The Real Estate AI Stack Has Become a Routing Problem

Local AI has come up in Agency Intelligence before, and for good reason. The last few weeks have added enough new evidence to make the question feel more urgent, not because the earlier argument was underdone, but because pricing, hardware, open-source models, routing tools, compliance, privacy and the way agents actually work are all pushing on the same point... where should the intelligence run?

For a while, the default assumption was fairly simple. The work went to a cloud model, the answer came back, and most users did not think too much about the real cost underneath it. That made sense while the first wave of AI usage was light, subsidised and experimental: listing copy, email drafts, summaries, social posts, basic research. A fixed monthly subscription felt close enough to all-you-can-eat that people used the tools loosely, and often quite casually.

The economics around that are now less forgiving. Commonwealth Bank chief executive Matt Comyn made the point last week when he warned that corporate AI costs are getting harder to predict as tasks involve more reasoning, more tools and more context. The cost does not behave like a neat line item once a model is running a longer workflow, pulling in documents, checking data, using tools and moving through context again and again. That is a very different thing from asking for a cleaner subject line or a shorter property description.

You can see the same tension in some of the larger company examples now being reported. Walmart has put token limits around an internal AI coding tool, partly because staff were apparently asking it to solve similar problems again and again. Uber has also been reported as putting limits around employee use of AI tools after spending moved faster than expected. None of that reads like a retreat from AI. It reads like a market learning that usage and value are not the same thing.

Real estate has its own version of this problem. A lot of the early industry conversation treated AI usage as an obvious good... more drafts, more notes, more follow-up, more summaries, more content, more automation. As the work gets heavier, and more of the value depends on context moving through the system, that assumption becomes harder to defend. The agency does not need AI everywhere. It needs a sharper view of which work deserves the most expensive model in the stack, and which work can be handled closer to the business.

There is a Jevons-like quality to this, but with a slightly awkward twist. Cheap AI encouraged loose usage. Metered AI forces specificity... then, once the toll becomes visible, the bypass gets more attractive. Smaller models, specialised models, local models, office-network models, laptop-class AI machines and routing layers all become more commercially interesting when the premium cloud model looks less like a utility and more like a toll road.

That toll road idea is familiar in real estate. When one layer of the market becomes powerful enough to charge heavily because everyone feels they have to use it, the market usually looks for routes around it. The dominant path does not disappear, but alternatives get more attention... direct channels, private databases, pre-market strategies, cheaper listing pathways, competing interfaces. The toll road extracts value, but it also energises competition.

AI may follow a similar shape. If the cloud model is cheap, simple and good enough, most people will use it without thinking too hard. If it becomes expensive, metered and harder to control, the commercial case for routing more work around it gets stronger. OpenAI, Anthropic, Google and Microsoft still have an important role, but fewer tasks should automatically be pushed through the most expensive lane.

The model market is already pointing that way. Microsoft said at Build that one tuned MAI model was comparable to GPT-5.4 on public and private benchmarks while being up to ten times more efficient, and that a McKinsey-tuned MAI model outperformed GPT-5.5 on quality while being ten times lower on cost. Those are Microsoft’s claims, and they should be treated as claims, especially because they relate to tuned models rather than a generic model dropped into any workflow. Even with that caution, the signal is useful... once a model is tuned around specific work, the gap between “frontier” and “commercially good enough” can close very quickly.

Routing tools fit the same pattern. Google’s FunctionGemma is a useful signal because Google describes it as a small open model built for function calling, able to act as an offline agent for private tasks or as a traffic controller for larger connected systems. In plain terms, common commands can be handled locally while harder work gets escalated. That is a much more mature shape than one model trying to do everything.

NVIDIA’s RTX Spark is the hardware version of the same argument. It carries more than the usual faster-PC message. NVIDIA and Microsoft are positioning it around personal AI agents on Windows machines, with up to 128GB of unified memory, local security controls and enough capacity to run very large models with substantial local context. The first wave will sit closer to premium users, developers and creators than ordinary agency desks, but the signal still counts. The Windows PC market is now being pointed toward local AI, not cloud access alone.

That widens the conversation beyond Apple Silicon, iPhones and more technical local-network setups. Plenty of real estate offices still live on Windows laptops and office desktops, and plenty of agents still use the laptop as their mobile production base. Campaign work, appraisal preparation, reporting, image handling, database work and admin already happen there. If those machines can carry more local intelligence, the laptop becomes more than a screen for cloud software. It becomes part of the operating layer.

The Handy example is a small version of the same thing, which is why it is useful. A speech-to-text tool that could easily become another monthly AI subscription can, in some cases, be replaced by a free local app running models such as Whisper or Parakeet on the machine itself. It is just transcription, but that is the point. One ordinary task moves from a cloud subscription to a local utility, with practical benefits: faster in some settings, more private, less dependent on another monthly bill, and unmetered at the point of work.

Repeat that pattern across enough small tasks and the economics feel different. Not free, because hardware, setup, training, security and maintenance still count. But much less exposed to per-use pricing for work that happens all day.

