AI property description generator: build, buy, or hire?

Most estate agency software now ships with a button that writes a property description in seconds. When Alto launched its version in June 2024, it reported that 80% of its customers had seen and used the feature, and some listings on Zoopla were already carrying text it had produced. The question for an agency is no longer whether AI can write listing copy. It is whether a generic tool writes copy that is accurate, compliant and recognisably yours.

That is where the choice between building, buying or hiring comes in. A free feature inside your CRM is a sensible starting point for many agents. A custom agent that knows your house style, your portal rules and your compliance checks is a different product, and it suits a different kind of agency. This guide compares the three routes so you can decide which one fits your volumes, your risk appetite and your team.

Quick answer

Use the generator built into your agency software if you list a modest number of properties and an experienced negotiator edits every draft. Consider a custom agent when you need consistent house style, automatic checks against portal fields and advertising rules, or copy across several branches. In every case, a person must approve the final text.

What an AI property description generator actually does

A property description generator takes structured facts, such as bedrooms, property type, location features and condition, and turns them into marketing prose. Some tools also read floorplans or photos, and some rewrite a description for a different portal, social post or print brochure.

Agency software vendors have built this into their products. Street's AI tool writes descriptions from details agents enter, lets them refine a draft with prompts, edit it by hand or regenerate it, and is included for all Street.co.uk customers. Alto's equivalent is free for all Alto users, according to the trade press. These are useful because they sit where the property record already lives, so there is no copying and pasting between systems.

The limit of any generator is the same: it writes from the facts it is given and from patterns in its training. If the record says the garden is south-facing and the vendor's surveyor says otherwise, the tool will not know. If the record is thin, the model fills gaps with plausible-sounding detail. That is the behaviour that creates compliance risk, and it is the reason the rest of this guide spends so much time on checks rather than on prose quality.

The compliance rules every generated description must meet

Whoever writes the copy, the agency publishes it and owns it. Two sets of rules matter most for UK agents: consumer protection law and the advertising codes. For agents elsewhere, the principle is the same even if the statute differs.

Consumer protection and material information

Since 6 April 2025, the Digital Markets, Competition and Consumers Act has replaced the old Consumer Protection from Unfair Trading Regulations. Government guidance confirms that the CMA and courts can fine non-compliant traders up to 10% of their worldwide turnover, and that giving required information too late or in a way consumers are unlikely to notice can be a misleading omission.

For property, that sits on top of the material information work that began with National Trading Standards. In July 2022 its estate and letting agency team published guidance on Part A material information, covering price, council tax band and tenure, and developed with the major portals. Propertymark, the professional body for agents, has since noted that this guidance has been withdrawn from official channels and that the burden is now on agents to judge what the average consumer needs to know. It adds that omitting material information is now automatically an unfair commercial practice.

In practice, a generator should never be allowed to be the only author of facts that matter to a purchase. Tenure, lease length, service charge, ground rent, flood risk, construction type, parking and restrictions should come from the verified property record, not from model output. The description can mention them, but a check should confirm that the text matches the record.

Advertising codes and exaggeration

The advertising regulator's guidance for the sector makes the same point from another angle. The ASA notes that property ads should not exaggerate features and that objective claims need adequate substantiation. In one upheld case, a claim of a private drive was found to break the rules because the drive was shared with other residents.

That example is a good test for generated copy. Phrases such as private drive, south-facing garden, walking distance to the station, newly refurbished or no onward chain are all objective claims. A model may produce each of them from a vague prompt. A human who knows the property would catch the shared drive. An unsupervised tool will not.

Fair housing and discriminatory wording

In the United States, the equivalent risk is fair housing law. A HousingWire checklist warns that agents usually get into trouble through subtle steering, because AI tools learn from past listing copy that can contain outdated wording or assumptions about buyers. It recommends avoiding buyer-profile phrases such as ideal for families, and describing verifiable features rather than the ideal occupant. Its closing point applies everywhere: once your marketing goes public, you own it.

