Microsoft Copilot for legal: a briefing for Australian law firm principals

Quick answers

Questions principals are asking

Access and licensing

Since general availability on 15 October 2026, the legal skills reach every user with a commercial Microsoft 365 Copilot licence, rolling out worldwide through mid-December 2026 (MC1509799).

Activation is mostly automatic. The Word plugin is on by default in Word desktop; Copilot detects a legal document and applies legal skills. Your IT administrator can disable it tenant-wide through the policy "Legal Skills in M365 Copilot in Word" in the Microsoft 365 Apps admin center. For Cowork, each user adds "Legal for Cowork", and administrators govern it on the Agent Tools page of the Microsoft 365 admin center.

Does working inside Word really matter?

Yes, and it is the strongest argument for deployment. Most failed legal AI rollouts fail on adoption, not capability: a separate platform is used by one enthusiast and abandoned by everyone else. Copilot for legal puts the assistant in the Copilot pane beside the contract, and its edits land as native Word tracked changes that preserve formatting.

It also works in documents that already carry tracked changes, keeping its proposals separate from the other side's revisions (ComplexDiscovery). For a negotiating lawyer that is the practical difference: the redline, the explanation and the counter-proposal sit in one document.

Is it unique, and what does it connect to?

It is not unique. Spellbook, and several contract tools since, have run inside Word for years. What is new is distribution: Microsoft ships it to every Copilot-licensed lawyer with no separate procurement, inside a security and compliance framework most firms have already approved.

It does not connect to consumer ChatGPT or Claude. Microsoft 365 Copilot runs on OpenAI GPT models and, where enabled, Anthropic Claude models as a Microsoft subprocessor. It does not connect to the consumer ChatGPT or Claude products. Harvey and Legora are separate partner plugins, which need their own subscriptions.

Specialist tools still have a place. Dedicated platforms offer curated research databases, verified citations and deeper workflow. Copilot for legal suits firms whose needs centre on contracts, correspondence and matter documents. Firms with heavy research or litigation workloads will likely keep a specialist tool alongside it.

Whole-of-matter work, knowledge and precedents

Copilot for legal is a document assistant, not a practice system. It does not run matter intake, conflicts, time recording, trust accounting or AML/CTF workflows. Those stay in your practice management system (LEAP, Smokeball, Clio and similar).

Its knowledge is your Microsoft 365 knowledge. Through Work IQ, Copilot reasons over documents a user can already access in SharePoint, OneDrive, Outlook and Teams. Drafting draws on approved precedents stored in SharePoint. LEAP and Smokeball keep matter documents in their own document stores, and Copilot works from Microsoft 365. Unless a document is also held in SharePoint or OneDrive, or a connector or partner plugin brings it in, Copilot will not see it. Firms that want Copilot to draw on matter documents need a controlled process for deciding which copies are held in Microsoft 365 and which system remains authoritative.

In our experience few firms have a clean, owned precedent library. Most hold several versions of each precedent across personal folders. Copilot will draft from an outdated one as readily as from a current one, so the precedent library needs to be cleaned up before any rollout.

Traceability, verification and accountability

Traceability is reasonable at document level. Summaries carry numbered, clickable citations that highlight the source text. Playbook reviews show which rule informed each finding. Every amendment is a tracked change the lawyer must accept or dismiss; nothing is applied without approval. Obligation trackers link back to the clause they came from.

There are gaps. Microsoft's announcement does not explain how Copilot's revisions are attributed in Word's revision history. The accepted change will typically show the accepting lawyer as author, so firms should record AI involvement in their own file notes.

There is no built-in second-model check in the legal skills. Microsoft's own transparency material says outputs, including citations, may be inaccurate, and that Copilot is not jurisdiction-aware by default. Jurisdictional awareness can be created by specifying the SharePoint libraries and sources Copilot may use.

Accountability stays with the practitioner. In NSW, Supreme Court Practice Note SC Gen 23 restricts generative AI in affidavits and expert reports and requires verification of citations in submissions. Treat every authority Copilot names as unverified until it has been checked in Westlaw, Lexis, AustLII or another authoritative source.

Confidentiality, data retention and privilege

1. Microsoft-hosted processing. Prompts, responses and documents are processed within the Microsoft 365 service boundary, respect existing permissions and sensitivity labels, and are not used to train foundation models.
2. Anthropic models. In Australia, Anthropic models have been on by default since January 2026 and are excluded from Microsoft's in-country processing commitment for Australia (IDM).
3. Partner plugins. When Copilot calls Harvey, Legora or another partner, that provider's own terms, permissions and privacy controls apply. Each plugin needs its own due diligence.
4. Retention. Copilot prompts and responses are stored in a hidden folder in the user's mailbox and are searchable by eDiscovery. A "private discussion" about a clause is invisible to the counterparty, but it is a firm record that may be retained, held and discoverable.

