AI workflows for real estate.

AI Workshops for Real Estate Agencies: Pick One Agency Task, Build the Workflow, Keep the Licensee in Control

Bring the one job your office repeats every week and build a working AI workflow for it in the room. For most agencies that is a vendor update, a listing description, a buyer enquiry reply, an inspection follow-up or an arrears message. You write the instructions, run them on your own material, check the draft against that material, and the licensed agent still approves anything that goes to a vendor, a buyer or a tenant. Price is the one output we never hand over.

That last line does most of the work in a real estate session, so I will start with the tasks and come back to it.

TL;DR

• One repeated agency task beats a tour of tools. Vendor updates, listing copy, enquiry replies, inspection follow-ups and property management messages are the five to train on, and each of them needs a different source layer. Appraisal preparation is the sixth thing agencies ask about and the one that has to be handled differently.
• Property marketing is a regulated output, not ordinary marketing copy. Under the Australian Consumer Law a claim about a property has to be accurate and based on reasonable grounds, and it makes no difference whether the business intended to mislead. 1 A model wrote the sentence, and your licence carries it.
• The price never comes out of the model. In Queensland the advertised price has to line up with the minimum the seller will accept on the Form 6 appointment, and setting a price the vendor will not sell at is treated as false advertising. 2 An appraisal figure is a professional judgement, so it stays with the licensed person.
• Property management is the most rule-bound area and the easiest to get right, because the timeframes are published. A general breach notice runs seven days, unpaid rent runs seven days once rent has been unpaid for seven days, and entry for repairs needs 48 hours. 3 4 Put those numbers in your rules, not in the drafter's memory.
• The public full-day workshop suits one or two people from an office who want to build a workflow and take it back. Private onsite training suits a whole team who need the same workflow running the same way. If you would rather the system was built and connected for you, that is a separate engagement and we scope it.

Which agency task should you bring?

Bring the one that keeps landing back on your desk after hours. In a sales and property management office that usually comes down to six candidates, and they are not equally good starting points.

A weekly vendor update is the strongest first task in most agencies. It repeats on a fixed cycle, the inputs are the same every week, and the current version of the job is usually a principal or a sales associate retyping portal numbers into an email at nine at night. A listing description repeats constantly and has the highest value per minute saved, but it also carries the most regulatory exposure, so I would rather a team had one workflow already working before they take it on. Buyer and renter enquiry replies repeat more than anything else in the office and are the easiest to improve, because most of the time the reply is short and the failure is slowness rather than wording.

Then the second tier. Open home and inspection follow-ups are a good second task once the first one is running. Maintenance and arrears communication in the property management department is rule-bound in a way that suits written instructions unusually well. Appraisal preparation is the one people ask about first and the one I push back on, for reasons I will come to.

Pick by repetition and by how much of the job is assembly rather than judgement. A vendor update is roughly eighty percent assembly. An appraisal is roughly eighty percent judgement. That ratio, not the tool, decides whether the exercise works.

How does a vendor report actually get drafted?

You supply the campaign facts, the instructions describe the report you already write, and the model assembles the draft. It does not go looking for numbers, and it does not decide what the numbers mean.

The source material for one week of a campaign is small and you already have it: portal views and saves, enquiry count, inspection attendance, the buyer feedback you took down, offers received, and what was done on the campaign that week. The instructions name the vendor, the property, the stage of the campaign and the tone you use with that particular seller, then set the limits: use only the supplied figures, never estimate a number that is missing, flag a gap instead of filling it, and keep the recommendation section empty for the agent to complete. We teach that shape as five written inputs that turn a repeated job into instructions you can run again, and there is a longer walkthrough of the agency version in our guide to AI vendor report workflows for real estate.

The interesting part is the sentence most vendor reports end with. Any statement about where the campaign is heading is a claim about the future, and under the Australian Consumer Law a business making a claim about future matters has to have reasonable grounds for it when it makes the claim, with the burden of showing those grounds sitting on the business. 1 "We expect strong competition at the four week mark" is a forecast. If a model generated it from nothing more than a pattern in the training data, there are no grounds behind it at all. So the rule I write into an agency's vendor report workflow is blunt: the assistant reports what happened, and the agent writes what happens next.

This is the same shape as the proposal work in my own business. At Searchmaxxed we built AI workflows for blog creation, SEO and custom proposals, and the proposal one only started producing something we would send once we stopped asking for a good proposal and started supplying the actual material a proposal is assembled from. A disconnected chat guesses at your business. An assistant given the right context produces a draft you recognise. That difference is most of the value in the room, and it is not a tool you buy, it is instructions you write.

