Ed Molyneux argued last week that the next 18 months in conveyancing will be defined less by what AI can do, and more by which firms have the management muscle to adopt it. This week he turns to the layer below that: how a firm adopts AI without the adoption itself becoming the liability.
It’s six o’clock on a Friday. A junior has a draft enquiry letter she promised would be on her supervisor’s desk by close of business, and she is one paragraph short. She pastes a page of the client’s bank statement into a consumer AI chatbot, asks it to summarise the source-of-funds story, and copies the result into the email. The letter goes out, she closes the laptop, and nobody noticed. The work got done. The client never asked.
This is playing out in every conveyancing firm in the country, often several times a week. It is not malice and it is not laziness; it is competent people getting work done with the tools easiest to reach. The question for the firms reading this column is not whether shadow AI is present — it is — but what to do about it.
A lot of the AI conversation in property has been about whether the models are good enough. That is the wrong question. The honest answer is “mostly yes, and getting better.” The harder question, and the one that matters for a regulated profession, is whether the firm can tell a right answer apart from a confidently wrong one.
A concrete example. Earlier this year we put a £510,000 first time buyer SDLT calculation to several frontier AI models. Several returned around £4,250, calculated against relief thresholds that applied before April 2025. The correct answer under current rules is £15,500: the price exceeds the new £500,000 cap on first-time-buyer relief, so the relief does not apply and standard rates kick in. None flagged uncertainty. None said, “I’m not sure this is current.” They were confidently, plausibly wrong by more than £11,000.
The danger in property AI is not that it is dumb. It is that it is plausibly correct. A fee earner reading a hallucinated answer in the right ballpark and the right register will sign it off four times out of five. The fifth time is the one that ends up in front of the SRA.
Two kinds of AI, routinely conflated
There are two very different kinds of AI tool entering conveyancing, sold under the same label. The first makes the AI smarter about the work — feeding it the current SDLT thresholds, live lender Part 1 requirements, the right lease-eligibility rules — so it gives a better answer. The second makes the AI safer in use — guardrails that watch what is being typed, flag when client-sensitive data is about to leave the firm’s environment, and refuse to allow it. Different jobs, different evidence required, both important. The firms that treat them as one thing are making a category mistake that will cost them.
The silent-failure trap
The trap with safety guardrails is silent failure. A check that runs reliably catches what it was meant to catch. A check that runs only sometimes is worse than none, because the firm now believes it is covered, and the half of the incidents the check missed sail straight through. The guardrail has to live in the environment fee earners actually work in, not bolted onto good intentions at the end of a busy week. You cannot reasonably ask the individual to turn on the tool meant to catch their own slips.
There is a related nuance worth airing. A good data-safety check has to be capable of judgement, because not all client data is equal. Owner names on the title register are public information; a tool that throws a red flag every time one is mentioned will be turned off by week two. A client’s bank statement, an unredacted ID, a will draft — different category, different response. “Not all AI is the same” cuts both ways: the tools that make AI safer in conveyancing have to know what is genuinely sensitive and what only looks sensitive on first glance.
What firms should actually do
First, do not ban AI. Channel it. Banning never works — the fee earner with the deadline will use it anyway — and the only question is whether the safe, sanctioned route is also the easiest one. If it is not, the unsafe route wins.
Second, ask suppliers a sharper question. Not “do you use AI?” — every supplier will now say yes — but “can you show me that the answers are current, and can you guarantee the safety checks always run?” Treat any AI claim the same way you would treat any other control on a file: unmeasured, unproven, unsourced — discount it.
Third, frame data-safety prompts as coaching, not policing. A fee earner who sees “this looks sensitive — would you like to use the firm’s approved tool?” learns the right reflex. One who hits a punitive block learns to route around it.
No system is airtight. Someone can always use a personal phone or email. The aim is not to pretend leakage is impossible; it is to make the sanctioned route the path of least resistance, so the fee earner choosing it on a Friday afternoon is choosing the easier option.
The firms that lead the next decade of conveyancing will not be the fastest to adopt AI. They will be the ones that adopt it accountably — with evidence behind the smartness, and enforcement behind the safety. The most important AI tool a firm runs may turn out to be the one that, at six o’clock on a busy Friday, knows when to say “stop.”
About the author
Ed Molyneux is co-founder and CTO of Moverly, the property intelligence platform working with LMS and Connells Group to bring structured, verified data to property transactions. He is the architect of the Property Data Trust Framework (PDTF), the open standard for machine-readable property data now being adopted across the industry. Ed writes about AI, property data infrastructure, and the future of conveyancing.
