Where AI Fits in an Agent's Week, and the Fair Housing Line

The job is usually described as selling property. In practice an agent's week is dominated by three other things, and knowing the split determines where tooling helps.

Content production. Listing descriptions, brochures, social posts, market updates, email newsletters. High volume, moderate quality bar, repetitive, and it consumes hours that generate no direct revenue.

Communication. Responding to enquiries, chasing, updating vendors, coordinating between parties who each need different information. This is the largest single time consumer in most agencies and it is largely reactive.

Analysis. Pricing a property, assembling comparables, preparing a valuation argument. Lower volume, higher stakes, and it is the part that most resembles expertise.

And underneath all three, relationship work: being known, being trusted, being the person someone calls in four years. That is where the business actually comes from and none of it compresses.

So the fit is straightforward. Content production compresses substantially, communication compresses partially, analysis compresses conditionally, and the relationship work does not compress at all.

What makes this profession different from the others in this catalogue is not the capability question. It is that two of those three areas sit inside consumer protection and anti-discrimination law that predates AI by decades and applies with full force, which is the subject of the next step.

This is the constraint that makes real estate marketing different from ordinary marketing, and it catches agents who have no intention of discriminating.

Across most jurisdictions, housing discrimination law prohibits treating people differently on protected characteristics, and crucially it reaches advertising as well as decisions. In the United States the Fair Housing Act prohibits making, printing or publishing any advertisement indicating a preference or limitation based on a protected characteristic. European jurisdictions have equivalent provisions under equal treatment directives and national law.

What that means for generated copy. Phrases indicating who a property is suitable for can constitute an unlawful preference even when written innocently. Perfect for a young professional couple, ideal for a growing family, in a quiet Christian neighbourhood, close to the synagogue, suits an active person: each of these signals a preference about who should live there.

A model generating listing copy has learned from a corpus of listing copy, which historically contained a great deal of this language. So it will reproduce it, fluently, without any awareness that the phrasing carries legal weight.

The practical rule that follows: listing descriptions describe the property, not the buyer. A useful test is whether a phrase says something about the building or something about the person who should live in it. The second is where the exposure lives.

And the liability is the agent's. A model produced it is not a defence available to anyone.

Sorting agent tasks by exposure, which is a different sort from sorting by how well AI performs.

High exposure, high AI capability, which is the dangerous combination. Listing copy, where discrimination law reaches the wording. Targeted advertising, where delivery can skew even with neutral parameters and housing is a regulated ad category on major platforms. And any statement about a property's condition, which becomes a representation to a buyer.

Moderate exposure. Comparables and valuation, where the risk is a wrong number a client relies on rather than a discrimination claim.

Low exposure, high capability. Internal drafting, scheduling, transcribing viewings, summarising a long chain, preparing your own briefing notes. Nothing here reaches a consumer as a representation.

And low exposure, low capability: the relationship work.

The practical implication is that the sort order for adoption is the inverse of the obvious one. The most visible use, listing copy, carries the most regulatory weight, and the least visible uses carry almost none.

Listing descriptions are the most obvious use and they need a specific discipline rather than a general one.

What a model does well here. Turning a set of facts, room count, dimensions, features, condition, into readable prose. Producing several tonal variants. Adapting one description across portal, brochure and social formats. Translating for international buyers.

The two failure modes.

Discriminatory framing, covered above, which the model will produce because its training corpus contained it.

And invented features. A model given a sparse brief will produce a fuller description by adding plausible detail: period features, natural light, a sought-after location, proximity to good schools. Each of those, if untrue, is a misrepresentation to a buyer, and in property that carries consequences well beyond an advertising complaint.

So the pipeline is: supply the verified facts, instruct the model to use only what it is given, and review specifically for two things, an added feature and a statement about who should live there. That is a targeted review of about a minute, and it is the difference between a time saving and a liability.

One further point specific to this profession. Portal listings persist and are archived, and a listing that misdescribed a property is retrievable long after the sale. The permanence raises the cost of a fabricated detail relative to most marketing.

And where a jurisdiction requires specific disclosures, energy performance, tenure, known defects, those are not places for generated prose at all.

One category of listing content deserves separate treatment because it is where well-meaning agents most often cross a line.

