AI Consulting for Legal Firms

Real enough that courts in multiple jurisdictions, including Australia, have had to deal with it, and real enough that several Australian courts have issued practice notes or guidelines governing generative AI use in material filed with them. The mechanism is worth understanding because it explains why the problem is stubborn: a language model generates text that is statistically plausible, and a fabricated citation is extremely plausible text. It has a real-looking party name, a real-looking year, a real-looking court. Nothing in the model knows the difference between a case that exists and a case that would have been decided if the universe were slightly different. The mitigation is not a better prompt — it is architectural. Retrieval against an actual legal database, citation verification as a mandatory step, and a hard rule that no authority reaches a document without a human confirming it exists and says what the draft claims. Any vendor who tells you their model does not hallucinate is either misinformed or selling. Also worth noting: practice notes on this change, and they vary by court, so the current requirements in your jurisdiction are something to check rather than assume.

This is the question that should be settled before a single matter file goes near a model, and it is a question for your firm and your professional indemnity insurer rather than a technology consultant. What we can do is establish the facts a lawyer needs in order to answer it: where the data physically goes; whether the vendor retains prompts or outputs; whether inputs are used to train or improve the vendor’s models; who at the vendor can access the content and under what circumstances; what jurisdiction the vendor and its subprocessors sit in and what compelled-disclosure exposure that creates; and what the contract says versus what the sales team said. The Solicitors’ Conduct Rules on confidentiality do not contain an exception for convenient software. Our repeated finding is that firms make this decision on the marketing page rather than the data processing agreement, and the two are frequently different documents.

Overwhelmingly in the work around the law rather than the law itself. First-pass document review and discovery triage, where volume is enormous and a human still confirms everything consequential. Matter intake and conflict-check preparation. Correspondence triage — sorting and summarising the inbound flood before anyone opens it. Time recording, which lawyers hate and do badly, and where reconstructing the day from calendar, email and document activity recovers billable hours that were genuinely worked and never captured. Precedent retrieval — finding the clause your firm already drafted well three years ago, instead of drafting it again from scratch. Client status updates. None of it is glamorous, all of it is high volume, and it is where the hours and the margin actually leak. The firms that get value are the ones who accept that AI in legal practice looks like a very good paralegal for a narrow set of tasks, not a junior solicitor.

It matters more than the technology choice. Law firm partnerships have a specific structural challenge with technology: the people with the authority to approve investment are often the people furthest from the work being automated, and the benefits accrue unevenly across practice groups. We have seen good projects die because litigation wanted it and property did not, and the compromise was to do a diluted version of both. Our practical advice is to scope the first project to a single practice group with a single willing partner, make the measurement honest, and let the result do the arguing. A written audit helps here in a way that a demo does not, because it gives sceptical partners something to disagree with in specific terms rather than in principle. Partnership politics is a legitimate input to the sequencing decision, and any consultant who pretends otherwise has not worked with a firm.

It depends on which problem you are solving, and the answer is often both. Legal-specific platforms are worth their premium where the value is in the legal content and integrations — a research tool with a properly maintained, verified case database is doing something you cannot replicate, and should not try to. General-purpose tooling wired into your own systems tends to win for the firm-specific admin: your intake, your precedents, your document management system, your time recording, your correspondence flow. That work is not legally special; it is just yours. Where firms go wrong is buying a legal AI suite to solve an admin problem, paying legal-software prices for workflow automation, then discovering it does not connect to the practice management system they actually run. The audit exists to sort those two categories before the purchase order, not after.

It starts with a free initial consultation: a conversation about where the hours are going and what the partners actually want to change. If there is a real opportunity, the next step is generally the AI Opportunity Audit at around $3,000 — a written map of your workflows, the systems and data behind them, the confidentiality and professional constraints attached to each, and a ranked shortlist with realistic effort and payback. It is deliberately useful on its own; you keep it whether or not you engage us further. Where it makes sense, we then build and ship the first workflow into production rather than handing over recommendations, and we stay involved while it beds in. The audit generally earns its fee at around 20 or more staff. A five-person firm usually needs one or two well-chosen tools and no consultant, and we will say so.