Legal AI has become a partnership-level governance question; Harvey, Legora and Claude made sure of that. Airclerk helps NZ firms put Claude to work on real matters, governed and evidenced, on the systems the firm already runs. We read all 151 skills in Anthropic's open-source Claude for Legal suite and turned that review into a New Zealand install profile: which skills run as they are, which run patched, which never get switched on, and where our own 22 go in their place. The firm ends up running the suite, configured for the jurisdiction it practises in.
Use Copilot? We can implement on your preferred platform.
A chronology records what the documents say happened. It does not tell you which pleaded issue each document bears on, or which proposition has nothing behind it yet. One of the four matter skills added in August 2026, with no counterpart anywhere in Anthropic’s 151. Illustrative matter.
Four propositions under one pleaded issue. One with material both ways, one where the contest turns out to be about authority rather than what was said, one resting on a named inference, and one where the document everybody had been assuming existed was never produced. None of that is a view on strength. It is the difference between knowing your position and remembering it.
Two disciplines hold the output together. Conflicts stay conflicts: the termination letter dates the site meeting 15 May and the minutes say 14 May, so both are cited and neither is adopted. And no test, section or case is stated as settled law anywhere in the matrix. The pleaded repudiation issue routes to the research skills rather than being answered here, and anything with the shape of a hearsay, opinion or privilege problem is flagged for a lawyer rather than decided.
The matrix feeds the witness statements, and twenty-one other skills sit around it →
New Zealand law firms use Anthropic's Claude models across document-heavy legal work such as matter chronologies, disclosure and privilege review, contract review against the firm's playbook, due diligence, witness statements and briefs, Privacy Act access requests and permission-aware precedent search. In legal practice Claude needs to be connected to the document management system, matter files, permissions and lawyer approval steps before it can be used safely on client work, with matter isolation so one client's material is walled off from another's.
The lawyer-referral lists name the Law Societies of Scotland, Northern Ireland, Ireland, Canada and Australia; the New Zealand Law Society is absent. Many suite outputs carry US-style "attorney work product" markings. The operative NZ questions are legal advice, litigation and settlement privilege under the Evidence Act 2006; a label can signal intended confidentiality, but it does not create privilege.
The termination review assumes at-will employment; NZ dismissals need justification and fair process. The patent intake treats the US one-year grace period as a green light; NZ's grace period is newer and narrower, and relying on one can still sink the overseas filings. Contract deadlines roll against US federal holidays, so Matariki and anniversary days silently don't exist. US demand-letter workflows do not map cleanly to NZ letters of demand: settlement privilege and costs strategy, professional-conduct duties and misleading-conduct risk all land differently here.
Deposition prep, in a jurisdiction with no US-style depositions. The New Zealand equivalent is the witness statement, and it has just moved: the High Court (Improved Access to Civil Justice) Amendment Rules 2025 replaced briefs of evidence with witness statements from 1 January 2026, exchanged near the start of a proceeding and confined to the witness's own knowledge. The District Court still takes briefs. A US skill cannot be patched into either. Subpoena triage, where the NZ equivalent depends on context: witness summonses, notices to produce, orders for particular disclosure against a non-party under HCR 8.21, regulator information-gathering notices. A DMCA takedown workflow and an EU AI Act register, neither of which is NZ law unless a client's overseas footprint brings it into scope.
The Airclerk NZ Law Pack: 22 New Zealand skills for employment, privacy, companies, litigation, private client and regulatory work, written to the suite's own structure and QA rubric. Eleven rebuild a deliverable outright: Employment Relations Act dismissal reviews, IPP 6 access requests on the 20-working-day clock, Companies Act 1993 board and shareholder resolutions, witness statements and briefs, letters of demand with the open or without-prejudice call flagged for the lawyer. Four read the matter file itself and five do research, both covered below. One runs the regulatory watch, and one interviews the firm on the way in and writes the practice profile every other skill reads. Every skill is source-linked with verification dates to current primary and approved materials: legislation, court rules, practice notes, regulator guidance and case-law references where the firm approves them. The pack is an accelerator, not a finished product: it is where an implementation starts, and each firm's version diverges from there as skills are adapted to the firm's own precedents, playbook and risk settings. The schedule behind it covers all 151: the original skill, the US assumption, the NZ risk, what replaces or patches it, and the verification date. That schedule is what a firm's install is built from, and it is covered next.
