// AI
Document review, first drafts and the inbox are where AI pays back in a NZ law firm. Legal research is where it burns partners. The uses, the risks, and a 90-day plan that respects privilege and the Law Society’s guidance.
The short version
Law firms were early to try AI and early to get burned by it — usually by a confident citation that did not exist. The firms using it well in 2026 have narrowed it to three things it is genuinely good at and fenced off the one it is not. Here is that split.
| Use | What it does | What it needs |
|---|---|---|
| First drafts from precedents | Drafts letters, clauses and standard documents from your own precedent bank and the matter file | Precedents in SharePoint; matter permissions correct |
| Document and bundle review | Summarises discovery, contracts and long correspondence; finds the clause or the date | Documents in the tenant; Copilot or an approved legal AI |
| Inbox and meeting management | Drafts replies, summarises threads, produces attendance notes from Teams | Teams transcription, with consent, where appropriate |
| Time-recording assistance | Turns the day’s documents and calls into draft time entries | Practice-management integration |
Cost. Copilot or an approved legal AI for the fee-earners who draft and review most; the matter-permission audit, which most firms need regardless; training on verification. A five-lawyer firm is looking at a low five-figure annual commitment after the pilot, plus professional advice on the engagement-terms clause.
Return. First drafts of standard documents in minutes instead of an hour; long bundles summarised before the meeting rather than during it; attendance notes written by the end of the call. Firms we work with recover the licence cost in recorded-but-unbillable time within the first two months.
Risk if skipped. A junior uses a free tool on a client matter, and the firm finds out from the client.
Before any of this: a one-page AI policy saying which tools are approved, what may never be pasted into them (client identifiers, health, financial and legal detail unless the tool is inside your Microsoft tenant), and who owns it. We cover the page itself in a separate article, and Be AI aligns the work to ISO/IEC 42001 and the NIST AI RMF so the answer to “is this governed?” is yes.
Every use above runs inside DigitalShield, on the Microsoft 365 you already pay for, with the permissions and backups done first. If you want the pilot scoped for your practice, see the industry plan or talk to a Be AI consultant — the readiness check is free.
Only inside tools bound to your own tenant with enterprise terms, where prompts and documents are not used for training and access follows your matter permissions. Consumer tools are not safe for privileged material and most firms now prohibit them in policy.
Not reliably as a source. General models are weak on New Zealand case law and legislation and will invent citations. Use AI to summarise and draft from authorities you supply from a proper legal database, and verify every citation.
The Law Societyu2019s guidance expects practitioners to remain responsible for the work and to consider disclosure where AI materially contributes. Most firms now add a short clause to terms of engagement. Take your own professional advice on wording.
This is the permissions problem again: Copilot surfaces what a user can already open. Matter-level permissions in SharePoint or your practice system must be right before AI is switched on, or information barriers fail quietly.
// More insights
A single commercial screen with a player and installation, a four-screen menu board, a reception video wall: realistic NZ budgets for each,…
Read more →The three laptops we specify for most NZ businesses, the spec floor that stops a machine being replaced in two years, why…
Read more →Hyperfibre sounds like the obvious upgrade. For most offices under 30 people it is the wrong money. How to size a connection…
Read more →// Ready when you are
Tell us what your business runs on today. We will map what DigitalShield would cover, what it would cost per user, and what we would fix first — free, and in plain language.
