The AI assistant on your website is you

On 7 October 2026 two things happened on the same day, and together they moved the AI assistant from a thing agents talk about to a thing agents have.

Bradleys Estate Agents switched one on and gave it a name. "Brad" runs on Homeflow's AI-powered Live Chat, and Kai Logan, a director of the firm, told Property Industry Eye that in about a month "he's already handled several hundred conversations", that the team have "given him a cheeky streak", and that "he's one of our best hires this year, especially given he doesn't even want a desk". Shawn Zvinis, Head of Product at Homeflow, said the assistant "gives genuinely useful answers straight away and turns anonymous traffic into contacts that agents can follow up".

The same day, The Property Franchise Group told the market it had taken a 25 per cent interest in Enteka Ai: £0.9m for a 15 per cent equity interest, plus a further 10 per cent issued upfront and subject to forfeiture against a three-year agreement making Enteka its preferred AI provider. The platform is already live across more than 210 TPFG offices, covering Belvoir, EweMove, Hunters, Martin & Co, Northwood and Whitegates, and it handles enquiries by telephone, email, SMS, WhatsApp and live chat.

Two hundred and ten offices is a small share of roughly 25,000 UK agency businesses. It is still the first time an automated assistant has been answering consumers in an agency's name at that scale, around the clock. How closely those conversations get reviewed is not something anyone has published.

Which raises a question almost nobody is asking out loud: when it says something wrong, whose statement is it?

What the position actually is here

The United Kingdom has no AI chatbot law for agents. It does not need one, because the law that applies attaches to what you do to a consumer, not to the machinery you did it with.

Since 6 April 2025, Chapter 1 of Part 4 of the Digital Markets, Competition and Consumers Act 2024 has replaced the old Consumer Protection from Unfair Trading Regulations, with one piece deliberately left behind: the consumer's own right of redress, the right to unwind a contract, to a discount or to damages, still sits in Part 4A of the 2008 regulations until replacement regulations are made. Two parts of the new Act matter here.

Section 226 says a commercial practice involves a misleading action if it includes "the provision of false or misleading information relating to a product, a trader or any other matter relevant to a transactional decision", or "an overall presentation which is likely to deceive the average consumer about a matter relating to a product, a trader or any other matter relevant to a transactional decision". Then section 226(3) adds the sentence that matters most to anyone running a chat window: "an overall presentation may be deceiving even if the information it contains is true".

Read that again with a transcript in mind. Every individual sentence your assistant produces can be factually correct and the conversation as a whole can still be a misleading action. That is not a stretched reading. It is what the provision says.

Section 227 covers omissions, and catches information given in a way that is "unclear or untimely", or so that the consumer is unlikely to see it. In deciding whether there has been an omission it weighs both the limitations of the medium used, which a narrow chat bubble certainly is, and any steps the trader took to get the information across some other way. A cramped format is not an excuse on its own.

There is no automation carve-out in any of it.

There is one defence, and it is not the one agents assume. Section 238 covers the act or omission of another person, reliance on information given by another person, a mistake or accident, or another cause beyond the trader's control. It applies to the criminal offences only, not to civil enforcement. And it requires the trader to prove they took all reasonable precautions and exercised all due diligence. An unmonitored assistant with no logging, no accuracy testing and no escalation route fails that second limb before anyone reads the first.

So the accurate sentence is not "a bot said it, so it is the vendor's problem". It is: we governed the thing properly and it still erred, which is a partial defence to a criminal charge, and only if you can show the governance.

Nobody has told agents any of this

Here is the part that should concern you more than the law itself.

The CMA's unfair commercial practices guidance, CMA207, updated on 18 November 2025, runs to 64 pages. The main guidance does not mention chatbots, artificial intelligence, automated systems or algorithms once.

The Information Commission, which succeeded the Information Commissioner's Office on 30 September 2026, inherited general guidance on AI and data protection that applies to a chat widget like it applies to anything else processing personal data. Its predecessor intervened on one consumer chatbot, Snap's. But there is no guidance for businesses putting a chat widget on a website: nothing on how long you keep transcripts, and nothing on the lawful basis for a bot collecting a phone number you will later market to.