The pack was never built to run alone. Most of Anthropic's suite is jurisdiction-neutral plumbing that would be daft to rewrite, and the parts that are wrong are wrong in three different ways that need three different answers. What a New Zealand firm needs is a decision on each of the 151, which is what the review produced.
Portfolio and workspace plumbing, tabular review, the builder-hub authoring layer, most of the study skills. None of it turns on a point of law, so none of it needs a New Zealand version. Rewriting these would be work for its own sake, and the builder-hub layer is the rubric our own pack is written to.
The engine is sound and the content is American: the wrong statutes in a delta checklist, deadlines rolled against US federal holidays, a US privilege header on the output. These run under the New Zealand practice profile, which carries the suite-wide substitutions in one place rather than skill by skill: Evidence Act privilege wording, the New Zealand Law Society referral, the NZ holiday calendar including Matariki and anniversary days, New Zealand research sources, and Law Style Guide citation form.
Where the underlying artefact or statute has no New Zealand counterpart, patching is not available. The termination review assumes at-will employment. The DSAR response grants rights the Privacy Act does not create. Deposition prep prepares for a procedure that does not exist here. These are disabled during setup, and that is the point of the exercise: a skill that is not installed cannot be run by accident at five o'clock on a Friday.
Most of the disabled slots get a New Zealand skill in their place: ERA dismissal reviews for the at-will one, IPP 6 access requests for the DSAR one, witness statements for deposition prep, Companies Act 1993 resolutions for Delaware consents. A few get nothing, because the New Zealand answer turns on context rather than a single document, and the schedule says so instead of pretending otherwise. On top of that sit the four matter skills, the five research skills and the profile interview, which have no upstream counterpart at all.
That schedule is the deliverable, not a brochure for the review behind it. It names every skill, its classification, the New Zealand risk of running it unmodified and what replaces or patches it, and it is versioned: the upstream repo moves, so rows are re-verified against the current commit before your firm relies on any of them. Classifications are ours, as at July 2026.
Every skill in the upstream suite starts from a task you have already defined: review this contract, draft this resolution, sort this discovery. The work that comes before that, reading into an inherited file until you know what the matter actually is, has no skill anywhere in the 151. Four of ours do it. They read the matter file rather than the public web, share one extraction vocabulary, cite every factual statement to a document they opened, and route every question of law to the research skills instead of answering it.
For the partner who has just inherited the file, the lawyer back after six months, or anyone with a client meeting in an hour. It reads the file and produces a Matter Brief: what the matter concerns, the parties and the capacities they hold, where things apparently stand, the key chronology, the issues in play, the documents that carry weight, what does not add up, what is missing, and which skill to run next. Limitation periods and deadlines get flagged prominently and are never computed to a final date.
A chronology says what happened. The matrix says what evidence exists for each thing that matters: every issue broken into the factual propositions it turns on, then each proposition mapped to its supporting material, its contrary material and the holes, with every cell cited and every gap named as a gap. It is the working spine for mediation preparation, pleading review and witness interviews. It organises evidence and never weighs it, so there are no findings and no credibility calls in the output.
A trust settled decades ago by another firm, an estate mid-administration, a family business whose succession intentions are spread across wills, deeds and memoranda. The review turns that into people and their capacities, entities and ownership, an instrument chronology, and the key provisions quoted to source rather than recalled, with the points that need a lawyer's attention listed. It is document understanding, not estate planning, and never a view on validity.