Privilege is unsettled. No Australian court has squarely decided the point. Using an enterprise tool under your firm's tenant is materially safer than consumer AI, but firms should not assume privilege is guaranteed. The same applies to AI tools built into practice management systems, which use the same foundation models.

Bilingual contracts

Treat bilingual review as unsupported. Microsoft has not published supported languages for the legal skills or any cross-language comparison feature. In one practitioner's published test, a frontier model missed all nine provisions in which three different English timing standards were rendered by a single Arabic expression. Where a contract has two authentic language versions, a qualified bilingual lawyer must confirm equivalence; Copilot can assist with first-pass translation only.

Cost for firms already on Microsoft 365

A firm on Microsoft 365 Business Premium can add Microsoft 365 Copilot and gain playbook review, precedent drafting and chronology building without a separate legal AI subscription. The caveat is that the value comes from the firm's own playbooks and precedents, which need to be documented. Budget time, not just licences.

Does it end prompt engineering, and is "for legal" just marketing?

It reduces prompting; it does not remove judgment. A skill is a packaged set of instructions and steps. Microsoft says its skills are "more than prompts layered on a general-purpose model" because they combine legal reasoning with Word features such as tracked changes and content controls. The label alone proves nothing, so test it.

The right evaluation method is simple. Take a previously reviewed agreement where you know the correct issues and acceptable amendments. Run it through Copilot with your playbook. Measure issues missed, incorrect findings, correction time and total cost. That gives evidence for your firm, which no vendor statistic can.

What the launch means for a law firm principal

The technology question is only the first layer. A principal is responsible for client outcomes, professional standards, people, profitability and the long-term value of the firm's intellectual capital. Copilot for legal therefore needs to be assessed as part of the firm's operating model: who performs each task, what knowledge they work from, who supervises the result, how the client is charged, and how the firm proves the new process is at least as reliable as the old one.

Three different kinds of value

Pricing, leverage and profitability

An efficiency gain does not automatically become a commercial gain. If an AI-assisted task takes materially less lawyer time, a time-billing firm may simply create fewer billable units unless it changes the way value is packaged or uses released capacity for additional client work. Principals should decide deliberately who receives the benefit of faster production: the client through a lower or fixed fee, the firm through improved margin, or both through a broader and more responsive service.

The strongest economic use cases are not necessarily the tasks with the largest theoretical time saving. They are tasks where the firm can define the expected output, supervise the result efficiently, and convert released lawyer capacity into better advice, faster turnaround, additional volume or a more predictable fee. Pilot economics should measure total lawyer intervention and correction time, not simply the time Copilot takes to generate an answer.

The precedent library becomes AI infrastructure

A precedent library has traditionally been treated as knowledge management. With AI-assisted drafting and playbook review it becomes part of the firm's production infrastructure. An authoritative precedent, approved alternative clause, escalation rule or playbook position can become reusable input into repeated work. The inverse matters just as much: duplicated, obsolete or ambiguously owned precedents can be reproduced faster and at greater scale. Before relying on precedent-driven skills, the firm should identify authoritative sources, archive superseded material, assign an owner, record jurisdiction and practice-area scope, and set a review cycle.

A practical maturity model

A licence can provide access to Level 1. It does not by itself create Levels 2 to 5. Those depend on the firm's knowledge, information architecture, controls and management choices.

Risk-adjusted supervision, not a generic human-in-the-loop rule

The purpose of a risk tier is not to make low-risk output automatic. It is to stop the firm applying the same superficial review rule to everything. Human intervention should rise with the consequence of an error and the difficulty of detecting it.

Client-facing implications

Firms should decide centrally how AI use interacts with client instructions, outside-counsel guidelines, contractual confidentiality requirements and engagement terms. They should also anticipate a commercial question: if routine production becomes faster, some clients will expect the benefit. A stronger value proposition than "we use AI" is that technology helps the firm apply its approved knowledge systematically while substantive legal judgment remains supervised by lawyers.

Training, juniors and the leverage model

Many traditional training tasks are also attractive AI use cases: first-pass reviews, chronology preparation, comparison, basic drafting and proofing. Removing those tasks without redesigning training risks removing part of the apprenticeship through which junior lawyers develop pattern recognition. A better model is deliberate: AI produces or organises a first pass, the junior identifies omissions and verifies the work against source material, and the senior lawyer reviews the judgment calls. The junior's learning objective shifts from producing every word unaided to detecting why a proposed answer is right, wrong or incomplete.

