AI Agents for Lawyers

Tools

Microsoft Word Legal Agent: Playbook Review and Redlines Inside Word

Microsoft's Legal Agent for Word reviews contracts against your playbook and writes tracked-change redlines within the $30 Copilot add-on. What to check first.

By Dr. Niklas SchmidtUpdated 13 November 20269 min read
On this page
  1. Built by the team Microsoft bought from a failed start-up
  2. How a review runs, click by click
  3. The playbook is the product
  4. The sub-processor question your IT has not answered
  5. Against Spellbook, Claude's commercial plugin and CoCounsel's Word agent
  6. What the benchmarks say about AI redlining
  7. A rollout plan for a contracts team

Microsoft's Legal Agent for Word, released on 30 April 2026 in the Frontier preview, reviews a contract clause by clause against a playbook you supply and writes its proposed changes as tracked changes with explanatory comments, without leaving Word. It costs nothing beyond the $30-a-month Microsoft 365 Copilot add-on, but it runs only in tenants enrolled in Frontier where an administrator has enabled Anthropic models as a sub-processor.

Built by the team Microsoft bought from a failed start-up

The agent has a history worth knowing before you trust it with a contract. Robin AI was for a while the UK's best-funded legal AI start-up, with about $69m raised including a Temasek-led Series B in November 2024. In October 2025 it failed to close a $50m round, cut around 50 roles and listed itself for sale on an insolvency marketplace (Non-Billable). Microsoft acqui-hired its technology team, reported on 9 January 2026 (Artificial Lawyer). Sixteen weeks later that team shipped the Legal Agent (Legal IT Insider, 1 May 2026).

Two things follow. The product is not a Copilot feature bolted on by generalists; it is a contract-review engine that had been sold to in-house teams for years, now distributed through the most widely deployed software in law. And the distribution is the point. ILTA's 2026 survey of 500-plus firms found Microsoft 365 Copilot in use at 76% of them, ahead of CoCounsel and Claude at 44% and Harvey at 43% (LawSites). Adidas's legal operations director Alex Herrity summarised the pitch at launch: "Word is where legal work happens." Freshfields' chief innovation officer Gil Perez had put the same fact less kindly a year earlier, describing legal as a field that "lacks optionality. It's a Microsoft dominated field" (Diginomica).

How a review runs, click by click

The mechanics, as documented by trainers who tested the preview in May 2026 (Chris Menard) and by Microsoft's own guide (Microsoft Support):

  1. Open the contract in Word on Windows and click the Copilot icon.
  2. In the task pane, click the plus sign and choose Legal Agent (Frontier). If it is absent, your tenant is not enrolled or Anthropic models are not enabled; that is an administrator's job, not yours.
  3. Paste your instruction, including the playbook or a reference to it, and the side you act for.
  4. The agent reads the document and returns a comment per clause: the issue, the risk, and replacement wording, each with a Scroll to clause link and Apply or Dismiss buttons.
  5. Work through the comments. Dismiss what you disagree with; apply what you accept.
  6. A follow-up prompt ("convert the applied suggestions to tracked changes with a short comment on each") writes the redline.
  7. Microsoft describes a "purpose-built insertion algorithm" that preserves lists, tables and existing tracked changes, the point at which earlier Word add-ins most often mangled a document.

The output is a Word document with tracked changes and comments, which is what the other side expects and what your DMS versioning handles. That is a real advantage over tools that redline in their own interface and export.

The playbook is the product

A playbook-driven reviewer without a playbook is a general model with a Word plug-in. The contract playbook glossary entry defines the artefact; the playbook automation guide shows how to build one for a whole contract type. For the Legal Agent a working excerpt looks like this, and it should exist in Word before the agent is enabled:

ClausePreferred position (we are customer)FallbackEscalate to partner ifComment to counterparty
Limitation of liabilityCap at 12 months' fees; carve-outs for confidentiality, data protection, IP indemnityCap at fees paid in prior 12 months with the same carve-outsSupplier seeks cap below 6 months' fees or removes data-protection carve-out"Cap aligned to our standard; carve-outs reflect statutory exposure."
Governing law and forumEngland and Wales; English courtsIreland; Irish courtsAny US state law or arbitration"Our group standard; no substantive change to your position."
AssignmentSupplier may not assign without consent; we may assign intra-groupMutual consent not to be unreasonably withheldSupplier demands free assignment to any affiliate"Consent standard is mutual."
Termination for convenienceCustomer on 30 days' notice, no fee60 days with pro-rated refund of prepaid feesAny early-termination fee"Reflects the subscription nature of the service."
Data processingOur DPA attached; supplier is processor; EU sub-processor list; 30-day notice of changesSupplier DPA if it meets Art. 28 GDPR with objection rightNo sub-processor notification right"Required under Art. 28 GDPR."

