Disclosure, Duty, and What an Agent Still Sells

Before any question about disclosure, it helps to be clear about what an agent already owes, because those duties are what determine the answer.

Accuracy about the property. An agent making a statement about a property to a buyer is making a representation, and a false one has consequences ranging from a regulatory complaint to a claim. This duty is old and it is strict in the sense that good faith is not always a defence.

Loyalty to the client. An agent instructed by a vendor acts for that vendor, which shapes what may be shared and with whom. Where an agent acts for both sides, the disclosure obligations tighten considerably.

Confidentiality. What a vendor tells you about their circumstances, their timeline, their bottom line, is not yours to share.

And competence. Holding yourself out as a professional means the client is entitled to rely on your judgement.

None of these change because a model produced a draft. But two of them become easier to breach.

Accuracy, because generated content states things confidently that were never verified. And confidentiality, because vendor information typed into a general-purpose tool has left your control, which is covered later in this lesson.

So the framing for everything that follows: the duties are unchanged, the failure modes are new.

Disclosure obligations divide into three tiers, and conflating them produces either over-disclosure that helps nobody or a missed requirement.

Tier one, required by specific rule. Virtual staging and materially altered images, where a growing number of jurisdictions and professional bodies require labelling. Automated systems interacting directly with a person, where the EU AI Act's Article 50 transparency obligation requires that people be informed they are interacting with an AI system unless it is obvious, applying from 2 August 2026. And synthetic image, audio or video content, which under Article 50(2) must be marked in a machine-readable format, with that marking obligation applying from 2 December 2026.

Tier two, required by the accuracy duty rather than by an AI rule. Anything material about the property that is wrong, regardless of how it got there. This is not an AI disclosure at all. It is the ordinary obligation, and generated content simply creates more opportunities to breach it.

Tier three, not required and usually unhelpful. Announcing that a listing description was drafted with assistance. Nobody expects an agent to have personally typed every word of a brochure, the same way nobody expected it of a copywriter, and the disclosure signals anxiety rather than transparency.

The distinguishing question is whether the person is being deceived about something that matters to their decision. A staged image and a bot they think is a person both meet that test. A drafted paragraph the agent verified and adopted does not.

There is a single idea that resolves most of the anxiety around generated content in a regulated profession, and it is worth stating plainly.

When an agent publishes a description, sends a valuation, or makes a statement to a buyer, the agent is asserting it. Not the tool. The act of publishing is an act of adoption, and after it the provenance is irrelevant to the agent's exposure.

This has two consequences that point in opposite directions, which is why it is useful.

It means there is no safe harbour. A misdescription generated by a model is your misdescription, and no complaints body or court is going to treat the tool as the responsible party. Agents who feel protected because they did not write it are mistaken about how liability works.

And it means there is nothing inherently improper about using the tool. If you have verified the content and you are prepared to stand behind it, you have done exactly what a professional does with any draft from any source: a junior colleague's, a copywriter's, a photographer's.

So the operative question is never was this generated. It is have I verified this and am I willing to assert it.

Which is why the two-point listing check from lesson one is the whole compliance programme for content. It converts an unverified draft into an adopted statement, and that conversion is the professional act.

The quietest risk in this profession is not the listing. It is what gets typed into a chat window.

An agent handles material that is confidential by obligation and often personal by law: a vendor's reason for selling, their financial position, their timeline, buyers' names and contact details, offer histories, and in lettings, applicants' financial and identity documents.

When that is pasted into a general-purpose consumer tool, several things happen at once. It leaves the agency's control. It may be retained. Under some terms it may be used to improve the service. And under data protection law it is a processing operation that needs a lawful basis, a processor relationship, and a place in the agency's records.

The common failure is entirely mundane. An agent pastes a chain of emails into a chat window to get a summary, and that chain contains three parties' names, a solicitor's reference, and a vendor's explanation of their divorce.

The practical positions, in order of preference. Use tools procured by the agency with terms that address retention and training. Strip identifying details before pasting where the task does not need them, which is usually. And keep a clear rule about which categories never go into a general tool at all: identity documents, financial statements, anything about a person's circumstances.

