When Real Estate AI Runs All Day

Most agents still meet AI at the visible edge of the business... listing copy, email drafts, social posts, call summaries, maybe CRM notes. That was always the easy doorway because the task was obvious, the result was quick, and the cost did not feel especially threatening when one person asked for one piece of output at a time.

The harder commercial question comes when AI moves underneath that visible layer. Not in the sense of a robot running the office, more in the ordinary sense that software begins listening to calls, reading enquiries, checking documents, matching buyers, preparing vendor updates, routing tasks, looking for appraisal opportunities and carrying more of the assistant work that used to sit around the agent. A chat box gets used in moments... an operating layer runs through the day.

That changes the cost model. Real estate software has mostly been sold by seat, office, module, listing, campaign or transaction. Token-based AI is different because every summary, match, check, draft, classification, voice instruction and workflow action has some kind of inference cost underneath it. When that happens occasionally, it looks like another subscription. When it happens all day, across a team, office, group or network, it starts to look like part of the operating model.

Analysing Cursor is useful because it gives a live example of that pressure playing out in a high-value workflow. Its Composer 2.5 model is built on Moonshot's Kimi K2.5 open-source checkpoint and priced at US$0.50 per million input tokens and US$2.50 per million output tokens (where the 2.5 model is widely seen as being 'comparable' for coding to GPT 5.5 Medium). That sits well below the headline pricing for the heavier frontier models from OpenAI and Anthropic. The comparison is not perfect because these are different models, different tasks and different products, but the commercial point is fairly clear. If a specialised model is close enough for the work that matters, at a materially lower cost, it becomes harder to justify sending every ordinary workflow step through the most expensive model available.

Real estate is a good industry for that distinction because a lot of agency work does not need the heaviest frontier model every time. There will always be moments where deeper reasoning, legal sensitivity, negotiation context or high-stakes client communication justify a stronger model. But much of the operating layer is classification, summarisation, retrieval, matching, extraction, drafting, routing and checking. In that layer, a cheaper model that understands the work may be more useful than a general model that knows more about everything else.

That is not an argument for every agency to go and build its own model. For most agents and smaller offices, that would be a distraction, and probably an expensive one. The more useful lesson from Cursor is the pattern: take a strong open model foundation, adapt it around a specific workflow, wrap it in a product people already use, and make the economics work at scale. Coding is the example today. Law and finance are already seeing their own versions of it. Real estate will almost certainly get one too.

Kiraa points to a different version of the same underlying problem. Errol Brandt and the team are building private enterprise intelligence that runs on-premise and works from a company’s own data, terminology and business logic. Kiraa is not a consumer on-device assistant, and it is not a small agency chatbot. It is closer to a private intelligence layer for businesses that want answers grounded inside their own operating environment rather than constantly pushing proprietary data through a generic external model layer.

For large real estate networks, that kind of thinking becomes very relevant. A Ray White, Harcourts Real Estate, LJ Hooker, Stockdale & Leggo or other major group has enough data, internal process, listing activity, appraisal history, buyer enquiry volume, property management information and office-level knowledge to think differently from an individual agent. At that scale, the question is not only what AI can do, it is whether the network should keep renting intelligence one call at a time, or own more of the capability itself through a private model layer, a partner like Kiraa, or a real estate-specific vertical system built on lower-cost models.

That would not be cheap. Hardware, data architecture, security, model evaluation, licensing, workflow design, maintenance and governance all cost real money. A large network could spend millions before the result is dependable. But the comparison changes once the system is running across hundreds of offices and thousands of users. The cost moves away from paying a frontier provider every time the business thinks, and closer to running an internal or semi-private capability with hosting, maintenance and improvement costs around it.

Most of my audience is not sitting inside a national head office with that kind of budget. For individual agents, boutique offices and smaller groups, the more practical question is what part of this eventually becomes available through trusted vertical providers. A smaller agency is unlikely to train a Kimi-style model, maintain a private inference layer, or build its own compliance engine from scratch. But it may use a real estate-specific product that has done that work already, with better security, state-based compliance logic, cleaner permissions and a lower operating cost than a loose collection of general chat tools.

That middle path is probably where a lot of the market ends up. Not fully public, not fully in-house, but more specialised and controlled than a personal AI subscription. A smaller group may end up using a private-cloud style product, a real estate-specific model layer, or an on-device capture tool connected to a governed agency system. The benefit is not that the agency becomes a technology company. The benefit is that it gets access to better workflow intelligence without carrying all the technical weight itself.

The privacy and compliance side may matter even more than the token bill. It is important to be precise here because the major commercial API providers generally say they do not train on business data by default unless the customer opts in. That is worth acknowledging properly. But training is not the only issue. Data still moves outside the agency boundary unless the architecture prevents it. Connectors need permissions, logs exist somewhere, third-party software layers may learn workflow patterns, agents can be over-authorised, and prompt injection or poorly designed access rules can expose information that should never have been exposed.

Real estate is not a casual data environment. Sales files are sensitive enough, but property management can be much more sensitive again. Agencies handle tenancy applications, landlord details, tenant communication, identity material, maintenance issues, arrears conversations, routine inspections, dispute notes and long-running personal information. In Australia, privacy obligations, tenancy rules, commercial confidentiality and plain client trust all point in the same direction: unnecessary data movement should be treated as a risk, not just a convenience.