LEAP and Smokeball: the matter-system boundary

For firms using LEAP or Smokeball, the key architectural point is that working in Microsoft Word is not the same as Copilot knowing the whole practice management matter. Both vendors integrate their products with Microsoft applications, but their matter stores remain distinct systems. Principals should map exactly which documents, emails and metadata are presented to Word or Microsoft 365, which remain inside the practice management platform, and which system is the authoritative matter record, before designing Cowork or cross-matter workflows.

LEAP

LEAP documents that its Microsoft 365 integration creates legal templates in Word, populates firm data into content controls, automatically saves the resulting document to the relevant matter, preserves versions, associates emails with matters, integrates Teams calling, and provides a Microsoft-certified Power BI connector. That is highly complementary to Copilot for legal at the document layer: a lawyer can work in Word on a matter document while LEAP continues to handle matter-centric document management and practice functions.

That integration should not be treated as evidence that Copilot for legal can reason across the complete LEAP matter. Microsoft's current description says Copilot for legal reasons across Microsoft 365 context and supported third-party sources. The sources reviewed for this briefing do not document a first-party Copilot for legal connector giving Cowork general access to a LEAP matter store. Until such access is documented and tested, the safe design assumption is that LEAP remains a separate matter system and that a document being open in Word does not make the rest of the LEAP matter available to Copilot.

The practical implication is a two-system design question. If the firm's authoritative precedents live only in LEAP while playbook review expects a supplied playbook or document context in Copilot, the firm must decide where its authoritative AI-ready knowledge will live and how duplication will be controlled. Do not create an unmanaged shadow precedent library in SharePoint merely to feed AI. If duplication is necessary, assign ownership and versioning, and set a rule that identifies the authoritative source.

Smokeball

Smokeball documents a Word integration that lets users draft, edit and save documents directly to a Smokeball matter. Its Grow and Prosper+ desktop offerings use desktop Word and include document automation and AutoTime; Smokeball Boost uses Microsoft 365 web applications and its Word add-in. Smokeball also markets Archie, its own AI assistant for Microsoft Word, which can read and change the current document and keep the work connected to Smokeball matters.

That creates a more explicit overlap question than LEAP. A Smokeball firm may have Copilot for legal and Smokeball Archie available in the same drafting environment. Principals should define which assistant is approved for which task rather than letting fee earners choose casually. Compare the products on the firm's own documents for accuracy, matter context, saving and version behaviour, confidentiality terms, supervision, cost, and whether time capture or other Smokeball-specific functions matter to the workflow.

There is also a client-version issue. Microsoft's legal plugin is currently a Word desktop capability, while Smokeball Boost is documented as a web-oriented product using Word on the web. A Boost firm's existing web workflow should not be assumed to provide the same Copilot for legal experience. Test the firm's actual Smokeball plan, Word client and Microsoft 365 Copilot licence combination before promising a rollout.

LEAP and Smokeball decision matrix

A pilot scorecard for principals

A 90-day pilot before firm-wide rollout

A staged 90-day pilot lets a firm prove value on its own matters before committing every fee earner. The firm-wide decision rests on a benchmark the firm sets in advance, not on vendor statistics. Each phase ends at a gate that principals sign off.

Recommendations for principals

1. Decide your data position first: whether to allow Anthropic models, and whether in-country processing is a client or insurer requirement.
2. Fix the inputs before the tool: one owned precedent library and one written playbook for your most common agreement.
3. Pilot on defined, measurable use cases, benchmarked against matters already completed.
4. Keep verification mandatory and record AI use on the file.
5. Retain specialist research tools where your work depends on verified authorities.
6. Review results, plugins and settings quarterly. The rollout runs to mid-December 2026 and partner plugins are still being added.

How JurisIT helps

JurisIT works with Australian law firms to put AI tools, including Microsoft 365 Copilot, under governance the firm can prove.

• AI Risk Snapshot (free): a 10-minute self-assessment that scores your firm across six areas and shows where your AI use needs attention before Copilot reaches every desk.
• AI Risk Review: a consultant-led review that verifies your Snapshot answers against evidence in your Microsoft 365 environment, including permissions, data settings and how AI use is recorded.
• Staff AI use: policy, approved-tool settings and Microsoft 365 configuration for Copilot, with supervision and sign-off records on AI-assisted work.

See how JurisIT governs staff AI use →