Real estate is a good industry for this because the work is mobile, context-heavy and increasingly sensitive from a compliance point of view. An agent carries a practical world model that no AI system automatically has. They know the owner, the street, the property’s awkward room, the buyer who said very little but asked serious questions, the vendor who says they are relaxed but clearly watches every comparable sale, the local objection that keeps appearing at opens, and the difference between a soft enquiry and genuine intent.

AI can help organise that context, but first it has to get access to it. From there the cost and control questions become more than technical details. If every useful workflow requires voice notes, emails, CRM records, buyer histories, appraisal notes, inspection comments, property folders, photos and vendor concerns to be pushed into a premium cloud model, the agency is taking on cost and privacy risk every time. If more of that processing can happen on the phone, on the laptop or inside a controlled office environment, the trade-off changes.

A sensible real estate setup probably has layers. The phone is the capture layer... voice notes, inspection comments, simple image guidance, reminders and quick follow-up. The laptop is the mobile production layer... property folders, transcript clean-up, image handling, campaign drafts, appraisal preparation and local file search. The office network or private agency environment is where shared memory makes more sense... CRM intelligence, buyer matching, database opportunity search, compliance records, vendor-report assembly and campaign feedback. The cloud frontier model then becomes the escalation layer for harder reasoning, deeper research, strategy, complex writing and edge cases where extra capability is worth the cost.

That is a different way to think about AI than simply asking which model is best. The best model for a pricing strategy discussion may not be the best model for transcribing an inspection note. The best model for complex research may not be the right model for grouping buyer feedback. The best model for coding a new internal workflow may not be needed for first-draft listing copy. A good system needs to know the difference.

Property marketing is one of the more obvious places where the distinction shows up. AI can clearly help with copy, photo selection, image assessment, video workflows and campaign material. But real estate has rules and expectations around representation, and those do not disappear because the software suggests a nicer version of the room. A model can say a satellite dish would look better removed, or that a powerline should disappear, or that a crack, shadow or visual obstruction makes the image less appealing. The agent still has to know what can and cannot be changed.

The same logic applies to appraisal work. Some of the data-heavy preparation could become much easier with local or private models: pulling comparable sales, checking suburb activity, reviewing old appraisals, finding past contact history, summarising buyer demand, surfacing relevant notes and preparing the agent before the appointment. But the recommendation still belongs with a person. The model can prepare the ground. The agent reads the owner, the motivation, the timing, the risk, the competing properties and the local reality.

Inside the CRM, the case for local or private intelligence may be even stronger. Many agencies already have useful signals in their database: past appraisals, old buyer enquiries, owners who nearly sold, buyers who inspected repeatedly, vendors who paused, people who asked for contracts, suburbs or property types that keep appearing in enquiry patterns. A lot of that work does not need the most advanced model in the world. It needs clean records, structured retrieval, sensible rules, enough intelligence to find the signal, and a workflow that brings the opportunity back into view at the right time.

That work also sits close to the client relationship. In Australia, the timing is worth paying attention to because real estate businesses providing designated services will have AML/CTF obligations from 1 July 2026. Cleaner intake, better records, stronger audit trails, controlled access and more disciplined handling of client information are going to become harder to avoid. That does not mean every AI workflow must run locally, but it makes the old habit of casually uploading sensitive material into whichever tool is easiest feel less defensible.

For individual agents and smaller agencies, this could become a useful advantage. Larger businesses have bigger budgets, more IT support and more vendor relationships, but they also have more inherited process, more approvals and more difficulty changing how work moves. A smaller agency can be more deliberate. It can decide that voice capture runs locally, that CRM intelligence stays inside a controlled environment, that routine drafting uses a cheaper model, and that only the harder edge cases go to the frontier provider.

There are trade-offs. Local models can be weaker, more awkward to maintain, less polished and less reliable for certain work. Open-source tools still need setup and oversight. A workflow that works well for one person on one laptop may fall over when five people need the same record, the same permissions and the same audit trail. A model that is fine for internal sorting may not be good enough for client-facing advice, and hardware has a real cost even when the tokens feel free.

For most real estate work, being two or three months behind the frontier probably is not the issue. The gap matters in advanced coding, complex research, high-end technical reasoning or difficult creative work. It matters much less for transcription, open-home feedback grouping, CRM note clean-up, first-draft vendor updates, buyer categorisation, routine file search, internal summaries and a lot of property marketing preparation.

The agency stack probably lands as a mix of phone, laptop, office network, CRM, portal, cloud model and human judgement. Some work will be done locally because it is faster, cheaper or more private. Some will sit inside the agency system because it needs shared memory and control. Some will go to the cloud because the extra capability is worth the fee. Some should stay with people because the risk, context or relationship demands it.

So the useful question is no longer just what AI can do for an agency. It is where the work should live, what should be routed around the toll road, and when the toll is worth paying. The agencies that get that right may not be the ones using the most AI... they may simply be the ones wasting the least of it.