Regulators are watching how AI touches housing advertising. In May 2024 HUD released guidance on how the Fair Housing Act applies to advertising through online platforms that use targeted ads, issued as part of its work on artificial intelligence. Whatever your market, a good generator should be configured to avoid describing who a home is for, and to describe what the home is.

Option one: use the generator in your agency software

For many small and mid-sized agencies, the built-in tool is the right first step. It costs little or nothing extra, it reads the property record, and negotiators can adopt it in an afternoon. Our own comparison of the best AI tools for estate agents covers where these features sit in a wider stack.

The strengths are convenience and adoption. Alto's figures suggest negotiators will try a tool that is one click away. The weaknesses show up as you scale. Tone of voice is usually generic, though Alto has said it plans controls over tone. You cannot normally add your own compliance checks. You cannot usually see why a sentence was written. And because the tool belongs to the vendor, your copy standards depend on their roadmap.

This route works when three conditions hold. Volumes are modest. One experienced person edits every description. Your CRM record is complete and accurate, so the model has real facts to work from. If any of those is missing, the tool will still produce fluent text, which is exactly the problem: fluent text hides gaps.

Option two: buy a specialist or general-purpose tool

The second route is a standalone product, either a listing-focused service or a general AI assistant used with a prompt template. This gives you more control over tone and format than a CRM feature, and it can work across systems.

The trade-off is process. A general assistant knows nothing about your property data unless someone pastes it in, which creates copy-and-paste risk and sends client details to a third-party service. Before using any such tool, decide what personal data, such as vendor names or addresses, may be entered. The ICO has said that the more people rely on a model as a source of information rather than inspiration, the more central accuracy should be to its design and testing. A listing description is information, not inspiration, so treat accuracy as a requirement.

It is also worth reading the ICO's point about deployers. It suggests they consider how inaccurate outputs could cause harm, put mitigations in place such as output filters, and monitor how a tool is used. For an agency, that translates into a written process: who may use the tool, what they may enter, and who signs off the text. Our guide to GDPR and custom AI agents goes deeper on that side.

Option three: build or hire a custom listing agent

A custom agent is not a better writer than a good off-the-shelf model. In most cases, it uses the same kind of language model underneath. The difference is everything around the model.

A well-designed listing agent pulls facts directly from your CRM or property management system, so it never has to guess. It writes in your house style, using examples of descriptions your best negotiators have approved. It checks the draft against the record and flags any claim that has no supporting field, such as a private drive that is not recorded as private. It confirms that required portal fields are filled in. It produces variants for portal, brochure and social posts from one approved source. And it logs what it produced and who approved it.

This is the same pattern we describe in our comparison of custom AI agents and off-the-shelf tools: the generic product is quicker to start, and the custom one pays off where your process is specific and your risk is real. It also depends on connecting cleanly to your existing systems.

A custom build tends to make sense when you have several branches that must sound like one brand, when you list at volume, when you handle lettings and sales with different rules, or when a compliance officer wants an audit trail. It rarely makes sense for a single-office agency with a handful of new instructions each week.

Build, buy or hire: a decision framework

Use these questions in order. The first honest no usually tells you where to stop.

Hiring a partner rather than building in-house is mostly a question of capacity. Writing the agent is the smaller part of the work. The larger part is mapping your fields, writing the checks, testing against real listings and training staff.

What good checks look like

Whichever route you choose, the quality of the checks matters more than the quality of the first draft. A description that reads beautifully but contains one unsupported claim is worse than a plain one that is correct.

The first check is source grounding. Every objective claim in the text should trace to a field in the record or a note from the negotiator. If it does not, the draft is held for review rather than published.

The second is a banned and risky phrase list. This should cover buyer-profile wording, unsupported superlatives, and claims you cannot substantiate, such as best in the area. It should also catch words that imply certainty about things like planning potential, which a negotiator cannot promise.

