AI in regulated professional work, Mid-2026

A family lawyer’s affidavit, a valuer’s report, and a building inspector’s pre-purchase report have the same structure. A professional attends to the facts, applies judgement, and signs a document that someone else will rely on: a court, a lender, a purchaser at the moment of maximum commitment. The name at the bottom is the unit of accountability. The professional indemnity attaches to that name and to no other.

That structure is why these three professions are worth reading together in mid-2026. Each of their regulators has now been asked the same question (what happens when AI touches a document a professional has signed?) and each has answered differently. The courts wrote rules and enforced them. RICS wrote principles and made them mandatory. Building inspection in Australia has, so far, no answer at all. The distance between those three answers, when the underlying technology is identical in all three cases, tells an operator most of what they need to know.

This note consolidates and updates the three separate industry notes we published in June 2026. The evidence base for every claim below is the State of AI hub and the full Mid-2026 literature review, which surveys ten capability categories across applied AI and went through three independent fact-check passes before publication.

What do family law, valuation, and building inspection have in common?

In each of them, a professional’s name attaches to a document that someone with no obligation of charity will rely on. A family law affidavit is read by opposing solicitors who are paid to find errors, and by judicial officers under time pressure. A valuation is read by a lender, an owner, or an opposing expert who will test the comparable evidence. An inspection report carries a purchasing decision and is read, when things go wrong, by a tribunal. In all three, the reader’s reliance runs to the professional’s name and credentials, not to any tool that contributed to the document’s production.

The liability profile has the same shape too, and it is uneven in the same way. Inspection liability concentrates around the defects that were present, within scope, and missed, not the concealed ones a competent inspector could not have seen. A legal citation that does not exist is a categorical failure, not a marginal one. A valuation figure on contested property gets tested line by line by an opposing expert. In each field, a tool that is excellent at the easy cases and weak at the hard ones inverts the risk profile the professional actually carries.

That is the frame in which every AI question in these professions has to sit. Not “is the tool impressive” but “what reaches the signed document, and who verified it on the way”.

How has each regulator answered the AI question?

Three different ways: the courts wrote rules and have enforced them, RICS wrote principles and made them mandatory, and building inspection in Australia has no answer yet.

The courts: rules, with enforcement. On 16 April 2026 the Federal Court issued practice note GPN-AI, requiring the practitioner responsible for a document to confirm that legal authorities cited in it exist and support the propositions stated, alongside the facts and evidence relied on. NSW had already prohibited generative AI for affidavits and expert reports without leave from February 2025. Victoria, Queensland, and South Australia have parallel notes and guidelines in force. The enforcement is not hypothetical. In late 2025, in Mertz & Mertz (No 3) [2025] FedCFamC1A 222, the Federal Circuit and Family Court ordered a solicitor to pay $10,000 in costs thrown away after AI-fabricated citations were filed in a list of authorities, and referred the solicitor, along with two counsel, to their state conduct regulators. The question “is AI use permitted in legal work” has moved from open to answered: permitted, with a verification obligation that attaches personally to the practitioner.

Valuation: principles, made mandatory. RICS made its responsible-AI standard mandatory for regulated surveyors on 9 March 2026. The standard does not single out automated valuation models or defect detection by name. It requires risk-based governance, client disclosure of material AI use, and clear human accountability for the outcome, for any AI application a regulated surveyor relies on. In Australia, the API and the state property regulators have not yet matched RICS with a binding professional standard, but the direction of travel is visible, and the sensible planning assumption is that the Australian floor rises to meet the RICS one. A valuation-specific reading of the evidence, including the strongest Australian study and the build rules that follow from production work, is in AI in property valuation: the evidence and the design rules.

Building inspection: silence, so far. No Australian regulator has issued a binding position on AI in inspection work. The only legislated pathway anywhere is in the United States, where Florida permitted software-based plan review via House Bill 683, effective 1 July 2025. No independent accuracy evaluation of the plan-review products in that market (CodeComply.AI, CivCheck, PlanCheckPro.AI) has been published, and the regulatory pathway in Australia has not opened. Silence is not permission. It means the standard an inspector will eventually be measured against is being set now, elsewhere, and the legal profession’s lesson is available to be learned privately rather than publicly.

What does the independent evidence say about the tools?

In all three fields, vendor accuracy claims run ahead of the independent evidence, and the failure modes concentrate exactly where the liability does.

Legal AI has the most independent measurement, because the products are the most mature. Harvey raised US$200 million at an US$11 billion valuation in March 2026 and Thomson Reuters announced one million professional users of CoCounsel in February 2026; these are serious commercial products in wide use at sophisticated firms. Against that, the preregistered Stanford RegLab study, published in the Journal of Empirical Legal Studies in 2025, found Lexis+ AI hallucinating in 17 per cent of responses and Westlaw’s AI research product in 33 per cent, both marketed at the time with hallucination-free language. Those figures come from a peer-reviewed study by an institution with no commercial interest in the result, which makes them the credible baseline for any purchasing decision.

Valuation has the most honest vendor figure. CoreLogic’s own account, given by a company spokesperson to trade media rather than in a formal published report, puts almost 90 per cent of its Australian AVM estimates within 15 per cent of sale price since early 2024. That is a vendor claim, not an independently audited figure, but it is the most credible accuracy anchor in the Australian market. A widely circulated figure of “94.2 per cent within 5 per cent” could not be traced to any credible source, vendor or independent, and should be treated as marketing. Read plainly, the honest figure supports automated valuation as a triage and sanity-check layer, and does not support it as the final word on any contested property.

