6 min read. Updated 20 October 2026

Using AI for policy papers in New Zealand

Begin with an approved source pack and a defined policy question. Build a claim ledger, distinguish evidence from inference and draft one section at a time. Open every material reference and check the original context. Retain the review and approval record under your agency’s requirements; the analyst remains responsible for the paper.

Read the guide’s sources.

See Copilot training NZ, Copilot courses NZ and Copilot workshops NZ.

Illustration generated with AI.

How can policy analysts use AI without losing the evidence?

Begin with an approved source pack and a defined policy question. Build a claim ledger, distinguish evidence from inference and draft one section at a time. Open every material reference and check the original context. Retain the review and approval record under your agency’s requirements; the analyst remains responsible for the paper1,2.

Key points

  • Separate an observation from the explanation inferred from it.
  • Trace material claims to the original passage.
  • Ask agency records staff what the real workflow must retain.

What policy question should the drafting task answer?

Write the decision the paper is intended to support and identify its audience. A request to write a policy paper leaves too much unstated: the problem, available evidence, scope of advice and authority to recommend an option. Give the analyst a bounded task before asking an assistant to organise or draft anything.

Use a fictional exercise such as comparing alternative ways to publish routine service information. The source pack might contain an invented service description, a feedback summary and a list of constraints. Label the pack as training material. It should not look like research about an actual agency or imply that a real consultation has taken place.

The Government Chief Digital Officer’s responsible-AI guidance covers public-service use, including governance and accountability1. Read it alongside the agency’s own policy. A general training workflow cannot decide which information the agency permits in a tool or who has authority to approve the resulting advice.

How should we assemble the evidence pack?

Choose sources that directly address the policy question and record their origin, date, purpose and limitations. Keep the authoritative documents available to the reviewer. An extracted paragraph may lose an exception, definition or population boundary that matters to the argument. The assistant’s source pack should make those limits visible rather than hide them behind a short summary.

Separate source statements from the team’s interpretation. For the fictional service example, a feedback note might describe difficulty finding a form. That does not establish why the difficulty occurred or how many people are affected. Record the observation accurately and identify the further evidence needed before making a broader claim.

Ask the source owner to clear the material for the proposed account and purpose. If there is doubt, replace it with invented information for the workshop. The exercise can teach evidence handling without using sensitive policy development, personal information or documents whose sharing conditions have not been resolved.

What belongs in a claim ledger?

Give each proposed claim a row containing its source passage, the interpretation being made and the check required. Include whether the claim is a direct statement, an inference or an unanswered question. This structure helps a reviewer identify where the reasoning moves beyond the source instead of treating every sentence as equally established.

Ask Copilot to draft the ledger from a small approved extract, then compare it manually. Look for a claim that merges different populations, changes the time period or omits a qualification. A citation attached to the sentence is only a pointer; the reviewer still needs to check that the linked material supports the words used.

For the invented exercise, include a statement that staff prefer a shorter form, but no evidence about completion rates. The ledger should preserve that distinction. A recommendation may still be considered, but the paper must not claim that a shorter form has been shown to improve completion when the source pack contains no such finding.

Ledger fieldFictional entryReview test
ObservationStaff report difficulty locating the formDoes the source say this directly?
InferenceNavigation may need improvementIs this labelled as an interpretation?
UnknownEffect on completed applicationsIs the absence of evidence visible?
Next evidenceA permitted review of the service journeyWho owns the follow-up?

How do we draft an options comparison?

Set the options and the evaluation questions before asking for prose. These might concern implementation effort, accessibility, operational responsibility and evidence gaps. Use the agency’s approved method where one exists. The model can help organise the comparison, but the analyst should decide which considerations matter and whether the available evidence is sufficient.

Ask for a table with a short rationale and a source reference in each cell. Permit a cell to say not established by the supplied evidence. This is especially useful in the fictional exercise, where the pack deliberately leaves some questions unanswered. A fully populated table can conceal uncertainty rather than resolve it.

Review whether the draft uses a consistent basis across options. One option should not receive a quantified benefit while another is described only with vague concerns unless the evidence explains that difference. Record assumptions explicitly and ask the responsible analyst to decide whether more work is needed before a recommendation can be supported.

How should we check the argument and citations?

Read the draft from the decision maker’s perspective, then trace its important statements back to the ledger. Confirm that the recommendation follows from the evidence and stated assumptions. An elegant transition can disguise a missing step in the reasoning, such as moving from an observed problem to a specific solution without examining alternatives.

