AI staff policy assistant.

What a staff policy assistant does in a school

A staff policy assistant answers questions such as “What is the approval process for an overnight trip?” or “Can I photograph students on my own phone?” by quoting the school’s own documents. It is retrieval over a small, curated library, not a general chatbot, and it always shows the policy name, section and review date so staff can check the source.

Many schools publish policies through SchoolDocs, which says about 2,200 New Zealand schools use it, so the starting text is often already structured. This page is part of our guide to AI for education administration.

The workflow at a glance

Inputs

Approved policies and procedures, the staff handbook, current collective and individual agreements from the Ministry of Education site, a named owner for each document, and staff questions typed into Teams, Google Chat or a web page.

Systems

A document store in Microsoft 365 or Google Workspace, a private retrieval index, a business-tier model that does not train on inputs, school sign-in, and a question log.

Steps

Check the question against a refusal list, retrieve passages from approved documents only, answer in plain language with citations, say plainly when nothing supports an answer and name who to ask, then log the question, sources and feedback.

Human review

The principal or business manager approves the library and refusal list before launch, then samples 20 logged answers a week for the first term and fortnightly after that.

Outputs

Cited answers, referrals to a named person, and a monthly report of unanswered questions and stale documents.

Measures

Correct answers with correct citation in the sample, referral rate, documents past review date, and repeated questions about one policy.

Risks

Superseded agreements, answers that read as personal employment advice, HR material leaking into the index, and staff trusting an answer without opening the source.

Illustrative worked example: a 40-staff primary school

Illustrative example: a primary school with 28 teachers and 12 support staff. Assumptions: the office and principal field 30 policy or procedure questions a week in term time, each takes about 5 minutes once interruptions are counted, and 60% can be answered directly from documents. All figures are synthetic.

The model price is Anthropic’s API list price as listed on 19 October 2026. Model fees are negligible and the time returned is modest. The stronger case is consistency: staff stop relying on someone’s memory of a policy that changed last year.

Which NZ documents and rules shape the answers

Agreements change. Teacher pay, leave and allowances sit in collective and individual agreements that the Ministry publishes. Its page on primary teachers’ agreements notes the recently ratified PTCA 2026–2028 and newly promulgated IEAs. An index holding last year’s PDF will answer confidently and wrongly, so index one current version and print its date.

Grievance clocks keep running. The Ministry’s problem resolution page explains that a personal grievance must be raised within 90 days (12 months for sexual harassment). A message that reads like a complaint about treatment must go to the principal the same day, and the assistant must tell the staff member it has not raised anything on their behalf.

Boards have their own advice line. The NZ School Boards Association runs GovHub, which offers free governance and employment advice to state and state-integrated boards. The assistant can point managers there; it should not interpret contested clauses itself.

AI policies are thin. ERO’s 2026 Ready or not report found 45% of primary and 53% of secondary leaders use AI to generate policies, and about seven in ten lack a privacy and security policy for teacher AI use. Index only board-approved policies, never AI drafts. For the AI policy itself, use the Ministry’s generative AI guidance and our AI use policy template.

What stays with the principal and the board

Anything that decides something about a named person stays with a person.

• Working out any individual’s entitlement, pay or leave balance.
• Discipline, competency and complaints about colleagues.
• Anything touching child safety, which goes to the person named in the child protection policy.
• Interpreting ambiguous clauses, and approving what goes into the library.

Where it goes wrong and how to catch it

Most failures start in the document library, not the model.

The last row reflects the Ministry’s own warning that AI models are likely to be weak on mātauranga and te reo Māori.

How to start

Begin with a document inventory: every policy, its owner and its review date. Remove duplicates before indexing, write the refusal list, and pilot with five staff for three weeks against the accuracy bar in the FAQs. Our guides to building a searchable knowledge base and private AI search cover the build, and RAG development is the service when you want it built for you. Related workflows: parent enquiry triage and administrative document workflows.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.