A separation file, a contracting-out review, or a dispute where trusts and companies tangle with the pool. Out comes a relationship chronology with disputed dates left disputed, an asset and liability schedule with names attached, the structures involved, each party's position set beside the other's, what the valuations actually cover, and where the record is thin. Entitlement, property classification and the validity of any agreement stay with the lawyer.
Two of these run in areas where the pack's research is deliberately limited and specialist-review flagged: complex trusts and estates, and family law. The reviews still run, because organising the documents is not the same job as advising on them. They say which one they are doing, on the face of the output.
The one deep-research skill in the upstream suite is scoped to US law and Westlaw US, which is the cleanest statement of the problem in the whole review. There is no localising your way out of it. Five of the twenty-two skills in the pack are the New Zealand answer, and they run one method between them.
Every run is routed Lookup, Quick, Standard or Deep from the complexity of the question and how far the firm intends to rely on the answer, and the output names the depth it ran at. A $2,000 dispute does not get a $20,000 memo.
A search result is not a source. Each run separates what it relied on from what it reviewed and set aside, and discloses the leads it never opened at all. legislation.govt.nz refuses a good share of automated requests, so there is an order of fallbacks behind it, and the output names which one supplied the text rather than implying it read the section. This is the failure mode the courts have already dealt with, approached from the other end.
Every research output goes looking for authority that cuts against the position it is building, and records what it found. A question about an earlier version of a statute either confirms the version in force at the relevant time or says it could not, instead of quietly answering with current law. A Bill or a consultation is never described as law.
Public judgment coverage in New Zealand thins out below the senior courts, and NZLII, where Taxation Review Authority decisions usually sit, blocks automated access. The note-up skill approximates from public sources, classifies how later judgments treated the case, and states that limit in the output rather than leaving a lawyer to assume completeness. Client names and matter details never enter a public search query.
The five split by the job in front of you: a question to answer, a single case or provision to status-check before it is cited, matter documents to work through before any conclusion is formed, someone else's draft to check, or a date-bounded sweep of what changed in a window. Every output is draft research for the reviewing lawyer. None of it is advice.
The patterns that produce the cleanest path from pilot to production. Each is a candidate first workflow, and each says where it comes from: a skill from our pack, an upstream skill running under the New Zealand configuration, or connector work.
A source-linked Matter Brief on an inherited file: synopsis, parties and capacities, key chronology, issues, inconsistencies, gaps, and which skill to run next. Orientation before any narrower task, with limitation dates flagged and never computed.
NZ Law PackDe-duplicated, significance-tagged chronologies from the matter file, every entry cited to source. Runs on the upstream skill with the NZ deltas applied: the HCR 8.30 restriction on using disclosed documents for other purposes, and privilege flagged under ss 54 and 56.
Configured suiteEach issue broken into the propositions it turns on, then mapped to supporting material, contrary material and the holes, every cell cited. The working spine for mediation preparation, pleading review and witness interviews.
NZ Law PackDeeds, variations, wills, EPAs and company records reconstructed into capacities, ownership, an instrument chronology and key provisions quoted to source. Document understanding, never planning and never a validity view.
NZ Law PackRelationship chronology with disputed dates preserved, an asset and liability schedule, the structures involved, both parties' positions side by side and what the valuations cover. Entitlement and classification stay with the lawyer.
NZ Law PackFirst-pass sorting toward the affidavit for disclosure under HCR 8.15, or a District Court affidavit of documents, with potential privilege left flagged for a lawyer's call under ss 54 and 56 rather than silently stripped.
NZ Law PackTerm-by-term deviation analysis against the firm's own positions, carrying the NZ deltas: CCLA 2017, Fair Trading Act unfair contract terms, the CGA, and renewal dates that roll on the NZ holiday calendar.
NZ Law PackOne row per document, every cell cited, straight from the data room to a review-ready grid. Runs on the upstream skill with the NZ issue set substituted: ComCom clearance, OIO consent, Companies Act ss 110 to 111, the PPSR.
