AI listing description drafting.

What the workflow drafts, and from what

AI listing description drafting produces the words for a new listing: the portal description, a headline, brochure text, social posts and the short summary agents read at open homes. The difference between a useful system and a risky one is the input. A model given “3 bed brick, great sun, handy to schools” will fill gaps with what similar homes usually have. A model given a fact sheet, and told to write only from it, cannot.

The fact sheet is the real work product. Each field holds one claim and its source: land area and legal description from the record of title, floor area from a measured plan, zoning from the district plan, chattels from the vendor’s list, and agent observations from the appraisal visit. Vendor statements carry a flag: verified, or not. REA’s marketing guidance says that if a vendor claim cannot be verified, such as that the property can be subdivided, “we recommend you do not make the claim”. Unverified fields never reach the model.

How the workflow runs

Inputs

The fact sheet with sources, the agent’s notes and voice memo from the appraisal visit, selected photo captions, the agreed method of sale and price wording from the agency agreement, and the agency’s style guide and banned-claims list.

Systems

The sales CRM (MRI Vault or Rex), the document store holding title and LIM, the portal feed, and a model API.

Steps

Check the fact sheet for missing required fields; draft each format; run a second pass that maps every sentence to a fact field and flags any that do not match; insert price wording verbatim from the agency agreement; queue for the agent with flags shown inline.

Human review

The listing agent checks each flagged sentence and the overall impression, the vendor approves the final copy in writing, and the branch manager or supervisor oversees it before it is published.

Outputs

Approved copy for each channel, a record of which fact supported each claim, and the vendor’s sign-off on file.

Measures

Agent minutes per listing, flagged sentences per draft, unsupported claims found after publishing (target zero), and vendor change requests.

Risks

Invented features, price wording that drifts from the agreement, claims about zoning or permitted use, and copy left live after facts change.

Illustrative example: a Nelson agency listing 18 homes a month

Illustrative example, synthetic numbers: an eight-salesperson agency in Nelson lists about 18 properties a month. Writing the portal description, headline, brochure text and three social posts, then reworking them after vendor comments, takes about 70 minutes a listing.

The fee line assumes 30,000 input tokens (fact sheet, notes and 15 photo captions or images) and 3,000 output tokens for the drafts, plus a checking pass of 10,000 input and 500 output tokens, on Claude Sonnet 5.5 at US$2 per million input and US$10 per million output tokens, as listed on 19 October 2026. That is about US$0.12 a listing. Building fact sheets adds time at first; the saving comes once agents fill them at the appraisal visit.

New Zealand rules the copy must meet

• No misleading claims. Rule 6.4 of the Code of Conduct bars misleading customers or withholding information. Section 14 of the Fair Trading Act 1986 bars false or misleading representations about land’s characteristics, price, location or the use it “may lawfully be put” to.
• Reasonable grounds at the time. Under s 12A, a claim is unsubstantiated if you lacked reasonable grounds when you made it, even if it later proves true. REA gives the example of “home and income” used where a restrictive covenant prevents it.
• Price. Rule 10.4 says an advertised price must clearly reflect the pricing expectations agreed with the client, and rule 10.9 bars advertising on terms the client did not authorise.
• Approval. REA says the client should pre-approve all advertising and a supervisor should oversee it, and that new contrary information means withdrawing and correcting material immediately.
• AI responsibility. REA’s generative AI guidance says licensees must fact-check AI material and that an AI error is no defence.
• Photos. REA’s chief executive told RNZ in July 2026 that photos should be “a true representation” of the property. REA calls it best practice to label aerial boundary lines “only indicative of the boundary”.

Where listing drafts go wrong

• Plausible extras. Heat pumps, double glazing and “sunny all day” appear because comparable homes have them. The checking pass is the control, not the prompt; our lesson on checking AI answers against sources shows the method.
• School zones and walking times. A zone can change, and “walk to school” depends on which school. Treat both as claims that need a source.
• Room counts. A sleepout or study becomes a bedroom. Use the counts on the fact sheet and the agent’s description of any unconsented space.
• Superlatives that read as fact. REA accepts obvious puffery but warns it is risky. “Best views in the street” invites a complaint.
• Te reo place names. Models drop macrons and misspell suburbs. Keep a checked list of local names for the region you sell in.

Listing copy feeds the other two sales workflows: buyer enquiry handling answers from the same fact sheet, and appraisal preparation produces the evidence behind the price conversation with the vendor. The real estate sector guide covers the wider rules, and our AI workflow automation and document processing services cover building fact sheets from titles and LIMs.

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