Real Estate Work Is Training Its Replacement

A lot of the AI jobs conversation is still being argued at the wrong level. One side keeps pointing to job losses, the other keeps pointing to employment data, new AI roles and the amount of work still being done by people. Both can be partly right, which is why the argument feels more slippery than it should.

The more useful place to look is what happens in the middle, where the job still exists but the work inside it is being made more visible. Some roles can look more resilient than expected because the person is still needed in the loop, correcting the system, carrying the exceptions, approving the final answer, cleaning up the edge cases and keeping the work safe enough to run. From the outside, that can look like the job has survived, but inside the workflow something else may be happening: the system is learning from the way the work is done.

Looked at this way, the jobs argument gets more complicated than a clean pendulum swing from “AI takes work” to “AI creates work”. A role can be useful today because the system still needs the person, while the actions of that same person become part of the evidence that makes the system more capable tomorrow. It is a slightly uncomfortable thought, but I think it is becoming one of the more important labour-market patterns to watch.

The OpenAI and Thrive tax example is useful because it is practical rather than speculative. OpenAI says Tax AI processed 7,000 returns across Crete firms, saved practitioners about a third of their preparation time, drafted returns with up to 97% accuracy and increased throughput by about 50%. The more important detail is how the system improved, with OpenAI describing a loop built around practitioner feedback, production traces and Codex-driven evaluation, where expert corrections become structured evidence the system can use to improve over time.

That is very different to a chatbot answering a tax question. This is AI sitting inside a live professional workflow, watching where the human had to correct it, preserving enough context to understand the correction, and then turning repeated patterns into something the product can improve against. Reuters had already reported the broader Thrive arrangement in similar terms, with OpenAI taking a stake in Thrive Holdings and the collaboration focused on professional-services AI that uses domain-expert feedback to train and improve specialised models.

You can see a similar shape in Meta’s internal work. Reuters reported that Meta was installing software on US employee computers to capture mouse movements, clicks, keystrokes and occasional screen snapshots to help train AI models for computer-based work. Meta’s stated aim was to give its models real examples of how people use software to complete tasks, including the small interactions that models often struggle to replicate.

Cursor is another version of the same pattern, this time in software development. Reuters reported that SpaceX is buying Anysphere, the company behind Cursor, and that Cursor’s access to developer data, including coding requests and design decisions, was part of what could help improve models such as Grok. The tool layer becomes strategically valuable because it sits close to the work itself, seeing the request, the draft, the correction, the accepted code, the rejected answer and the judgement around what was actually useful. Across each of these examples, the useful pattern is observation, correction and improvement inside the workflow.

That ports across to real estate more cleanly than it first appears. Sales is still comparatively protected by trust, judgement, negotiation, local context, timing, vendor confidence and the incentive structure around commission. The agent at the front of the relationship is not the easiest place for this to land first, partly because the value of the work is not only in the visible task, it is also in reading the person, holding confidence, managing tension, knowing when to push and knowing when to slow the conversation down.

Property management looks different, because while there is judgement in it, and plenty of moments where experience counts, a large property management operation also carries a huge amount of repeated coordination: rent collection, arrears, owner statements, supplier invoices, maintenance requests, lease renewals, routine inspection reports, compliance checks, tenant communication, owner updates, complaints, insurance notes, water usage, bond issues, repairs being quoted, approved, followed up and closed out. A lot of that work already runs through software, and a lot of it leaves a trail.

That makes property management feel much closer to accounting than sales. Not identical, because the physical and relationship elements are different, but similar enough in the way repeated decisions are made across a large base of clients, documents, requests, exceptions and approvals. Thousands of properties moving through the same platform can create a very rich picture of how the work is actually done.

A property management platform does not only hold records, it can see how work moves around those records: which arrears message gets sent and when, which maintenance issue gets escalated, which owner update is rewritten before sending, which supplier is selected, which invoice is queried, which tenant message needs care rather than automation, which inspection comment gets softened, clarified or turned into a repair request, and which task looks routine until a property manager intervenes. That is the training set, and the value sits less in the static record than in the work around it.

The rent roll is still important, but the more valuable signal may be the timing, the sequence, the correction, the escalation, the human override and the repeated pattern of what good looks like. In the Thrive example, the accountant’s correction is not simply a correction, it becomes evidence. The same logic can apply anywhere a system can see enough repeated work and enough expert adjustment.

I am not suggesting a specific Australian property management platform is doing this in the way Meta is reportedly doing it with employee computers. The point is structural. A platform sitting across trust accounting, maintenance, inspections, communication and reporting is much closer to the production layer than most people probably think. If enough work passes through it, and if enough corrections and approvals are captured, the platform can learn more than the agency may realise.

The labour picture can almost look like Schrödinger’s cat. The job can appear safe because the person is still sitting there doing the work, while the work itself is becoming more legible to the system. In year one, the AI drafts the owner update and the property manager fixes it. In year two, the system has seen enough corrections to draft it more accurately, route the maintenance better, classify the urgency more reliably and prepare more of the routine work without as much help. In year three, the person may still be there, but the number of properties they can carry, the support they need, and the nature of the role may look quite different.

