In his previous article Ed Molyneux discussed how shadow AI is embedded in everyday conveyancing, In it he argues the real issue is not capability but accountability: distinguishing accurate outputs from confident errors, and ensuring data safety always functions. Firms that succeed will channel use, prove reliability, and build guardrails that work when it matters most. In this latest article Molyneux says the recently published reform roadmap will give the market data it can trust, but a model reasoning over perfect facts can still get it wrong… here’s where the real work begins.
The Home Buying and Selling Reform Roadmap published last week has rightly been read as the biggest structural change to our market in a generation. Most coverage went to the sales pack and binding contracts. But the quieter revolution sits in Chapter 10, where the government legislates not for more paper, earlier, but for data that carries its origin with it: gathered once to a known standard, reused with permission, trusted by a lender months later because everyone can see where it came from. The roadmap also actively invites AI-supported conveyancing, an invitation that only makes sense once you understand why, today, you cannot simply ask an AI to do the job.
Here is the uncomfortable thing about large language models, the technology behind every AI assistant your firm is being sold this year. They do not know conveyancing. They are extraordinarily capable pattern-matchers trained on the public internet, producing the most plausible next words for any question. Plausible is not correct. Ask a general-purpose model a precise conveyancing question and it answers with total fluency and confidence whether or not it is right, because confidence is a property of the writing style, not the underlying knowledge.
We see this in our own testing. Put one well-specified Stamp Duty Land Tax scenario to several leading models and they do not merely disagree at the margins; some arrive at the same wrong figure, stated as confidently as the ones that get it right. That error is not noise you train away with a bigger model. It is structural. The model has no access to the actual title, lease or search for the property in front of you. It reasons about a generic transaction that exists only in its training data, dressed in the language of certainty. For a profession whose entire value is being right about this specific property, that is exactly the wrong failure mode.
So the instinct to keep a qualified human in the loop is sound, and regulators are converging on that point. But keeping a human in the loop is a safeguard, not a strategy. It tells you what to guard against, not how to make the AI genuinely useful. If the model is confidently wrong and a tired fee-earner is the only thing between that answer and the certificate of title, you have not solved the problem. You have relocated it.
The real answer is not a cleverer model, and it is not provenance alone. Trusted data is necessary but not sufficient. A model can reason over impeccable facts and still reach the wrong conclusion, because knowing a lease term is not the same as knowing how that term, the ground rent and a lender’s criteria combine to turn a saleable flat into an unmortgageable one. Closing that gap takes encoded judgement: lawyer-verified, risk-grounded playbooks that set out how each risk should be assessed, tested against real anonymised transactions, and required to explain their reasoning rather than assert a result. Give a capable model trusted, provenanced inputs and those playbooks, and the output changes character. It stops being a plausible guess about a generic property and becomes a defensible position on this one, every conclusion tracing back both to the document it rests on and to a rule a conveyancer would recognise.
This is why the reform roadmap, for all its significance, only does half the job. It legislates for the trusted inputs, the provenanced data machine intelligence cannot safely run without. What it cannot legislate for is the judgement applied on top: what a fact actually means for this buyer, this lender, this risk. Provenance tells you a fact is true; it does not tell you what to do about it. That layer has to be built, and audited.
That is the principle we have spent two years building on at Moverly: open, inspectable foundations on the Property Data Trust Framework, not a black box, so an agent’s work is auditable rather than mysterious. Our open-source conveyancing toolkit, a small set of published, lawyer-shaped skills, is the visible proof of the pattern. The next thing we are building applies it across every risk in a transaction: each one a lawyer-verified playbook, evaluated against real anonymised deals, required to show its rationale rather than answer from training data. Not a tool that answers for the conveyancer, but one that does the diligence groundwork against the real facts and hands the professional a defensible, traceable position to sign off. The model never has to “know” conveyancing. It only has to reason correctly over data it can prove, against rules a lawyer has signed. More on that very soon.
Take one lesson into every AI conversation this year. Do not buy the model. Buy the trusted data it stands on and the verified judgement it reasons with. Intelligence you can trust is built on facts you can trace and rules a professional has signed, and nothing else.
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.