The Property Ombudsman has spoken about AI, but from the consumer side: it has said complainants need not declare having used it. We could find no published guidance from TPO on agents' own assistants. Propertymark's most recent published piece on AI and agency, from 10 December 2025, is about consumers using AI to generate complaints against agents, which is the arrow pointing the other way.

The Advertising Standards Authority has spoken, at least twice in 2026, and usefully. Its position is that "if an ad falls within our scope, our rules will apply, regardless of how the ad was created", that advertisers "must always sense-check outputs and remember that they are responsible for the final result", and that responsibility does not shift to the technology provider. But the ASA also says "the CAP Code only applies to ads, not the products advertised", adding that "we cannot comment on the (in)actions of the AI product itself". So the Code reaches what you claim your assistant can do. It does not reach what the assistant then says to a customer in conversation. That falls to the Act and the CMA.

Meanwhile the vendor marketing promises accuracy. Homeflow's own Live Chat page says "every visitor gets the same polished, accurate experience every time". We could find no accuracy disclaimer and no limitation of liability anywhere on it.

A confident promise from the supplier, near-silence from the regulators, and a statutory test that bites on overall presentation. That is not a comfortable place to stand.

What a defensible assistant looks like

None of this is an argument against having one. We built one for this site, because an agency website that can answer a question at eleven at night is better than one that cannot. It is an argument about how it should be built, and the rules are not complicated.

It answers from your pages, not from its own general knowledge. This is the single decision that determines everything else. A model answering from training data will produce something plausible about your fees, your areas and your process. An assistant constrained to your own published pages can only repeat what you have already decided to say, which is the same thing your staff are trained on.

It must be able to say it does not know. An assistant that cannot decline has to invent. Ours says so in its opening line, before anyone has asked it anything, because the visitor should know the limit up front rather than discover it.

It never states property-specific facts it cannot source. Price, tenure, lease length, service charge, EPC band, flood history, parking, restrictions. This is the material information area, and it is currently the most exposed ground in agency compliance. The old National Trading Standards guidance was withdrawn on 8 May 2025, a month after the Act came into force. The CMA has said it has no plans to produce sector-specific guidance of any kind. The government consulted in late 2025 and committed in its June 2026 home buying and selling roadmap to publish non-statutory replacement guidance during 2026, and as at 7 October 2026 it had not appeared. Under section 225(4)(b) of the Act, omitting material information from an "invitation to purchase" is unfair without anyone having to show it changed a consumer's decision, and a priced property listing will normally be one. A bot guessing in that space is the worst exposure on the site.

It never gives legal advice, on Renters' Rights, on tax, on what a tenant can be charged. Those are exactly the questions consumers most want to ask a machine at midnight.

Every conversation should be logged, and somebody should read a sample. You cannot demonstrate all reasonable precautions from memory.

There is a route to a human, and the assistant takes it rather than filling the silence.

People are told it is a bot. Naming it is good marketing and we do it too. A name is not a disguise, and an assistant that lets someone believe they are talking to a negotiator has already failed the overall presentation test.

The sales objective is where it gets interesting

Bradleys' description of Brad is admirably honest: it "always steers the conversation towards a valuation". There is nothing wrong with that. Every page on your site should be doing the same thing.

The care is needed at the join. An assistant with a commercial objective, answering a factual question, in a way shaped by that objective, is the shape of the thing section 226 is aimed at: an overall presentation likely to deceive, assembled from individually true statements. "Is this house in a flood zone?" answered with a cheerful pivot to booking a valuation is not a lie. It may still be an omission that is unclear or untimely.

The test to apply when you read the transcripts is not "did it say anything false". It is "would a reasonably observant person come away from this conversation with an accurate impression". Those are different questions, and the second one is the one the Act is asking.

A last word on timing

There is, as far as we can find, no publicly documented case of a United Kingdom estate agency's AI assistant giving a consumer wrong information and a regulator or ombudsman acting on it.

That is not reassurance. Ombudsman decisions are not published case by case, and a chat error would surface as an ordinary complaint about misleading information with no mention of the software. It means only that this is early. The assistants started going in during 2026. The complaints, if they come, come after.

Which makes now the cheap moment to get the design right, rather than the expensive one.