Configured suiteERA 2000 s 103A justification and fair process, personal-grievance exposure, redundancy consultation and Holidays Act final pay. The at-will version it replaces is switched off at install.
NZ Law PackERA s 65 mandatory terms, trial-period validity including who may lawfully be on one, and whether a restraint is reasonable enough to hold.
NZ Law PackChronologies, document references and inconsistency checks from the witness's instructions, on the right form for the forum: a High Court witness statement confined to personal knowledge since 1 January 2026, or a District Court brief. Evidential issues are flagged for counsel, and the evidence stays settled by the witness and lawyer.
NZ Law PackIPP 6 access requests on the 20-working-day clock, with ss 49 to 53 withholding grounds analysed document by document.
NZ Law PackWhether a breach is notifiable, what goes to the OPC through NotifyUs and what goes to affected individuals, with the assessment reasoning kept on file.
NZ Law PackCompanies Act 1993 Sch 3 cl 3 and s 122 resolutions with interested-director disclosure handled, in place of the Delaware written-consent template.
NZ Law PackThe open, without-prejudice or Calderbank call flagged for the lawyer rather than made for them, the NZ costs regime applied, and a Fair Trading Act risk check on IP demands.
NZ Law PackStanding monitoring across the NZ Gazette, legislation.govt.nz, select committee submissions and regulator consultations, tracked with owners and deadlines.
NZ Law PackA one-off sweep of what changed in a stated window, by practice area, with a Bill or a consultation never described as law. Deliberately a different skill from the standing watch.
NZ Law PackA question worked at a stated depth across legislation, judgments and regulator material, every claim cited to a source that was opened, with a contrary-authority pass.
NZ Law PackOne case or provision noted up from public sources before it is cited: later citing judgments, how they treated it, and what the public record cannot tell you.
NZ Law PackA junior's memo or opinion checked claim by claim against opened authorities, with issue coverage, case treatment and overstatement flagged. It corrects and never rewrites.
NZ Law PackPermission-aware search across the firm's precedent bank, styles and internal knowledge, scoped so it returns only what the person running it may already see. Connector and permissions work rather than a skill.
ConnectorsBehind every workflow sits AI Process Assurance: the reasoning, sources and sign-off retained as an AI review record, the AI Workpaper, so the firm can show what the AI did, how it was checked and who signed off.
The first production workflow is chosen by volume, source availability, permission complexity and ease of lawyer review. For most firms that means bounded, source-heavy work: matter chronologies, due diligence tables, contract review against the firm's playbook, Privacy Act access-response grids, regulatory watch. Court filings, settlement correspondence, witness evidence and final advice stay behind stricter gates.
Each workflow defines who reviews what: partner, supervising lawyer, privacy officer or practice lead. The AI Workpaper records which sources were attached, what changed in review and whether any required gate was missed.
We run the workflow on held-out examples before a lawyer relies on it: does it cite the right sources, refuse when sources are missing, respect matter permissions and route consequential outputs to the right reviewer. Prompt-injection, cross-matter-leakage and hallucinated-citation checks are part of the production gate.
The configured workflow on your systems, a Workflow Charter setting out stages, evidence and approval gates, the evaluation results from the held-out tests, a runbook with training run separately for lawyers and support staff, a retained AI Workpaper for each material run, and reviews at 30, 60 and 90 days to check the workflow is being used, not just installed.
The workflow you have in mind, the systems it touches and who owns it. Fixed fee quoted after the call.
First governed workflow into production.
Controls, approvals, audit trails.
Monitor, improve, expand.
Everything is deployed to your own Claude instance, under your own agreement with Anthropic.
A Claude implementation plan your partners and risk committee can stand behind, in two weeks.
