing you, but it is a reasonable yardstick for what a customer, insurer or regulator may ask about later.

Privacy and the OAIC

The OAIC’s guidance on commercially available AI products says organizations should consider whether using personal information in an AI system is necessary and the best solution, and should not use AI simply because it is available. It also points to internal policies on permitted and prohibited uses, human oversight, and ongoing monitoring of AI output.

This cuts both ways for audit trails. A trail supports oversight, but it can also quietly collect personal information. Redact or hash sensitive inputs and outputs before storage, and review the trail regularly instead of setting it and forgetting it. If your records touch personal information, talk to a privacy lawyer or your privacy officer about retention and access before you switch logging on.

Why most small businesses stall before they start

The usual advice assumes you already know where AI is used. Many owners do not. Staff sign up for free assistants, paste text into browser tools, and use AI features switched on inside software you already pay for. None of it appears in a procurement list.

That gap creates three problems for any audit trail project:

No scope. You cannot decide what to log when you cannot list the tools.
No ownership. Nobody is named as the authorizing person for a tool nobody approved.
No honesty. If staff fear punishment, they will tell you what you want to hear, and your inventory will be fiction.

Our own AI security assessment guide goes deeper on finding shadow AI, and the one page vendor risk checklist helps once you know which vendors are in play.

A sensible order of work

Six steps for starting an AI audit trail

Step 1: Build the inventory

List every AI tool in use, who uses it, for what task and with what kind of data. Include AI features inside existing software. This is the same inventory Guardrail 9 expects, so the effort is not wasted.

Step 2: Classify by consequence

Sort uses into three bands: low (brainstorming, internal summaries), medium (client-facing drafts a person reviews) and high (anything that decides, scores, filters or acts on its own). Spend your logging effort on the high band first.

Step 3: Write down the rules and approvals

Record which tools are approved for which tasks, who approved them and when. Document how the process works around the AI step too. Tools that let you capture and structure processes in plain language can make this easier, because a trail is far more useful when a reviewer can also read what the intended procedure was.

Step 4: Match the tool to the task

Thomson Reuters draws a useful line between three categories of tooling: document-focused assistants for evidence extraction, AI-enabled audit platforms for end-to-end workflows, and broader governance tools for inventory and risk. Buying a narrow extraction tool when you need inventory and risk management, or the reverse, is a common mistake. Decide which problem you have before you look at products.

Step 5: Turn on logging where the stakes are highest

For each high-consequence use, capture the fields in the table above. Check retention periods, who can read the records and who can edit them. Redact personal information before storage where you can.

Step 6: Review on a schedule

Put a quarterly check in the diary. Ask: are new tools in use, are the logs capturing the right depth, and has anyone actually read them? A trail nobody reviews is just storage cost.

For agents and automated workflows

If you use AI agents that call tools or send messages by themselves, the standard of record rises. Research on audit trails for large language models notes that LLMs are increasingly embedded in consequential decisions across healthcare, finance, employment and public services, yet accountability remains fragile because process transparency is rarely recorded. For an agent, log the authorizing person, the exact action and the outcome, so you can answer who let it do that.

Our perspective

We see the same pattern in small and mid-sized Australian and New Zealand businesses. Leadership asks about audit trails and logging, then discovers the prior question is unanswered: what are people actually using, and for what?

So we start with the people. We survey every staff member, anonymously, because anonymity is what gets honest answers. Nobody installs software and we collect no individual data. The result is an organization-wide map of tools and workflows, a scored report, and a 90-day plan. We make no compliance or legal claims, and we do not produce your audit trail. What we give you is the scope, the inventory and the priorities that make a trail worth building.

If you take one thing from this article: do not buy logging before you can name your tools. The sequence matters more than the product.

— Anthony

Where the readiness assessment fits

Our assessment is a structured, anonymous staff survey that produces a combined scored report: existing capabilities, risks, and a tailored 90-day action plan with operational and security recommendations. It is a sensible first step before you commit to audit software.

Report: A$1,990 one-off. See pricing.
Membership: A$2,388 per year for recurring assessments and reports, so you can see whether usage and risk are moving in the right direction. Details are on the Calibre Forum site.

You can look at a sample report first, read what the report contains, or see how it works from setup to report. Staff privacy is covered on our security and privacy page.

Sector matters too. If you run an accounting firm, a law firm or an aged care or NDIS provider, the records you will want most differ, and we have pages for each.

FAQ

What is an example of an AI audit trail?

A record showing that a named employee asked an assistant to summarize a contract, which document it retrieved, which model version answered, what it returned and who reviewed the summary before it went out. Each entry has timestamps and links to the next. The point is that someone can rebuild that single decision later.

Is an AI audit trail required in Australia?

There is no general law that requires one for every business. The Voluntary AI Safety Standard does ask for records and an AI inventory under Guardrail 9, and OAIC guidance expects oversight and monitoring where personal information is involved. For your specific obligations, ask a lawyer or your industry regulator.

Do AI audit tools replace an audit?

No. They collect and preserve evidence. Independence and judgment still come from your governance and from competent reviewers, as the tool category analyses make clear.

How do I audit AI use if I do not know which tools staff use?

Start by asking them, in a way that is safe to answer honestly. An anonymous survey across the whole team gives you the first inventory, which you can then refine with IT checks of software and subscriptions. Then decide what to log for each high-consequence use.