And note this is the one risk in the cursus that is entirely invisible. Nothing goes wrong, nobody complains, and the exposure sits there.

A decision path that separates the three tiers.

Start with the artefact. If it is an image that has been materially altered, including virtual staging, label it. That is a specific requirement in a growing number of places and an expectation nearly everywhere.

If it is an automated system interacting directly with a person, tell them, because Article 50 of the EU AI Act requires it where it would not otherwise be obvious.

If it is synthetic media, mark it in machine-readable form under Article 50(2).

Otherwise, ask the accuracy question rather than the AI question. Is every material statement verified. If not, verify it or remove it. That is the ordinary duty and it applies to all content regardless of origin.

And if the content is verified and adopted, no AI-specific disclosure is required, because the person is not being deceived about anything that affects their decision.

The path is built around a single distinction: deception about the artefact requires disclosure, assistance in producing an accurate artefact does not.

The structural question for the profession is what happens when producing a polished listing costs nothing.

The honest answer is that it removes a differentiator that was already weak. Agents have not, for a long time, won instructions on the quality of their brochure copy. They win on local presence, on the price they achieve, on being recommended, and on being the person a vendor trusts with the largest transaction of their life.

What does change is the floor. When every agency can produce competent listing content, competent listing content stops being a reason to choose one. It becomes table stakes, which is the normal fate of any capability that gets commoditised.

Two second-order effects worth anticipating.

Volume rises. When content is cheap, more of it gets produced, and the marginal social post is less effective than the last. The agencies that benefit are not the ones producing the most.

And buyer trust in listing content declines. As buyers come to assume that descriptions are generated and images are enhanced, they discount both, and weight the viewing and the agent's word more heavily. That shift arguably increases the value of the in-person part of the job.

So the direction is not that the work disappears. It is that the marketing layer gets thinner and the relationship layer gets thicker, which is where the fee was actually justified anyway.

The useful exercise for anyone in a service profession is to name precisely what is not substitutable, because that is what the fee is for.

Having been inside the building. Every automated valuation works from recorded attributes. An agent who has walked through the property knows the things that are not in any dataset, and that knowledge is the entire gap between a range and a price.

Being accountable. A vendor who is badly advised has someone to hold responsible, professionally and legally. A tool provides no such recourse, and for a transaction of this size the existence of an accountable party is a substantial part of what is being purchased.

Negotiating. Reading whether a buyer will stretch, knowing when to hold and when to take the offer on the table, managing a chain where four parties each have partial information and competing interests. This is judgement under uncertainty with money at stake and no clean information.

Holding the relationship. Absorbing a vendor's anxiety at week nine, telling them something they do not want to hear about their price, and being the person they call in four years and recommend to their sister.

And local knowledge that is not written down anywhere. Which developer builds well, which block has a management problem, which street floods, what the council is likely to approve.

The pattern across this catalogue's profession cursus holds here: the tools compress production and leave judgement, accountability and relationship intact. In real estate that residue is unusually large, because the transaction is infrequent, high-value, emotionally loaded, and legally consequential.

What a small agency can write down in a page, which is the realistic form of governance at this scale.

Content. Listing descriptions are generated from a verified property record and checked on two points before publication: nothing added that is not in the record, and nothing describing who the property suits. Named person signs off.

Images. Enhancement is permitted where it corrects the photograph. Alteration that changes what the property is, is not. Virtual staging is labelled in every channel.

Enquiries. Automated response covers acknowledgement, published facts and viewing slots only. Anything about condition, vendor position, neighbourhood or price routes to a person. The system identifies itself as automated.

Valuation. Automated output is an input to a range, never a figure presented to a client on its own. Adjustments are written down by the agent.

Screening, where the agency does lettings. Not automated to a decision. A named person decides, the basis is recorded, and applicants can contest. Treated as a separate and heavier regulatory category.

Confidentiality. A named list of tools that are approved. Client personal data, identity documents and financial information do not go into anything outside it.

And one line that carries most of the weight: whatever we publish, we are asserting. Provenance changes nothing about that.

Six paragraphs is proportionate for an agency of ten people, and it is six paragraphs more than most have.