That does not mean property managers disappear, which is too blunt to be useful. The better conclusion is that the role gets thinned and reorganised around the parts where human judgement is still needed. The routine coordination compresses, the exception handling becomes more important, and the person is increasingly there for judgement, escalation, relationship management and accountability rather than every small piece of process.

The same pattern applies to sales support, but it probably plays out differently. The sales assistant does not need to disappear for the economics to change. One assistant might support more agents because the system can carry more of the CRM clean-up, buyer follow-up, inspection feedback, vendor reporting, portal updates, campaign preparation and listing production. Then, over time, some of that work may move into the CRM, the portal, the outsourced provider, the marketing system or a digital assistant working across the whole agency.

Outsourcing is worth watching for the same reason. A service provider handling the same task for hundreds of agencies does not just have labour, it has pattern recognition. If it edits thousands of real estate images, prepares thousands of listing campaigns, handles thousands of maintenance requests or produces thousands of vendor reports, it sees the repeated work at a scale no individual agency can match. If AI can observe enough of that work and learn from enough human corrections, the provider may eventually sell back less labour and more system.

This is where the phrase “AI takes your job” becomes more precise. In many cases, AI does not arrive from outside the building and take the whole role in one movement. It learns the role from the inside, through the work people are already doing. Self-driving has followed a version of this logic for years, using real-world driving data rather than trying to solve the whole problem from theory alone, and white-collar work is now getting its version of that pattern. The model learns less from a textbook description of the job and more from watching the job being done, corrected and approved in production.

The uncomfortable part is that the person can be most valuable during the period where the system is not yet good enough. They are needed because the AI still makes mistakes, misses context, mishandles edge cases and needs someone to protect quality, but every correction can also help reduce the need for that correction next time, especially where the work is structured, repeated and captured well.

Real estate has plenty of work like that. Vendor reporting is structured enough to be assembled by a system but still needs judgement before it reaches the client, buyer feedback can be grouped, summarised and compared across opens, but the agent still needs to know which comment carries weight, property management maintenance can be triaged, drafted and routed, but a human still needs to know when a tenant, owner or property requires a different touch, and trust accounting and compliance will need strong controls, even if the checking, flagging and preparation layers are not immune just because the final responsibility stays human.

The old “help me with this” version of AI is not the more important end state. That version is useful, but it is still mostly synchronous: ask for a draft, ask for a summary, ask for a checklist, take the output and move it somewhere else. It helps the person, but the person still holds the workflow together.

The more serious version resembles a digital employee sitting inside the workflow. It sees the record, the instruction, the message, the correction, the approval and the next step. It can carry more of the routine process, ask for approval where the risk is higher, and improve from the way people keep changing its work. That is a different operating model. The system is no longer just a writing tool sitting beside the work, it is part of the way the work moves.

For real estate, this is important because so much of the business is made up of small, repeated decisions that rarely look important on their own: how an owner update is worded, when a buyer is followed up, which maintenance request is urgent, which tenant message needs care, which vendor needs a call rather than another email, and which open-home comment is noise rather than signal. A lot of agency work lives in those small judgements. Humans still carry most of that now, but over time the systems around them will learn more of the pattern.

The privacy and governance question sits under all of this, and property management makes it especially hard to ignore. Tenant applications, payment details, owner records, arrears, identity checks, maintenance histories, inspection notes and communication between parties are all sensitive. Even if a vendor says data is anonymised or aggregated, agency principals should be asking sharper questions about what is captured, who consented, whether staff activity is being monitored, whether tenant or owner data is used to improve a model, and whether the benefit flows back to the agency or mainly compounds inside the vendor’s product.

That is also where local models, private deployments and stronger data controls may become more than technical preferences. They may become part of the commercial architecture of the business. Agencies will still need modern platforms, and most will not build their own AI stack from scratch, but the agreement around data, training, retention, auditability and model improvement will be more important than it has been in the past.

For principals, the practical question is not only which AI tool to use, it is who gets smarter when your team does the work. If your people correct every draft, clean every workflow, fix every exception and teach the system what good looks like, does that intelligence come back into your agency as a real advantage, or does it mainly improve the vendor’s product for everyone else? If your outsourced provider handles more of the back office, are they just supplying capacity, or are they also building the training layer for a software product that eventually replaces part of that same capacity?

None of this is an argument against software vendors, outsourcing or AI. Agencies need all three in some form. It is an argument for being more commercially awake to where value is moving. The valuable asset is no longer just the database, the rent roll, the listing, the buyer enquiry or the client record. Increasingly, it is the workflow around those things, especially the corrections and decisions that show how the work should actually be done.

This next phase of AI in real estate may be quieter than people expect. It may not arrive as a robot replacing an agent. It may arrive as one property manager carrying more properties, one assistant supporting more agents, one offshore team doing more work with fewer people, one platform absorbing more of the coordination layer, and one agency after another realising that parts of the support structure that used to feel necessary now feel heavier than they should.

The human side of real estate is still central to the best parts of the business, and may become even more important as the surrounding process work becomes cheaper and more automated. But the work wrapped around that human judgement is becoming more visible, more repeatable and more trainable, and once that happens it stops being only work. It becomes the model’s next lesson.