New Zealand law firms use Claude across document-heavy legal work: matter chronologies, disclosure and privilege review, contract review against the firm's playbook, due diligence, witness statements and briefs, Privacy Act access requests and permission-aware precedent search. Airclerk designs and implements these as governed workflows, connected to the firm's document and practice management systems, matter files, permissions and approval steps, whether that is SharePoint and Microsoft 365, iManage, NetDocuments, Actionstep, LEAP or another stack, so a lawyer reviews and signs off every client-facing, advice, filing, settlement or otherwise consequential output. We are Claude-preferred but platform-aware, and can implement on Copilot where that is your standard.
Claude for Legal is Anthropic's legal offering for Claude, launched in May 2026; its plugin repo is open source under an Apache 2.0 licence, with 12 practice-area plugins, scheduled agents and legal-system connectors. It is well engineered, and almost all of its hardcoded law is American. Airclerk reviewed every skill in the repo (151 as at July 2026) and found New Zealand appears nowhere in it; several defaults would actively mislead an NZ firm. We built an NZ Law Pack to the suite's own structure: 22 New Zealand skills for employment, privacy, companies, litigation, private client and regulatory work, source-linked with verification dates. Eleven rebuild a deliverable the US version gets wrong or has no equivalent for, among them at-will termination review, GDPR-shaped access requests, Delaware-style written consents and deposition prep. Five are New Zealand research skills, added in August 2026: the one deep-research skill in the upstream suite is explicitly scoped to US law and Westlaw US, which is not something a firm can localise its way out of. They route a question by depth, cite only sources they actually opened, run a contrary-authority pass against their own conclusion, and state the limits of public New Zealand case coverage rather than implying a citator. Four more, also added in August 2026, read the matter file rather than the public web: a Matter Brief on an inherited file, an issues and evidence matrix, a trust and estate structure review, and a relationship property review. One runs regulatory monitoring across the NZ Gazette, legislation.govt.nz, select-committee submissions and regulator consultations, and one interviews the firm on the way in and writes the practice profile every other skill reads. The pack is an accelerator rather than a product: it shortens the path to a firm's own build, and each skill is adapted to the firm's precedents, playbook and risk settings before production use.
Both, and that is deliberate. Most of Anthropic's suite is jurisdiction-neutral plumbing that would be pointless to rewrite, so a New Zealand firm runs the suite with a New Zealand configuration over it rather than a replacement for it. On Airclerk's July 2026 review of all 151 skills, roughly 55 install untouched (portfolio and workspace plumbing, tabular review, the builder-hub authoring layer, most study skills), roughly 75 install and are patched by the New Zealand practice profile, which carries the suite-wide substitutions in one place: Evidence Act privilege wording, the New Zealand Law Society referral, the NZ holiday calendar including Matariki and anniversary days, New Zealand research sources and Law Style Guide citation form. Roughly 21 are never switched on, because the artefact or statute behind them has no New Zealand counterpart: at-will termination review, GDPR-shaped DSAR response and deposition prep among them. Disabling those is the safety step that matters, since a skill that is not installed cannot be run by accident. Airclerk's NZ Law Pack of 22 skills fills most of the disabled slots, adds the research layer the upstream suite scopes to US law and Westlaw US, and adds a matter layer the suite has no equivalent for at all. That is why the pack is 22 skills rather than 151: it is the New Zealand-specific part, and the rest of the suite does the rest of the work. Classifications are Airclerk's own and are re-verified against the current upstream commit before a firm relies on them, because the repo moves.
It can do the research; a lawyer still has to check it, and the pack is written so that check is possible. The New Zealand courts have published guidelines on generative AI use and have already dealt with AI-invented citations in filings, so the question worth asking is not whether a tool answers but whether you can audit the answer. The NZ Law Pack's five research skills route a question by depth (Lookup, Quick, Standard or Deep) from its complexity and how far the firm intends to rely on it, and name that depth in the output. They cite only sources the run actually opened, separating what was relied on from what was reviewed and set aside, and disclosing leads that were never opened at all. Where legislation.govt.nz refuses an automated request, the output names which fallback supplied the text rather than implying it read the section. Every output runs a contrary-authority pass against its own conclusion, and a point-in-time question either confirms the version of the law in force at the relevant time or says it could not. Public judgment coverage in New Zealand thins out below the senior courts and NZLII blocks automated access, so the note-up skill is a public-source approximation and states that on its face rather than presenting itself as a citator. Everything produced is draft research for the reviewing lawyer, and none of it is advice.