There is also a competitive layer that is easier to miss. Even where a model provider is not training on your business data, the software layer wrapped around the model can still learn from usage patterns, workflow demand and industry behaviour. Cursor’s advantage came from owning the coding environment where the work was happening. Anthropic is now pushing more directly into finance and legal workflows through Claude Cowork, plug-ins and industry templates. The lesson for real estate is not that those companies are doing anything wrong. It is simply that whoever sits closest to the workflow learns where the value is.

That matters because real estate compliance is local, practical and often expensive when it goes wrong. Victoria is an easy example. A residential listing needs a Statement of Information, the indicative selling price has to be a single price or a range of up to 10 per cent, comparable property rules apply, and advertising language like “offers above”, “from” or “+” is restricted. A current CRM might check whether a PDF has been uploaded before the listing goes live. A more intelligent system should be able to read the Statement of Information, compare it against the listing price, compare it against the search price, check the range, look for missing comparable sales or required explanations, and flag the issue before the agent has a compliance problem.

That kind of checking is not glamorous, but it is exactly where localised intelligence becomes useful. The same system could understand that Queensland, New South Wales and Victoria do not treat advertising, disclosure, search price, contract process or buyer communication in the same way. A general language model can answer those questions if prompted well, but a real estate operating layer should not need to be reminded every time. The rules should sit inside the system, close to the workflow, so the agent is protected while the work is happening.

On-device AI adds another piece. A phone or laptop model will not carry the full knowledge of a real estate network, and smaller local models still have real limits around memory, reasoning and context. But the device can still become the edge of the system. It can capture the live moment, turn a voice note into structured feedback, summarise a call, classify buyer intent, prepare a follow-up, tidy images or short video, and send the right information into the agency system without every minor action needing a cloud model.

That is probably where field work gets interesting... an agent finishes an open home, speaks a few notes into the phone, and the device turns those comments into usable campaign intelligence while the context is still fresh. The agency or network layer then checks buyer history, property records, vendor context, compliance rules and next steps. The local model does not need to be the whole brain. It needs to capture the moment properly, reduce friction, and call the stronger layer only when the work requires it.

The Jevons-style part matters as well. When intelligence becomes cheaper and easier to run, the business does not simply do the same work for less money. It finds new work worth doing. A vendor report can become less like a weekly document and more like a live campaign view. Buyer matching can run continuously against new listings and old enquiries. Property management triage can classify and route routine issues before a human touches them. Appraisal opportunities can surface from old enquiries, ownership patterns, past conversations and local movement.

That is a different kind of capacity. It is not just cheaper admin. It is work that might not have happened at all because it was too expensive, too manual or too dependent on someone remembering at the right moment. A frontier model running every one of those checks all day may be too expensive for many agencies. A local, private, vertical or smaller task-specific model may make some of those loops practical.

For a small agency, that could be a serious competitive advantage. A larger office can sometimes absorb messy process because it has more people to carry the slack. A smaller one cannot, which is why cleaner loops matter so much. If the system remembers more, prepares more, follows up faster, checks compliance earlier and brings warm opportunities back into view, the smaller office can feel much larger than it is without copying the cost base of a larger competitor.

That is where the old 'two-people-running-from-the-lion' analogy still works. The aim is not to outrun the whole industry in one jump. It is to be better prepared than the competing agent when the next listing opportunity appears. If your system has already pulled the owner history, checked recent market movement, surfaced previous buyer interest, prepared the appraisal notes and flagged the compliance issues, you walk into the conversation with more leverage.

The labour question needs more care than it usually gets. Some support work will come under pressure, especially where the role is mostly repeated drafting, chasing, logging, copying information between systems or assembling reports. But a localised or vertical AI layer also creates a different kind of work. Someone still has to design the loops, manage exceptions, test outputs, protect client data, maintain the rules and understand when the system is technically correct but commercially clumsy.

That role is not purely technical. It is closer to an art director sitting above a designer, or an experienced operator sitting above a process team. The person needs enough real estate judgement to know when a vendor update is too blunt, when a buyer classification is too shallow, when a compliance alert is too noisy, or when a property management response is efficient but likely to annoy the tenant. The better agencies may not simply have fewer people... they may have different people carrying more leverage across cleaner systems.

For individual agents, the practical lesson is not to wait for enterprise AI to trickle down perfectly. It is to look at where the business is already leaking information and where lower-cost, more specialised intelligence could eventually carry more of the load. Buyer follow-up, vendor reporting, open-home feedback, compliance checking, appraisal preparation, property management triage, database mining and campaign monitoring are all ordinary enough to matter. They are also repeated enough that the economics become meaningful once AI is running through the workflow rather than sitting beside it.

That is why I think token cost is only part of the story. The larger issue is where the intelligence lives, who controls the data, who owns the learning loop, how compliance is handled, and whether smaller agencies get access to the same kind of leverage that larger groups may eventually build for themselves. The agents and principals who understand that earlier will probably make better decisions about which tools deserve their attention.

Real estate AI is moving beyond the chat box. The next useful layer will sit much closer to the work, and the best version of it will not just produce more drafts. It will help the agency think, check, remember, prepare and act more often, with less unnecessary data movement and a cleaner loop between information, action and judgement.