Listings, Leads, and Comparables: The Actual Workflows

The first real workflow is not writing a listing. It is producing the eight versions of it.

A single instruction typically needs: a portal description constrained by character limits, a brochure version with more length and warmth, two or three social captions in different formats, an email to the applicant list, a window card, and possibly a translation.

Historically this meant writing once and rewriting seven times, badly, at the end of a long day. It is the clearest genuine time saving available to an agent.

The structure that makes it work is a single source of truth. Assemble one factual record of the property: measured dimensions, room count, tenure, age, condition notes, features you have personally verified, and any required disclosures. Then generate every format from that record.

Why that matters beyond convenience. When each format is written separately, they drift, and a discrepancy between the brochure and the portal listing about, say, whether the loft is a bedroom is exactly the kind of inconsistency that becomes a problem later. Generating from one record keeps them consistent by construction.

And it makes the review tractable. You verify the record once, carefully, and then check each generated format only for the two things covered in lesson one: nothing added, nothing about who should live there.

Image tools are now standard in property marketing and they occupy a spectrum from routine to deceptive, with a legal line somewhere in the middle.

Uncontroversial. Correcting exposure, straightening verticals, colour balance. This is ordinary photography and always has been.

Generally accepted with disclosure. Virtual staging, where furniture is added to an empty room. Widely used, and the near-universal expectation and in some jurisdictions the explicit requirement is that staged images are labelled as such.

Contested. Removing clutter, removing a neighbouring building from a view, replacing a grey sky with a blue one. Each of these changes what a buyer believes about the property, and the sky substitution is common enough that many buyers now assume it.

Clearly over the line. Removing or concealing a defect, altering the apparent size or proportions of a room, changing an outlook to hide what it actually faces. These are misrepresentations regardless of the tool used, and the fact that the alteration was generated rather than physical does not change the analysis.

The workable rule. An image may be improved in ways a better photographer with better light could have achieved. It may not be altered in ways that change what the property is. Virtual staging sits outside that rule and is handled by labelling it.

And note the asymmetry: a buyer who discovers at a viewing that the image was altered has learned something about the agent, not just the property.

Inbound enquiry response is where automation has the clearest measurable effect, for a simple structural reason.

A property enquiry has a short window. Someone browsing at ten in the evening is browsing several properties, and the agent who responds first is disproportionately likely to get the viewing. Response speed has been treated as a primary lead-conversion driver in sales practice for decades, and property is a strong case of it because inventory is unique and the buyer is comparing simultaneously.

So the value of automated first response is not that the reply is good. It is that the reply is immediate, which converts a dead overnight window into a live conversation.

What an automated first response should do. Acknowledge, confirm the specific property, answer the two or three factual questions that are answerable from the listing record, and offer viewing times. That is genuinely useful and requires no judgement.

What it must not do. Answer questions about condition, chain position, why the vendor is selling, whether an offer would be accepted, or anything about the neighbourhood. Those are either judgement calls, confidential, or the discrimination risk from lesson one wearing a different hat.

So the design is a narrow bot with an explicit handoff: the automated reply covers acknowledgement and logistics, and anything outside that is flagged for the agent. A bot that tries to hold the whole conversation will eventually make a representation about a property nobody authorised.

The boundary that keeps automated response safe.

An enquiry arrives. The automated layer handles it if and only if it falls into a narrow allowed set: acknowledgement, facts already published in the listing record, availability and viewing slots.

Everything else routes to the agent. Questions about condition, because an answer becomes a representation. Questions about the vendor's position or motivation, because that is confidential and negotiating information. Questions about the neighbourhood and its character, because that is where discrimination risk lives. Anything about price flexibility, because it is a negotiation. And anything the classifier is unsure about, which is the important default.

The design principle is that ambiguity routes to a human rather than being resolved by the bot. A system that answers when uncertain will occasionally answer wrongly about a specific property, and in this profession that is a misrepresentation rather than a bad customer experience.

For agents handling lettings rather than sales, there is a category of AI use that is regulated far more heavily than anything in marketing, and it needs to be recognised as such.

Automated tenant screening, scoring applicants on creditworthiness, rental history, or a composite risk score, is a decision about access to housing. That places it in a category regulated well beyond ordinary business software.