Listing copy is the highest-risk output in the office

Everything in a property advertisement is a representation, and the law does not care which keyboard typed it.

The ACCC's position on this is not subtle. Information about prices, images and descriptions of what is offered, and claims about the value, benefits, qualities or performance of what is being sold must be accurate, truthful and based on reasonable grounds, and that rule applies to any communication by a business, including advertising, social media, websites and information provided by staff verbally or in writing. It makes no difference whether a business intends to mislead or not. 1 There is no carve-out for a first draft, and "the model wrote it" is not a state of mind that helps you, because intention was never the test.

Queensland adds its own layer on top. Licensed property agents must not engage in misleading or deceptive conduct when advertising property for sale, an offers-over price is supposed to be the minimum the vendor is willing to accept, and setting a property price knowing the vendor will not sell at that price is false advertising and gets called bait advertising. The reference point is the vendor's minimum selling price recorded on the Form 6 appointment under the Property Occupations Act 2014, and it is a false or misleading representation if the agent knows the seller has put incorrect details on that form. The maximum penalties under the Australian Consumer Law run to $2.5 million per breach for an individual and, for a corporation, the greater of $100 million, three times the value of the benefit received, or 30% of annual turnover during the breach period where the benefit cannot be calculated. 2

So the listing workflow we build in the room is deliberately narrow. The source layer is the agent's own inspection notes, the contract and title details, the building report if there is one, and the Form 6 figures. The rules layer says: describe only features present in the supplied notes, never infer a room count or a land size, never describe a renovation, an approval or a school catchment that is not in the source, carry the price exactly as supplied, and list anything the copy needed but did not have. Then the agent reads the draft against the notes before it touches a portal. We cover that reading step on its own in how to check an AI draft against your source material, because it is a skill, and it is the part most people skip once the drafts start looking polished.

That polish is the actual hazard. A generated description is fluent before it is accurate, and fluency is what stops people reading carefully.

AI should never produce the price

An appraisal figure is a professional judgement made by a licensed person who has stood in the property. No part of that belongs in a generated draft, including the version where someone plans to check the number afterwards.

Three reasons, and the third is the one people miss. First, the number is a representation about price, and price claims sit squarely inside the accuracy rules above. 1 Second, in Queensland the advertised figure is tied to the minimum the vendor will accept on the Form 6, so a number invented anywhere in the chain contaminates the appointment and the advertising at once. 2 Third, anchoring. If a draft arrives with $845,000 already in it, the review is no longer a valuation, it is a negotiation with a sentence. People check numbers they generated far less rigorously than numbers they had to produce.

What AI can do around an appraisal is the preparation. Take the comparable sales the agent has already pulled and arrange them into a clean table. Turn the agent's dictated walk-through into structured property notes. Draft the covering letter around a blank the agent fills in. Assemble the market commentary from data the agent supplied and dated. The presentation gets faster, the figure stays human, and nobody has to explain later how a number reached a vendor.

International courses are less careful here than Australian practice requires. The Coursera program built by AI CERTs devotes a full module to AI in property valuation and price prediction, including automated valuation models and case studies on Zillow's Zestimate and CoreLogic's valuation models, and lists "improve property valuation and pricing accuracy" among its outcomes. 5 That material is real and it is interesting, but an automated valuation model is a data product, not an appraisal by a licensed Queensland agent, and a team that blurs the two has a compliance problem rather than a productivity gain.

What does an enquiry reply workflow look like in an agency?

It answers the question that was actually asked, from facts the office already holds, and it stops where the licensed judgement starts.

Most agency enquiries are narrow. Is it still available, when is the next inspection, does it allow pets, what are the rates and body corporate fees, is the price negotiable. The first four are lookups. The fifth is not, and the boundary between them is the whole design. So the rules layer gets an explicit stop list: do not comment on what the vendor will accept, do not interpret a contract clause, do not give advice on finance or building and pest, and do not commit to an inspection time that is not on the supplied schedule. When an enquiry touches one of those, the workflow drafts a short holding reply and routes it to the agent. That routing decision is the part teams find genuinely new, and we go through the general pattern in AI workflows for handling business enquiries.