Both, and the second part is the newer half of the pack. Anthropic's suite assumes the task is already defined: review this contract, draft this resolution. Four skills added in August 2026 do the work that comes before that. One produces a Matter Brief on an inherited file: what the matter concerns, the parties and the capacities they hold, the key chronology, the issues in play, the inconsistencies, the gaps, and which skill to run next. One builds an issues and evidence matrix, breaking each issue into the factual propositions it turns on and mapping every proposition to its supporting material, its contrary material and what is missing, for mediation preparation, pleading review and witness interviews. Two run document-understanding reviews on private client files: an existing trust, estate or succession structure reconstructed from deeds, variations, wills, EPAs and company records; and a relationship property file organised into a relationship chronology, an asset and liability schedule, the structures involved, each party's stated position and what the valuations actually cover. All four cite every factual statement to a document that was opened, and all four stop at the point of judgement. They organise and orient rather than conclude: entitlement, property classification, agreement validity and every question of law route to the research skills and the reviewing lawyer. Trust law and family law research remain limited and specialist-review flagged in the pack, and the two private client reviews say so on their face.
No. Every output is a draft for a lawyer's review, with sources cited and uncertainty flagged. The reviewing lawyer takes professional responsibility for anything that leaves the building. The workflow is supervised like delegated work, with added AI-specific controls for source verification, confidentiality, disclosure and audit trail. Airclerk is an AI implementation consultancy, not a law firm, and the workflows we build are designed around that boundary: consequential actions stop for human sign-off, and supervision stays with the firm. We design workflows to support the firm's obligations under the Lawyers and Conveyancers Act 2006 and the Lawyers and Conveyancers Act (Lawyers: Conduct and Client Care) Rules 2008: competence, supervision, confidentiality, privilege, client-care information, billing transparency and duties to the court.
Two protections come before anything else. Matter isolation: one client's material is walled off from another client's work. Permission-aware access: the workflow sees only what the person running it is allowed to see in the firm's existing document stores. Before any client material is connected, we check the vendor terms: whether inputs are used for training, what is retained and for how long, who the subprocessors are, and where data goes offshore (IPP 12). Everything runs on the firm's own Claude instance, under the firm's own agreement with Anthropic; client material goes to Anthropic on the firm's terms, not to Airclerk. Some steps never move into the workflow at all: clearing conflicts, giving undertakings, filing documents, sending settlement offers and crossing information barriers stay with the firm's authorised lawyers. And where people at the firm are already using personal AI accounts on client matters, we treat that as the first thing to fix, not a fact of life: an engagement starts by finding where AI is already in use and moving that work onto governed rails.
Harvey and Legora are strong products, sold as platform subscriptions. Airclerk's approach is different: we implement Claude on the systems the firm already runs, starting from Anthropic's open-source Claude for Legal plugins localised for New Zealand, so the firm controls its configured workflows and prompts, and keeps the record of what the AI did. For many NZ firms the honest comparison is a platform subscription against a build the firm keeps. We think the record of AI-assisted work on client matters is worth owning.
Because the repo is the easy part. What is left is the work that decides whether it sticks: deciding which of the 151 skills go on, which get switched off and which need the New Zealand profile over them, drawing out the firm's playbook and encoding it, connecting the firm's document and practice management systems, designing the approval gates, testing the workflow before a lawyer relies on it, and training people so adoption is deliberate rather than informal. A firm with engineering capacity can do this itself; the repo is Apache 2.0 and our review of it is on this page. Most firms would rather buy the weeks back.