Open references yourself and inspect the passage in context. Check publication date, scope and any qualification that the draft omitted. If the source cannot be found or does not support the claim, revise the paper rather than simply asking for a different citation. The problem is the unsupported statement, not the appearance of the reference list.

Use a separate challenge review. Ask a colleague what evidence would weaken the recommendation and which uncertainty matters most to the decision. The assistant can suggest questions, but a human analyst must judge their relevance. Keep useful challenges and the response to them in the working record.

What should be retained about AI-assisted drafting?

NZ Digital government’s records guidance explains that using AI does not remove public recordkeeping responsibilities and points agencies to Archives New Zealand for authoritative advice2. Consult the agency’s records staff about what the actual workflow must retain. Do not invent a retention period based on an assumption about chat history.

For the training exercise, keep the invented source pack, the claim ledger, the draft and the reviewer’s corrections together. Add a note describing how AI was used. This makes it possible to inspect the reasoning and demonstrates a review method without making a claim about the agency’s formal recordkeeping decision.

If the proposed real workflow will use integrations or generate material in another system, include that in the records discussion. The location of the final paper may differ from where prompts or intermediate drafts are held. Ask the system and records owners to resolve those questions before expanding the practice into live policy work.

How does a workshop test this process?

Give participants a short fictional policy question and a deliberately uneven source pack. Ask them to identify what the evidence can and cannot establish before drafting. Then have one group prepare a claim ledger while another reviews it. Compare disagreements about interpretation before asking either group to write a recommendation.

Introduce a changed source midway through the exercise. Participants should identify which claims and options need another check. This reveals whether the paper has a usable evidence trail or whether facts were copied into prose without a way to maintain them. Keep the changed document clearly labelled as a training version.

Finish with a review discussion about the limits of the answer. Ask what additional evidence would be needed for a real decision and who would obtain it. The outcome is a checked practice paper and a list of open questions, not a claim that the workshop has approved a government policy or certified an agency’s process.

How should analysts choose training and prepare?

A foundations session can teach staff to give bounded instructions and inspect generated output. Teams already familiar with the product may need more time on source comparison, permissions and review responsibilities. Describe the policy work rather than asking for a broad tour of every feature available in the account.

Bring the people who draft, review and own the information rules. They may need different exercises: an analyst can practise a claim ledger while a manager reviews whether the decision and uncertainty are clear. Agree an invented or public pack that everyone can use, and confirm the account features required before the workshop.

The courses below offer starting points for Copilot practice and rollout planning. Tell thecolab.ai the intended audience, approved environment and evidence-handling questions. We can scope training around those tasks while leaving agency approvals, policy judgment and records decisions with the responsible officials.

Related

  1. 1Introductory

    Microsoft 365 Copilot foundations workshop

    A half-day of Microsoft 365 Copilot training: check your access, draft an email, review a document and test a spreadsheet answer using cleared examples.

  2. 2Leadership

    Copilot for managers

    Help managers catch up, brief from files they can open, and set a checking rule for the team.

  3. 3Organisation

    Copilot rollout and readiness

    Prepare the tenant, choose a pilot, and measure whether people use Copilot well.

Industries and teams

More on office team practice

See all guides.

Questions

How can policy analysts use AI without losing the evidence?

Begin with an approved source pack and a defined policy question. Build a claim ledger, distinguish evidence from inference and draft one section at a time. Open every material reference and check the original context. Retain the review and approval record under your agency’s requirements; the analyst remains responsible for the paper1,2.

Train your team

  1. ScopeHeadcount, sites, Microsoft 365 plan and licences, and the teams you want to start with.
  2. PlanWe propose a course, a pilot group and a format, and agree the files and meetings to use.
  3. TrainOn-site or over Teams, in Outlook, Word, Excel and Teams, with your own work.
  • In person
  • Remote
  • Half-day workshop
  • Full-day workshop
  • Team rollout

Independent training by thecolab.ai. Not affiliated with Microsoft.

Enquire

Tell us about your team. We reply by email with a suggested course and format.

Not sure which licence? We check before training.

Include the apps, staff roles, number of people and location you have in mind.

Team workshops from $4,000 + GST for up to 10 people. +$200 + GST per extra person.

Public sessions from $399 + GST per person. Schedule coming soon.

No payment now. We reply by email with a suggested course and format.

We use these details only to reply to you, and store them securely with thecolab.ai. Privacy

Sources