The third is a material information check. Before publication, confirm that price, tenure, council tax band and the other items your jurisdiction expects are present in the record and visible on the portal listing. The National Trading Standards article noted that fields left empty are flagged on the listing, so a missing item is visible to buyers as well as to regulators.

The fourth is human sign-off, recorded. The person who approves the text should be named in the log. This is a cultural point as much as a technical one: it keeps responsibility where the law puts it.

The fifth is feedback. When a negotiator rewrites a sentence, capture what changed. Over a few weeks you learn which kinds of claim the model gets wrong, and you can tighten the rules or the prompts.

How Klevere approaches AI property description work

Klevere has deployed more than 500 AI agents across over 50 projects and 12 industries, and we start every listing project by looking at the data and the process rather than the prose. Our first question is usually where facts about a property are held, how reliable they are, and who currently signs off marketing copy.

From there we design the agent around your systems and your standards. That means connecting to your CRM or property platform, writing checks that match your compliance policy, and building a review step that fits how your negotiators already work. Many agencies find they do not need a large custom build. Sometimes the right answer is to keep the CRM tool and add a thin checking layer, or simply to tighten the process around what they already have.

If you want to see what a build would involve, our page for AI for estate agents covers the wider set of workflows we automate in the sector, from lead handling to valuation follow-up.

Common mistakes to avoid

The first mistake is treating the model as the source of truth. A description generator should format and phrase facts it has been given. It should not discover them.

The second is skipping the template work. A generic prompt produces generic copy. A few well-chosen examples of descriptions your team is proud of will shape style far more than any instruction to be warm and professional.

The third is ignoring lettings. Rules and material information differ between sales and lettings, and a tool tuned for one can mislead on the other.

The fourth is forgetting data protection. Descriptions often draw on information about vendors, tenants and neighbours. Keep personal data out of prompts unless you have a clear lawful basis and an agreement with the provider.

The fifth is measuring the wrong thing. Time saved per listing is easy to measure, but the numbers that matter are corrections made before publication, complaints, withdrawn listings and enquiry quality. If copy gets faster and worse, you have not made progress.

Frequently asked questions

Can AI write property descriptions that comply with UK advertising rules?

It can draft them, but compliance depends on the checks around it. Objective claims need substantiation, and material information must be present and accurate. Use verified property data as the source, run automated checks for unsupported claims, and have a named person approve every description before publication.

Is an AI property description generator free?

Some are. Alto and Street both include AI description tools for their existing customers, according to trade press coverage. Standalone and custom tools differ, so check each vendor's own pricing page. The larger cost is usually staff time spent checking and editing output, rather than the licence.

Who is responsible if an AI-written listing is wrong?

The agency that publishes it. Regulators and portals treat the listing as your advertisement, whoever or whatever drafted it. That is why a recorded human sign-off matters, and why generated text should be checked against the property record before it goes live.

Should I build a custom listing agent or use my CRM's tool?

Start with the CRM tool if volumes are modest and someone edits every draft. Build a custom agent when you need a consistent voice across branches, automated compliance checks, portal-specific variants or an audit trail. Fix your property data first, because neither route works well on poor records.

Can AI listing tools discriminate against buyers or tenants?

They can if the wording steers readers by protected characteristics, for example by describing who a home is ideal for. Describe verifiable features instead. Review output for buyer-profile phrases and coded terms, and keep a banned phrase list that the tool checks automatically before text is published.

Can I put vendor details into a general AI assistant?

Only with care. Decide in writing what personal data may be entered, confirm how the provider handles it, and prefer tools connected to your systems over copy and paste. Data protection duties apply to the information in prompts as much as to the text that comes out.

If you are weighing a built-in tool against a custom listing agent, the quickest way to find out which fits your agency is a free AI audit. We review your listing process, your data and your compliance checks, and tell you plainly whether you need a build at all.