What Is a Real Estate Agency’s Alpha?

Two real estate agencies can buy access to the same AI model, connect much the same software and work from broadly similar market data. A year later, one system may understand how the office prepares an appraisal, reads buyer intent, diagnoses a campaign and handles exceptions, while the other still produces generic answers that have to be rebuilt every time. The gap will sit in what each business has captured, corrected and kept.

Financial markets use the word alpha to describe performance beyond the return available from simply following the market. In a real estate agency, I think of alpha as the repeatable advantage in how the business interprets information and turns it into action. It is the difference between holding a database and knowing what to do with it.

That advantage has always been difficult to isolate. Some of it sits in records, procedures and scripts, although much more lives in experienced people. A principal knows which appraisal is genuine and which owner is collecting prices. An agent hears a buyer describe the property they want, then recognises that their behaviour points somewhere else. A sales leader can look at the same enquiry, inspection and feedback numbers as everybody else and see whether the campaign needs patience, a change in presentation or a harder conversation about price.

AI is making more of that judgement visible to a system. Every instruction, approved example, correction, exception, evaluation and changed sequence can help explain how the agency works. The commercial value grows when those lessons improve the next appraisal brief, buyer follow-up, vendor report or negotiation plan. So does the importance of knowing where the lessons accumulate.

Where Agency Alpha Accumulates

Satya Nadella described a version of this problem in a 12 July essay as the reverse information paradox. His argument is that a business can pay for intelligence and then supply the proprietary knowledge required to make that intelligence useful. Prompts, tool use, corrections, evaluation criteria, traces and memory can all carry knowledge about how the organisation thinks and what it considers good work.

Nadella is an interested participant in this market, and his essay should be read that way. His argument arrived as enterprise AI moved closer to the operating layer. OpenAI's 21 July small-business announcement described an agent connected to company files and applications, with memory about how people think, write and work. It also promoted prompts, guides, partner skills and multi-step workflows for ordinary business tasks.

The product becomes more useful as it understands more of the business. In an agency, that understanding might begin with instructions about tone or report format. It becomes more valuable when the system can see which comparable sales experienced agents reject, how a weak appraisal lead is qualified, which buyer comments deserve more weight, when an automated follow-up should stop, or which change to a campaign requires the vendor's approval.

Richard Sutton's 2019 essay, The Bitter Lesson, helps explain why this shift is possible. Sutton's argument was about the history of artificial intelligence: general methods built around computation, search and learning have repeatedly outlasted attempts to hand-code expert knowledge. He was not writing about company ownership or real estate data.

Inside an agency, Sutton's learning argument shows up after the model reaches the office. A valuer removes a poor comparable and records the reason, a sales manager changes the point at which a vendor conversation should occur, an agent corrects a buyer classification after seeing what the buyer actually inspects. Those decisions can give the next task a better starting point.

Over time, judgement that once travelled through conversations and individual habits becomes legible through examples, corrections, outcomes and workflow changes. The agency can become more capable and less dependent on one person remembering everything. It can also leave that operating knowledge inside a product the agency may struggle to inspect or move.

I wrote last month about real estate work training its replacement. That issue followed the corrections and approvals made by staff and asked how a platform could learn more of the role over time. The company-level issue begins once the learning becomes useful. The agency needs it to survive an employee's departure, the end of a software contract and the arrival of a better model.

Ownership and the Exit Door

Provider training needs careful language. OpenAI's current business commitments say inputs and outputs from its listed business products and API platform are excluded from model training by default. Anthropic says retained data in the API arrangements covered by its documentation is never used for model training without express permission, although retention varies across models, features and deployments.

These commitments remove the broad assumption that every business prompt trains a public foundation model. An agency can still become dependent on memory, stored conversations, connected knowledge, instructions, agent skills, evaluation criteria and workflow logic that are difficult to move. It may own the underlying client and property records while facing a practical rebuild around them.

OpenAI's guidance still says data export is unavailable from a ChatGPT Business workspace. Anthropic allows a Claude Team or Enterprise Primary Owner to export organisation conversation and user data. Anthropic also says Claude memory data is included in exports and offers an experimental, text-based process for moving memory between services.

A Claude organisation therefore has an exit door for conversations and memory. A conversation export or memory summary will not recreate the retrieval structure that finds the right CRM record, the permission around a buyer's information, the evaluation set that catches a bad appraisal brief, the workflow version that governed a client action or the exception that changed the result.

