New National Archives Guidance Puts AI Recordkeeping in Focus
When does an AI prompt become a record? Should organisations keep Copilot chat histories? And what evidence needs to be retained when AI contributes to a decision?
These are increasingly practical questions for records and information management professionals, and updated guidance from the National Archives of Australia (NAA) provides clearer direction for Australian Government agencies.
But that does not mean everything generated by AI needs to be kept.
Not every prompt, output or chat needs to be a record
One of the most useful clarifications in the updated guidance is the emphasis on a consistent, risk-based approach.
Agencies do not need to retain every AI output, prompt, input or chat log. Instead, existing records authorities should be applied to determine what needs to be retained for business, legal, evidentiary and accountability purposes.
For everyday use of generative AI assistants such as Microsoft 365 Copilot, an output is more likely to require retention when it:
- forms a final record in its own right
- substantially contributes to a significant record
- provides evidence of a high-risk or sensitive activity
- supports a recommendation or decision
- has continuing business or evidentiary value.
Minor drafts, unused research and reference-only outputs may generally be disposed of under an agency’s normal administrative practice (NAP) policy when they are no longer required.
Context matters: prompts and inputs may need to be kept
The guidance also tackles a question many organisations are confronting: do we need to keep the prompt?
Sometimes.
Where prompts, inputs, source data or metadata are necessary to understand an AI-generated record or demonstrate how a recommendation or decision was reached, they may need to be retained alongside the output.
The principle is essentially proportional: the more significant the record, or the greater the reliance placed on AI in decision-making, the stronger the need to preserve supporting evidence.
For R&IM practitioners, this means determining retention requirements cannot simply be based on whether information was created by AI. The business context, risk, evidentiary value and role of the AI output all matter.
Beware the AI chat history
The NAA also highlights the opposite problem: keeping too much.
Generative AI assistants can rapidly accumulate enormous volumes of prompts, responses, drafts and source material. Prompt and response logs will generally be low-value, transitory or facilitative information and may be eligible for disposal under NAP.
However, agencies should establish appropriate retention periods rather than relying on the default settings of AI platforms, some of which may retain chat histories indefinitely unless configured otherwise.
Over-retention has consequences. The NAA identifies increased storage and management costs, greater exposure of sensitive information, difficulty locating authoritative records and potentially more complex and time-consuming FOI processes.
This makes AI disposal an information governance issue from the outset, rather than something to address once systems have accumulated years of content.
Higher stakes when AI makes or influences decisions
The recordkeeping requirements become more rigorous when AI is incorporated into business systems and automates or materially informs decisions.
In these environments, simply keeping the final output may not provide sufficient evidence.
Depending on the circumstances, agencies may need to retain AI recommendations and outputs alongside prompts, source data, inputs, instructions, metadata, system and audit logs, system design and testing documentation, and decision-making criteria.
The NAA recommends agencies undertake a Business Systems Assessment Framework (BSAF) assessment of business systems incorporating AI to determine whether adequate information management functionality exists.
The underlying accountability question is important: could the organisation later demonstrate how the AI-supported decision was reached?
Agentic AI presents the next recordkeeping challenge
The September update also addresses agentic AI, where autonomous or semi-autonomous AI agents can plan, make decisions and take actions in pursuit of defined goals.
While detailed recordkeeping guidance for agentic AI is still developing, the NAA identifies specialised records such as system prompts, base prompts, agent instructions, activity logs and audit trails.
Because an AI agent can take actions autonomously, retaining evidence of what the agent did and what instructed it to act becomes particularly important for accountability and transparency. The NAA says it is continuing to monitor developments and intends to provide further guidance as practices evolve.
What should R&IM practitioners be doing now?
The guidance reinforces the need for R&IM professionals to be involved early in AI governance, procurement, implementation and system design.
Organisations should be establishing clear business rules around which AI records must be retained, where they should be captured, what metadata is required, how long AI logs should remain available and when automated disposal should occur.
Staff also need practical guidance. Knowing when to capture an AI output is important, but knowing when not to keep one is becoming equally important.
Perhaps the most important message from the NAA update is that AI does not create an entirely new category of recordkeeping obligation. Existing information management principles still apply, but organisations need to interpret and implement them in environments where information can be generated, transformed and acted upon at unprecedented speed.
For R&IM practitioners, that puts familiar principles of authenticity, reliability, integrity, usability, context and accountability firmly at the centre of responsible AI adoption.
Original source: National Archives of Australia, Information management for records created using artificial intelligence (AI) technologies, updated September 2026. Read the full National Archives of Australia guidance