Agentic AI is arriving in conveyancing, and most of what it does is genuinely useful. But the risk calls that carry a conveyancer’s liability are exactly where an unbounded model is most dangerous. Ed Molyneux on what a system that can be trusted with those calls has to look like, and what Moverly has built to meet that bar.
For the past few weeks this column has argued that agentic AI is arriving in conveyancing whether firms are ready or not, and that what separates the firms who benefit is not the model they buy but two things they can trust: the data going in, and the judgement coming out. It is time I was more specific about the second of those, because it is the part we have spent the last year building.
Begin with the risk that ought to keep anyone honest. The most dangerous thing an AI can do in a property file is answer a liability bearing question fluently and incorrectly. Whether a restrictive covenant is enforceable, whether an indemnity policy discharges a defect, whether a lease is mortgageable on a particular lender’s terms: these are the judgements a conveyancer’s professional name and indemnity cover sit behind, and they are exactly where a general-purpose model is least accountable.
It will produce a confident answer because confident answers are what such models produce, and confidence is not the same thing as being right. An agent that gathers documents, chases missing information and drafts enquiries is a genuine help. An agent quietly trusted to make the risk calls on its own is a claim waiting to be made.
A demanding specification
That sets a demanding specification for any system that wants to make those calls, and it is worth stating in the abstract before naming any product. Such a system cannot lean on the model having read enough conveyancing to pattern-match its way to an answer.
It needs an explicit account of how each risk should actually be assessed, written by someone who carries that judgement professionally. It needs to answer every relevant question rather than only the ones it happens to find, including the checks that come back clear and the ones where the data is simply missing, because a risk no one asked about is the one that surfaces on completion.
It needs to reason over information whose provenance is known, so it is not confidently building on a mis-keyed figure or an unverified assertion. And it needs to be honest about the limits of its own confidence rather than smoothing over them.
Where expertise lies
That specification is what we set out to meet with what we now call the DiligenceEngine. It reasons over each transaction against an expert-authored playbook, written with the supervision of expert conveyancers across dozens of risk categories, more than three hundred individual checks and over two thousand worked scenarios, and it returns a graded grid: for every check, a status, a risk score, the scenario it matched, the legal rationale, and the action that follows, addressed to whoever needs to act on it.
The expertise lives in that playbook and in the trusted, provenance-tagged data the engine reads, not in the model itself, which we treat as a bought-in commodity that happens to improve every few months at no cost to us. The point is not that the model is clever. It is that the judgement it applies has been written down, checked, and made accountable.
The harder discipline is not building such a thing but deciding where to let it speak. Credibility is the scarcest asset in this market, and a system like this earns the right to make firmer statements only as it proves itself, one category at a time. So we have begun where the risk is lowest and the need is plainest: auction buyers.
The fall of the hammer
Someone bidding at auction today generally receives no legal review of the legal pack at all, because commissioning a full one, for a property they may not win, costs more than the risk appears to justify at the moment of bidding.
The result is that people commit six figures under a binding contract on the fall of the hammer, having read documents they were never equipped to understand. A tool that reviews that pack and surfaces the material risks in plain language is not competing with a solicitor’s considered opinion. It is replacing nothing, and that is precisely the counterfactual against which a young system should first be judged.
None of this rests on a claim that the engine is always right. The more interesting question, and the one this column will turn to next, is how you would even know: how a system like this should be measured against expert judgement, honestly and in the open, and why that measurement, rather than the model, is the real product.
For now the point is narrower. The judgement calls that carry the liability are the ones agents cannot safely make alone, and building the layer that can is the work in front of us.
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.