The closest thing I have to direct experience here is outside real estate. For Glen House Homoeopathy we automated SEO, enquiries and admin, and the lesson that transferred was about scope rather than speed: the enquiries that could be answered from the practice's own information were answered quickly and consistently, and everything else was better off arriving at a person with the context already gathered. An agency is the same problem with a licence attached.

Property management has the clearest rules to write down

Nothing in an agency converts into written instructions as cleanly as tenancy communication, because Queensland has already published the numbers.

A Notice to remedy breach for a general breach gives seven days. For unpaid rent it is seven days, once rent has been unpaid for seven days, and rooming accommodation runs on shorter timeframes again. 3 Entry notice periods are just as specific: seven days for a routine inspection and no more than once every three months, 48 hours for repairs or maintenance, 48 hours to show the property to a prospective purchaser, and entry has to happen between 8am and 6pm Monday to Saturday unless the tenant agrees otherwise. 4

The important distinction in the room is which document the workflow touches. The statutory forms are prescribed and they come from the Residential Tenancies Authority, so nothing is generated there. What gets drafted is the human message wrapped around them: the maintenance update to an owner, the SMS telling a tenant a tradesperson is coming and when, the arrears message that opens with a repayment conversation rather than a threat, and the covering note that goes with a Form 11. The rules layer carries the timeframe so that the covering message can never contradict the notice it accompanies, which is exactly the error a busy property manager makes at 4:55pm on a Friday. If property management is the department with the problem, the property management page covers that side of an agency in more detail.

Arrears deserves one more sentence. The tone of an arrears message is a commercial decision, not a drafting one. A tenant in genuine hardship who receives a well-written but cold notice becomes a QCAT application; the same tenant who receives a message offering a repayment schedule often does not. Write that preference into the rules, because a model left to its own devices defaults to formal and slightly legalistic, which is the wrong register for the first contact.

What gets checked before anything leaves the office

Names, addresses, dates, figures, and every claim that was not in the source. Then the licensed person approves it.

That check takes about ninety seconds on a vendor report and about three minutes on a listing, and it is the step the whole session is built to make possible. A workflow that produces a draft nobody can verify against anything is worse than typing it yourself, because it feels finished. So in the workshop we keep the source material beside the output the entire time, and everyone runs their first check on a safe example rather than a live campaign. Choosing that example properly matters more than people expect, and preparing safe material for AI training is better sorted out before the day than during it.

One agency-specific rule: decide in advance who the approver is for each output type. For a listing, it is the licensee or the listing agent. For a vendor report, usually the agent on the campaign. For an arrears notice, the property manager and often the department head. Write the names down. An approval step that belongs to everybody belongs to nobody, and that is how an unchecked description reaches a portal.

How the public workshop and private onsite training differ for an agency

The public workshop suits one or two people from an office. Private onsite training suits a department that has to run the same workflow the same way.

The workshop page is where the date goes up, and you can register your interest there in the meantime. For a whole sales team or a property management department, private AI training for your team starts with a conversation about the tasks, the team size and the location, and the quote is fixed after that. Team size, session length and travel are what move the number.

There is a second rung, and agencies ask about it often enough that it belongs here. Training teaches your team to build and run the workflow themselves. Implementation is where the system gets built and connected for you. Both are ours, and which one suits you depends on whether the office has someone who will own the workflow after the session. It is a separate engagement with its own scope and price, so the honest answer is to raise it in the enquiry and let us scope it rather than guess at it on a web page.

How this compares with the online real estate AI courses

The online courses are cheaper, longer and broader. This workshop is shorter, more expensive per hour and narrower, and you leave with one workflow running on your own material.

That is our own comparison, not an independent ranking, and we sell one of the options in it. Read the row that matters for your situation. If you want CPD points in Western Australia, Aspire is the answer and we are not; their own page is clear that points cannot be processed unless your registration there is current. 6 If you want a structured grounding in how AI is used across the industry and you have a fortnight at ten hours a week to give it, the Coursera program covers far more ground than a full day ever will. 5 If you want a certification to put on a profile, McKissock sells one. 7

What none of them do is sit next to your actual vendor report while you write the rules for it. Adventures in CRE put the gap well: more commercial real estate professionals are using AI every day, but using AI and understanding how to engineer an AI output are two very different things, and that distinction is going to matter a lot. 8 I agree with them, and it is the reason this is a full-day build rather than a course. There is a broader version of that argument in our comparison of an in-person workshop and an online course.

Frequently asked questions