Open-weight adoption moves the question beyond chat exports. Vercel's July production index reported that open-weight models handled 29 per cent of tokens moving through its AI Gateway in June, up from 11 per cent in April, while accounting for less than 4 per cent of spend. Roughly one in eight enterprise customers on that gateway used an open-weight model in production.

The top four US frontier labs still received 95 per cent of spend through the gateway. Anthropic received at least 72 per cent of spend in the higher-stakes use cases Vercel identified. Agencies are likely to encounter a mixed model market, with lower-cost or open-weight systems carrying some routine work and expensive frontier systems used where deeper reasoning or higher reliability justifies the cost.

On 27 July, NVIDIA and the Linux Foundation announced an industry alliance around open models and open tools for testing, tracing, auditing and governing AI agents. By 30 July, Microsoft's separate open-weight letter listed more than 230 company and organisation signatories, including model developers, cloud providers and enterprise software businesses.

Those organisations have commercial reasons to promote an open ecosystem, and their advocacy needs the same caution as any vendor claim. Their announcements extend into the harnesses, logs, guardrails and evaluations around the model, because that surrounding layer determines how an agent behaves inside a real workflow.

OpenAI's current gpt-oss documentation says its open-weight models can run on infrastructure controlled by the organisation or a hosting provider, and the weights can be adapted with open tooling. It also notes that surrounding infrastructure may remain proprietary and that self-hosted deployments are the operator's responsibility. Access to the weights creates options. The agency still has to build, secure and maintain the system around them.

What Has to Be There on Monday

Imagine the agency changes a model provider on Friday afternoon. Its CRM records, documents, transcripts and generated outputs are the familiar data layer. They need known storage, access, retention and export arrangements.

On Monday, the new model also needs to see how the agency works. The appraisal examples with accepted and rejected comparables, the corrected buyer classifications, the campaign decisions and the cases that define a useful vendor report form the learning layer. They show what experienced people accepted, changed and refused.

The work then has to move in the right order. Instructions, approval points, escalation rules, process versions, agent skills and system connections turn an answer into an agency action. They form the workflow layer. Copying a prompt into a document preserves only one piece.

Model control appears when the agency can reconnect those three layers to another suitable model and test the result before clients feel the change. The office may never exercise that choice. Retaining it keeps one supplier from becoming the permanent home of the agency's accumulated judgement.

Sovereign AI is often used for national infrastructure, private computing environments and models operated under tight jurisdictional control. The average real estate office has no need to build that version. Its practical version starts with keeping the canonical agency material somewhere the business controls.

Approved appraisal instructions, vendor-report standards, campaign-diagnosis examples, buyer-classification corrections and negotiation-preparation criteria should sit in an organised and versioned agency library. Individual chat histories are a poor substitute, especially when the person who created them can leave.

A small evaluation set can be built from representative work. It might include an appraisal brief with known errors, buyer notes that must remain observations rather than facts, a vendor report requiring careful tone, a campaign exception that needs escalation and a contract request where the document version must be checked. The set can be run before and after a model or workflow change to see what broke.

Corrections belong in the same library. When an experienced agent rejects a comparable, rewrites a buyer classification or changes the recommended next action, the business can record the reason, approved outcome and relevant process rule. De-identified or synthetic examples can carry the lesson when client information is unnecessary.

Australian guidance supports that discipline. The Australian Signals Directorate tells small businesses to review vendor settings, terms and privacy policies, identify sensitive and proprietary information, and understand data ownership, access, use and storage. The Office of the Australian Information Commissioner recommends due diligence throughout the product lifecycle and advises, as a matter of best practice, against entering personal or sensitive information into publicly available generative AI tools.

Contracted business products may provide stronger commitments and controls. The principal still needs to know what the agency can retrieve, export and reuse. A managed account with good privacy terms can protect client data while leaving the office dependent on workflow logic stored somewhere it cannot carry forward.

On Friday afternoon, changing a provider should mean replacing one part of the system. By Monday morning, the agency's instructions, approved examples, corrections, evaluations and workflow rules should still be available, and representative work should still meet the agency's standard.

Agencies will continue to rent intelligence from model providers and software platforms. That is sensible for businesses with no reason to build the underlying technology. The judgement accumulated through years of appraisals, buyer conversations, campaigns and negotiations should still be sitting inside the agency when the rented intelligence changes.