AI tenancy document search.

What tenancy document search answers

Tenancy document search answers the questions property managers otherwise answer by opening six PDFs: when does the fixed term at Flat 2 end, did we serve a notice to remedy in March, which heat pump was installed and when, what did the owner approve by email. It is a retrieval system over the firm’s own tenancy files that returns a short answer and the exact pages it relied on.

The design rule is simple: no citation, no answer. If the files do not say, the system says it could not find it and shows the closest documents. That one rule removes most of the risk of a confident wrong answer about a tenancy, and it is what our RAG development work is built around.

How the workflow runs

Inputs

Tenancy agreements and renewals, ingoing and routine inspection reports, healthy homes compliance statements, notices, Tenancy Tribunal orders, owner instructions, invoices and tenant correspondence.

Systems

The property management system and document store, a text extraction step for scans, a search index with embeddings, a language model for answers, and the firm’s staff directory for permissions.

Steps

Extract text and metadata (property, tenancy, document type, date, version); index; on each question, filter to tenancies the user may see; retrieve passages; answer with citations; log the question, sources and user.

Human review

Staff check the cited page before acting on any answer about dates, money or notices. A senior property manager reviews a weekly sample of answers and every answer used in a Tribunal application or a Privacy Act response.

Outputs

Short answers with linked citations, “not found” responses with nearest documents, and an audit log.

Measures

Answers with a correct citation in the weekly sample, “not found” rate, time to answer common questions, and retrieval misses found in review.

Risks

Superseded versions, cross-property confusion, permission leaks, and stale guidance on tenancy law.

Illustrative example: a Christchurch firm with 900 tenancies

Illustrative example, synthetic numbers: a Christchurch property management firm holds about 60,000 tenancy documents. Its 14 staff look something up about 40 times a working day, and a manual search averages 6 minutes.

Assumptions, using prices as listed on 19 October 2026: indexing 60,000 documents at an average of 2,000 tokens each (120 million tokens) with OpenAI’s text-embedding-3-small at US$0.02 per million tokens (OpenAI pricing). Each answer reads about 12,000 tokens of retrieved passages and writes 400, on Claude Sonnet 5.5 at US$2 and US$10 per million (Anthropic pricing), about 2.8 US cents per answer. Text extraction for scanned files and the time to set permissions properly will cost more than either line.

NZ rules that shape the index

• Retention. The Privacy Commissioner’s guidance for landlords notes that under the Residential Tenancies Act 1986 certain documents, including inspection reports and correspondence with tenants and prospective tenants, must be kept during the tenancy and for 12 months after it ends, and business records for seven tax years. Store each document’s retention date in the index so expired material can be reviewed for deletion.
• Unsuccessful applicants. The same guidance says you generally should not keep information about people who applied but were not successful, beyond a short period. Keep application files out of the general search index.
• Access requests. Tenants can request the information you hold about them, and the guidance warns that destroying tenant information after receiving a request is an offence. Search speeds up the gathering; a person decides what is released.
• Accuracy. Under the Privacy Act 2020, information must be accurate, up to date, complete, relevant and not misleading before it is used or disclosed. An answer from an old agreement version fails that test.
• Compliance records. Every new or renewed tenancy agreement must include a healthy homes compliance statement, which makes these statements a common search target. Tag them as their own document type.

What staff still judge

Staff decide what an answer means for the tenancy: whether a notice was valid, whether a term has been breached, what to tell an owner. Questions about legal rights should be answered from Tenancy Services or a lawyer, not inferred from past files. Our lesson on checking AI answers against sources sets out a quick verification habit.

Where document search fails

• Superseded versions. A renewal changed the end date, but the original agreement ranks higher. Index version and effective dates, and prefer the latest unless the question asks about history.
• Similar addresses. “Flat 1” and “Flat 2, 14 Rata Street” retrieve each other’s documents. Filter by tenancy ID before searching text.
• Permission leaks. A staff member asks about a portfolio they do not manage and gets a cited answer. Apply permissions at retrieval, not by hiding results afterwards; our guide to building private AI search shows how.
• Stale law. Tenancy rules change. Do not index old guidance as if current, and date any external page the system cites.
• Unreadable scans. Handwritten notes extract badly and then retrieve wrongly. Mark low-confidence text and show the image.

Document search pairs naturally with maintenance request triage and inspection report drafting, which create many of the documents it searches. Records handling is also part of the Property Management Agent Levy Street built for APL Property. The property management sector guide covers the rest.

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