Under the EU AI Act, creditworthiness evaluation of natural persons is listed in Annex III as high-risk, and the Act's Article 27 fundamental rights impact assessment obligation reaches entities providing services in this area. High-risk classification brings requirements on data governance, documentation, human oversight, accuracy and logging that a letting agency will not satisfy by adopting a vendor tool and assuming the vendor handled it. The high-risk obligations under Annex III apply from 2 December 2027 following the Digital Omnibus amendments, which is time to prepare rather than a reason to ignore it.

In the United States, tenant screening reports are consumer reports under the Fair Credit Reporting Act, with adverse action notice and dispute rights attached, and screening algorithms have been the subject of fair housing litigation over disparate impact.

The practical position for an agent. Marketing automation and screening automation are not the same risk category and should not be procured or governed the same way. Screening decisions need a documented human decision-maker, a stated basis, and a route for an applicant to contest. If you cannot explain to a rejected applicant why they were rejected, the tool is not deployable.

The comparables workflow is where an agent's judgement is most defensible, and where the tooling helps most if it is used in the right order.

The order that works. Pull the comparable set first, mechanically and inclusively, from transacted prices rather than asking prices. Asking prices tell you what vendors hoped for, and in a moving market the gap between the two is the whole story.

Then filter by physical similarity: size, type, condition, age, position. This is where an agent overrules the automated match, because the tool sees a three-bedroom terrace at the same square footage and does not see that one backs onto a railway.

Then adjust. Each comparable gets an explicit adjustment for its differences from the subject property, and the adjustments are where your local knowledge is stored. Writing them down is the discipline that makes a valuation an argument rather than an assertion.

Then state a range, with the reasoning for where in the range you sit.

What the model contributes across that. It assembles the candidate set faster and more completely than manual search. It formats the comparison table. It drafts the narrative around your adjustments. It flags a comparable you overlooked.

What it cannot do is the adjustment step, because the adjustments encode information that exists only in your head and your feet: you have been inside those properties.

A vendor presented with a range and explicit adjustments understands the valuation. A vendor presented with a generated number understands only that a computer said it, and will treat it accordingly the moment they see a higher number elsewhere.

The least glamorous workflow is arguably the most valuable, because it addresses the thing clients complain about most.

A property transaction generates a long, fragmented correspondence across email, phone, portal messages and solicitors' letters, spanning weeks to months, with several parties who each hold part of the picture. The common failure is not that something went wrong. It is that something went wrong three weeks ago and nobody noticed because the information was distributed across four inboxes.

What a summarisation workflow does here. Take the correspondence for a transaction and produce a current state: what has been agreed, what is outstanding, who was last asked for what and when, and what has been waiting longest without a response.

That last item is the useful one. Chains stall silently. A weekly view of every transaction sorted by longest-waiting item surfaces the stall before the vendor calls to ask why nothing is happening.

And it feeds the update. Vendors want to be told what is happening, and the reason they often are not is that assembling the answer takes twenty minutes per transaction. Reducing that to two minutes changes how often it happens, which changes the relationship.

This workflow has essentially no regulatory exposure, since it operates on your own records and produces output for you. It is the reason lesson one put the invisible uses first.

Collecting the conditions, since each workflow has a different one and applying the wrong safeguard is a common error.

Listing content across formats. Needs a verified single source of truth and a two-point review on each output: nothing added, nothing about who should live there. The safeguard is on the input record and the output check, not on the model.

Image work. Needs a stated line between improving an image and changing what the property is, plus labelling for virtual staging. The safeguard is a policy, applied consistently, because the judgement is per-image.

Enquiry response. Needs a narrow allowed set and an ambiguity-routes-to-human default. The safeguard is a scope restriction, not a review, because the point of the workflow is that no human is in the loop at the moment of reply.

Tenant screening. Needs a documented human decision-maker, a stated basis, a contest route, and a regulatory assessment appropriate to a high-risk classification. The safeguard is a governance process, and it is categorically heavier than the others.

Comparables. Needs the agent to own the adjustment step and present a range. The safeguard is a division of labour.

Chain summarisation. Needs essentially nothing, because it operates on your own records for your own consumption.

The general shape: the safeguard should sit where the failure would occur. Reviewing a listing does not help an enquiry bot, and scoping an enquiry bot does nothing for a valuation.