Why begin with the workflow rather than the tool name?
Describe what people will actually do. “Use AI in the office” is too broad to assess. A more useful proposal says that an administrator will draft an internal event notice from an approved brief, a manager will check it, and the final notice will be saved in the normal communication system. That description identifies inputs, actions and decisions.
The Privacy Commissioner advises organisations to complete a privacy impact assessment before using generative AI, to identify risks and plan how to reduce them3. The Privacy Commissioner describes a PIA as a way to identify and assess privacy risks in handling personal information and consider how to reduce them1. The official toolkit provides assessment resources2. Start there for the organisation’s formal process; the fictional worksheet in this guide helps staff prepare the questions they need to bring to that review.
Separate the proposed benefit from an established result. A team may hope to reduce drafting effort, but that is a hypothesis to test. Record the task’s current process and why an AI draft is being considered. If the same outcome can be achieved without supplying personal information, include that option in the discussion.
What information should the worksheet capture?
List the fields entering the workflow, where they come from and why they are needed. Inspect free-text notes as well as obvious identifiers. A brief may contain names, contact details, access needs or comments about a person even when its main purpose appears administrative. “Document upload” does not describe the information precisely enough.
Record who can see the input and output, which account or workspace is proposed and where copies may be kept. Ask the people responsible for the system to provide evidence for technical statements. Leave a question open when the retention or access arrangement has not been confirmed; do not fill it with a vendor slogan or a guess.
The downloadable worksheet is blank apart from an explicitly fictional practice row. It contains no personal records. Use it to organise questions for the privacy lead, system owner and business owner. Keep the completed assessment in the organisation’s approved location, especially if it later describes actual systems or sensitive information.
How does the fictional event-notice example work?
Imagine a team preparing a notice about an internal equipment demonstration. The approved brief contains a topic, a room request and the role that handles enquiries. A separate attendance list includes personal information that the drafting task does not need. The proposed exercise uses only the approved brief and leaves the attendance list out.
Ask participants to identify what could still go wrong. Someone might paste a participant’s email into the prompt, attach the wrong file or accept an invented accessibility promise. These are hypothetical training risks, not incidents at a real organisation. Give each risk a concrete cause, possible effect and proposed control so it can be discussed and tested.
A control might require a cleared practice pack and a check of the prompt before it is sent. Test whether participants can follow that rule when several similarly named files are present. If the rule depends on perfect memory, improve the preparation process. Record what the test showed rather than declaring the risk resolved because a policy sentence exists.
| Worksheet field | Fictional example | Open question |
|---|---|---|
| Purpose | Draft an internal equipment notice | Who approves the final notice? |
| Input boundary | Approved event brief only | Has the practice folder been checked? |
| Excluded information | Attendance list and individual access needs | How will the organiser collect necessary details separately? |
| Control test | Participant selects and reviews the cleared brief | What happens if the wrong file is selected? |
Who should own the decisions?
Identify a business owner for the task, a privacy reviewer for the assessment and a system owner for the account configuration. The exact roles depend on the organisation. Use real role holders in an actual assessment; in the training worksheet, role labels are enough. Avoid leaving every action assigned to a generic team.
The business owner explains the purpose and decides whether the proposed work is useful. The system owner confirms the technical setup and available controls. The privacy reviewer considers the information handling and unresolved questions. Where specialist advice is needed, record that need and the person responsible for obtaining it instead of asking ChatGPT to approve the arrangement.
An owner field is not a signature. Keep approval status, conditions and outstanding actions separately visible. A person may agree that a proposal is worth investigating without approving its use with real information. The record should make that distinction clear to someone who was not present in the initial discussion.
How should controls be tested before a pilot?
Turn each proposed control into an observable check. If the rule says only cleared files may be used, inspect the actual practice folder and ask a participant to choose the input. If access should be limited, ask the system owner to verify it in the intended account. Describe the evidence retained for each check.
Include a failure scenario in the exercise. Supply a fictional brief with an unnecessary personal-detail field and ask what should happen before drafting. The expected behaviour is to stop and resolve whether the field is needed and permitted. A well-written output does not compensate for an input that should never have been supplied.
Keep residual questions visible. Some risks cannot be answered in a workshop, particularly where the team needs contractual or technical evidence. The assessment should say what remains unknown, who will investigate and what work is paused meanwhile. Do not treat completion of a training activity as approval to move on to live data.
What records and review triggers should we agree?
Record the proposed workflow version, the evidence considered, the controls tested and the approval decision. Keep enough context to explain why the decision was made. A future reader should be able to distinguish the original pilot scope from a later expansion without reconstructing the history from scattered chat messages.
Agree what changes require another review. Examples include a new input type, a wider audience, an integration with another system or a different use for the output. These are suggested planning triggers; the organisation’s formal requirements govern the actual review process. The point is to avoid assuming that approval for one narrow task applies to every later use.
Also describe how staff raise concerns or mistakes. Give them an established internal contact and a clear instruction to pause the affected activity. Do not ask staff to resolve an information incident by continuing to experiment with the tool. Follow the organisation’s own incident process and obtain appropriate advice for the circumstances.
Can ChatGPT help prepare the assessment itself?
It can help organise a fictional or suitably cleared description into questions and headings, but it should not be the source of the approval. Ask for omissions to investigate, not a declaration that the process is compliant. Verify any reference to law, policy or vendor terms against the actual current source before including it in the assessment.
For practice, give ChatGPT only the invented event scenario and ask for a list of unanswered information-handling questions. Compare that list with the worksheet fields. Participants should identify both useful prompts and misplaced assumptions, such as an answer claiming that a particular retention period applies without evidence about the account.
A human reviewer should edit the resulting questions into the organisation’s normal assessment format. Keep the distinction between verified evidence, a proposed control and a matter requiring advice. That separation makes the draft easier to assess and prevents generated wording from appearing more authoritative than the underlying work supports.
How can a team practise the review conversation?
Assign participants the business, privacy and system-owner roles. Give each a different part of the fictional brief and ask them to identify what they cannot yet confirm. Then bring the findings together in the worksheet. The exercise shows why a single enthusiastic tool user cannot answer every question alone.
Finish by writing a bounded decision: what may be practised, with which invented input, under whose review, and what remains paused. Keep the reasoning alongside the decision. This is useful preparation for a real approval discussion while making no claim that the exercise has certified the organisation or its chosen product.
Responsible-AI training can help staff practise these conversations and recognise information they should not supply. The related course explains the scope. When enquiring, describe the workflow and the roles involved without sending personal information; the team can begin with a fictional version and agree the appropriate review participants.