Keeping records when AI helps with official work

Why records matter here

Government work depends on records. A file shows who decided what, on what basis and when. If an AI tool helped produce a summary, a draft or an analysis, that help is part of how the decision was reached. Someone reading the file later (an auditor, a court, a new officer or an inquiry) should be able to see it.

Good records protect you as well. If a question arises months later, a short note showing what the tool did and what you checked is far better than trying to remember.

The basic principle

AI does not change your record-keeping duties. It adds one more thing to record: the part the tool played. Your department's existing rules on files, retention and official correspondence still apply. Where your department has issued instructions on AI use, follow those first. This guide covers good practice where the rules are silent.

What to note on file

For any substantive work where AI helped, add a short note covering:

1. The tool used. Name the tool and say whether it was an approved departmental tool.
2. What it did. For example "produced a first draft", "summarised the attached 40-page report" or "translated the reply into Urdu".
3. What you gave it. Describe the inputs in general terms. Note if you removed personal or sensitive details first.
4. What you checked. Say how you verified the output: against the original document, the relevant rules or a colleague's review.
5. What you changed. A line is enough: "facts corrected in paragraph 3, tone revised".
6. Who is responsible. The officer who approved the final version. See accountability.

You do not need a note for trivial help, such as fixing grammar in an internal email. Use judgement: if the AI's contribution could affect a decision, a citizen or a public statement, record it.

Keeping prompts and outputs

Some work needs more than a note. Keep the actual prompts and the raw outputs where:

• your department's rules or an approved tool's terms of use require it
• the output fed directly into a decision about a citizen, a licence, a payment or a complaint
• the work is likely to be audited, challenged or requested under a right-to-information request
• the AI produced analysis or figures that others will rely on

Save them in the official file or records system, not only in the chatbot's history. Chat history can be deleted, can expire or may sit on a server your office does not control. Copy the prompt and output into a document, add the date and the tool's name and file it.

Never store classified or personal information in a tool that is not approved for it, even for record-keeping. If the prompt itself contained sensitive material, that is a sign the wrong tool was used.

Building an audit trail

An audit trail is the chain of records that lets someone follow how a piece of work was produced, step by step. For AI-assisted work, a good trail lets a reader answer:

• What did the tool produce?
• What did the officer accept, reject or change?
• Which version went to the decision-maker?
• Who signed it off and when?

Simple habits help:

• Keep version history. Save the AI draft and the final version separately, clearly labelled.
• Date everything. Include the date on notes, saved prompts and outputs.
• Use the file, not personal folders. Records on a personal laptop or phone are not records.
• Be consistent. Agree a standard line with your team, so notes look the same across files.

A sample file note

AI assistance: first draft of this reply prepared using [approved tool] on [date]. No personal details were entered. Draft checked against the original complaint and the relevant notification. Facts in paragraph 2 corrected. Final text approved by [name, designation].

Warning signs

• Officers saying "the system wrote it" without a record of what was checked
• AI outputs pasted into decisions with no saved source or date
• Prompts and outputs kept only in personal accounts
• No one able to say which version of a document was AI-drafted

Keeping records is not about distrust of AI. It is about being able to explain your work, which is a basic duty of public service.

Related guide: How to draft letters and notes with AI.