Building inspection has the most dangerous distribution. Vendor case studies for defect detection from imagery converge on roughly 90 to 95 per cent accuracy for visually distinct defect classes such as water staining and roof membrane damage; no independent benchmark confirms that range at the precision vendors quote it. (What video-capable models can and cannot actually see is surveyed separately in AI and video, Mid-2026.) The same systems perform poorly on subtle, internal, or occluded defects, and those are precisely the defects that produce inspection liability. The technology is strongest at the defects an inspector would have found easily and weakest at the defects an inspector might have missed. The exposure profile is the opposite of what the marketing implies.

Two cross-cutting findings complete the picture. First, the best independent evidence on grounded summarisation, the capability underneath every “ask your firm’s documents a question” tool, still shows leading systems hallucinating in roughly 13 per cent of responses, with omission rather than fabrication as the dominant failure mode. An omitted fact in family violence material, or an omitted qualification in a post-inspection answer to a client, is not a recoverable error. Second, any vendor whose published evidence rests on a single secondary source or an unspecified internal benchmark is showing you a procurement red flag, not a capability.

Where does AI fit when a signature carries the liability?

Upstream of the signature: assembling inputs, drafting, and retrieval, with verification built into the routine and the professional’s judgement authoritative on everything that reaches the document.

The defensible pattern is the same in all three professions. Research and retrieval, with the professional reading the underlying source before relying on it: the solicitor reads the authority before citing it, the valuer reads the comparable evidence, the inspector’s eye stays authoritative on what the photographs mean. First-cut drafting from structured input, with the professional editing rather than approving wholesale: routine legal documents from instructions and precedent, valuation reports from a structured field record, inspection reports assembled from the inspector’s on-site narration. The closest validated analogue for that last capability is clinical AI scribes, where a multi-site study across five academic health systems found around 16 minutes saved per 8 hours of patient care among adopters, with an error profile dominated by omission; both findings translate directly. And client communication templated against the firm’s own prior reasoning, so routine questions are answered consistently rather than from memory or from scratch each time.

The most underrated shift is also common to all three: knowledge synthesis at the firm level. Current frontier models can hold an organisation’s accumulated material in working memory at a scale they could not eighteen months ago. A boutique family law firm with twenty years of file notes, submissions, and consent orders; a valuation practice with two decades of reports and expert witness statements; an inspection firm with years of completed reports and twelve months of post-inspection client questions on every job: each is sitting on a corpus no off-the-shelf product can replicate, because no off-the-shelf product has access to the firm’s own work. The value is not that a tool replaces senior judgement. It is that the firm’s accumulated judgement becomes legible to every professional in the firm, on demand, subject to the same verification discipline as every other intermediate input.

What should a principal do while the floors keep rising?

Build the verification routine now, before it is mandated, and make it something you could describe to a judge, a regulator, or a tribunal without embarrassment.

For legal practitioners this is no longer optional. GPN-AI has made verification a professional obligation, and the costs order and conduct referrals in Mertz showed the courts are prepared to act on it. For valuers, the RICS standard is the visible shape of what is coming, and a practice that adopts risk-based governance and client disclosure before the Australian floor rises will find the transition uneventful. For inspectors, the absence of a rule is the opportunity: a firm that can describe in plain language what AI does and does not do in its inspection process is writing the standard it will eventually be measured against, on its own terms.

The operational moves are the same everywhere. Be specific, internally, about which tools are in use, on what kinds of matters or jobs, with what verification routine attached, and be willing to make that visible to clients. Build the verification step into the workflow, not into individual discretion. Treat the firm’s own material as the asset most worth making legible before licensing anything. And on timing: the frontier is moving quickly enough that nobody needs to be first, and slowly enough that eighteen months without forming a view leaves a practice visibly behind a peer who spent those months running small structured pilots and learning what to trust.

The question worth working through is the same in all three professions. Not “which AI product should we license” but “what is the verification routine around any AI-assisted output, what does our accumulated experience look like once every professional in the firm can ask it a question, and how would we describe both to the court, the regulator, or the tribunal that eventually asks”. We work through exactly that with professional services practices in Perth. If those sound like the right questions for your practice, start with a conversation.

Questions practices ask

Can Australian lawyers use AI to draft court documents? Yes, with obligations attached. Federal Court practice note GPN-AI requires the responsible practitioner to confirm that cited authorities exist and support the propositions stated, and NSW prohibits generative AI for affidavits and expert reports without leave of the court. Verified AI-assisted drafting is permitted; unverified output filed with a court has already produced costs orders and conduct referrals.

Are automated property valuations accurate enough to replace a valuer? No. The most credible Australian figure is CoreLogic’s own, which puts almost 90 per cent of its AVM estimates within 15 per cent of sale price, and it is a vendor claim rather than an independent benchmark. That supports using AVMs for triage and sanity checks, not as the final word on contested property, where a valuer weighs the model output as one input among several.

Can AI reliably detect building defects from photographs? Only the easy ones. Vendor case studies converge on roughly 90 to 95 per cent accuracy for visually distinct defects such as water staining, with no independent benchmark confirming that precision, and performance drops on subtle, internal, or occluded defects. Those are the defects that produce inspection liability, so the systems work as a second pass on photographs already gathered, not as a substitute for a trained inspector.

Do professionals have to tell clients when AI was used? It depends on the regulator. RICS requires regulated surveyors to disclose material AI use to clients under its standard, mandatory since 9 March 2026. Australian courts require practitioners to verify AI-assisted material rather than disclose it in every instance. Where no rule exists yet, being able to describe what AI did and did not do in your process is the posture that ages well.