Describing a neighbourhood is normal and expected. Describing it in ways that signal who lives there, or who would be comfortable there, is where discrimination law engages. And the signals are frequently indirect.

School quality is the clearest case. In many areas school catchment correlates strongly with the demographic composition of a neighbourhood, so promoting a property on school ratings can function as a proxy signal even when nobody intends it. Guidance in several jurisdictions treats school-quality claims in housing advertising with caution for exactly this reason, and some agencies prohibit them outright.

Similar issues attach to describing an area as safe, which frequently correlates with demographics, family-friendly, up-and-coming, or by reference to religious institutions.

A model will generate all of this readily, because it is standard listing language.

The practical positions available. Provide factual, verifiable information without characterisation: the school is at this distance, here is the published rating, from the official source. Or direct buyers to the public data and decline to characterise it, which several large agencies have adopted as policy.

The general principle worth carrying: describe the property and the verifiable facts about its surroundings, and let the buyer draw conclusions about suitability. That is both the safer position and, arguably, the more honest service.

Advertising a property is not ordinary advertising, and the platforms know it even where agents do not.

Housing is one of the categories major advertising platforms treat specially, restricting the targeting options available, following enforcement action over housing ads that could be targeted to exclude protected groups. So the tooling itself is more constrained than for other verticals.

What that means practically. Age, gender, postcode-level and interest-based targeting that is routine elsewhere may be unavailable or restricted for housing ads, and using a general marketing tool that does not know your ad is a housing ad can produce a compliance problem the platform later flags.

The subtler issue, which the marketing cursus covers generally and which is acute here. Even with neutral targeting parameters, delivery optimisation infers who is likely to engage and delivers accordingly, which has produced documented skew in housing advertising. An agent who selected no demographic targeting can still find their listing reached a demographically skewed audience, because the optimiser did that on its own.

The practical positions. Use the platform's housing-specific ad category, which exists precisely to apply the restrictions. Do not use general-purpose AI ad tools for property without checking they handle the special category. And where you can, review delivery reports for skew rather than assuming neutral inputs produced neutral reach.

The underlying point for the profession: the automation is more constrained here than in any comparable marketing context, and that constraint is deliberate.

Pricing is where an agent's expertise is most visible to a client, and it is where automated valuation has the longest track record and the clearest limits.

What automated valuation does well. Establishing a range from recent transactions of similar properties. Surfacing comparables you might not have found. Producing the supporting table for a valuation report. Tracking how a local market has moved.

Where it fails, and the failures are systematic rather than random. Automated models rely on recorded attributes, and property value depends heavily on attributes that are not recorded: the condition of the interior, whether the extension was done well, the aspect, the noise from the road that does not appear in any dataset, and whether the flat above has a problem. Two properties with identical recorded characteristics can differ substantially in value, and the difference is exactly what an agent walking through the door can see.

So the model provides the range and the agent provides the position within it, which is the same division as elsewhere in this catalogue.

Two cautions specific to valuation.

A client shown a generated valuation will anchor on it, and anchoring in property pricing has real consequences for a vendor's expectations and time to sell. Present a range rather than a number, and say what it excludes.

And in thin markets, few comparable transactions means wide uncertainty that automated tools tend to under-report. A confident number from a sparse comparable set is the most misleading output available here.

An adoption order that follows exposure rather than visibility.

First, the invisible uses. Transcribing and summarising viewings so you stop taking notes. Drafting your own briefing before a valuation. Summarising a long email chain so you can see where a chain transaction actually stands. Preparing the weekly vendor update from your own records. None of this reaches a consumer, all of it recovers real time, and the risk is close to zero.

Second, communication drafting with review. Enquiry responses, vendor updates, chase messages. Real time recovery, and the review is for accuracy about the property rather than for tone.

Third, listing copy with the two-check discipline: no invented feature, no statement about who should live there. This is the largest visible gain and it carries the regulatory weight, so it comes after the habit of checking is established.

Fourth, comparables assembly, with the agent setting the position in the range and presenting it as a range.

And with real care, or not at all without advice: targeted advertising, which is a special category with constrained tooling and a delivery-skew problem you cannot fully control.

The order is deliberate. Most agents start at listing copy because it is the obvious use, which means their first exposure to generated content is in the highest-risk artefact before any checking habit exists.