Then the instruction to paste, which does the work of telling the agent what it may and may not do:

You are reviewing this supplier services agreement for the customer.
Apply the attached playbook. For each clause the playbook covers:
state whether the draft meets the preferred position, the fallback,
or neither; if neither, propose replacement wording taken from the
playbook, not drafted freely; add the playbook's "comment to
counterparty" text as the comment. For any clause where the playbook
says "escalate", do not propose wording; insert a comment beginning
ESCALATE and summarise the gap in one sentence. Do not touch clauses
the playbook does not cover except to flag them as UNREVIEWED. Do not
cite statutes or cases. Preserve defined terms and numbering.

The two prohibitions in that prompt, no free drafting and no citations, are what keep the output checkable in minutes rather than hours. Anthropic's own commercial-legal plugin takes the same approach, with a RED, YELLOW and GREEN triage against the firm's positions before any wording is proposed (GitHub).

The sub-processor question your IT has not answered

The Legal Agent "requires Anthropic enabled as a subprocessor", and one commentator observed that this is a decision "firm IT has not formally evaluated" in most tenants (Platinum IDS). It is worth pausing on. Microsoft 365 Copilot runs under Microsoft's data processing agreement with enterprise data protection and Purview controls and does not train on prompts; consumer Copilot runs under consumer terms with no DPA and is, in Spellbook's blunt assessment, "not suitable for confidential or legal work" (Spellbook, June 2026). The Legal Agent sits on the enterprise side of that line, and then adds a further processor.

Four items for the vendor file, following the vendor due diligence checklist:

  • Where Anthropic processes the documents and for how long it retains them under Microsoft's arrangement, which is not the same as Anthropic's direct API terms.
  • Whether your client engagement terms or outside-counsel guidelines list permitted AI sub-processors; several corporate clients now require the tool to be named.
  • Whether the Frontier preview terms differ from the general Copilot terms; previews often do.
  • Whether Purview audit captures Legal Agent activity in the same way as other Copilot use.

Security is the fifth. Copilot has had two serious injection findings: EchoLeak (CVE-2025-32711, CVSS 9.3) exfiltrated Microsoft 365 Copilot data with no click (Cycode), and a further vulnerability reported on 18 August 2026, CoSnitch, was said to auto-execute attacker prompts in Copilot. A contract sent by a counterparty is untrusted input by definition; the agent security guide explains why a reviewer that only proposes wording, and never sends anything, is the right shape of tool to expose to it.

Against Spellbook, Claude's commercial plugin and CoCounsel's Word agent

Microsoft Legal AgentSpellbookClaude commercial-legal pluginCoCounsel Word drafting agent
Price, September 2026Within the $30 Copilot add-onAbout $180 per user per month; tiers roughly $99–350Free with a paid Claude planWithin CoCounsel Legal; published range $104–639 per user per month
Where it runsWord on Windows, Frontier tenantsWord add-in, any platformCowork, Claude Code, Managed Agents, Microsoft 365CoCounsel Legal, rebuilt on the Claude Agent SDK, GA in the US 20 August 2026
Core jobPlaybook review to tracked changesDrafting, review, redlining, clause library; autonomous contract management in early accessMSA, NDA and SaaS review to tracked changes; renewal tracker; cold-start interviewDrafting grounded in Westlaw and Practical Law; Deep Research Verify
Playbook handlingYours, supplied per reviewBuilt into the productWritten to a practice profile after an interviewFirm content via Workspaces
Verification of lawNone; do not ask it for citationsNone beyond grounding on your documentsCitations tagged [verify] unless a research connector is attachedWestlaw-grounded
Best fitFirms already on Copilot with a written playbook and US paperSmall teams and in-house counsel that live in WordFirms on Claude Team or Enterprise that will maintain the profileFirms with a Westlaw relationship doing US drafting

The Spellbook review and the Claude for Legal plugin guide go deeper on the two closest rivals, and the CoCounsel review on the third. Litera's counter-pitch also deserves a line: its redline engine is deterministic and thirty years old, and its slogan for 2026 is "AI is table stakes. Trust is the moat." For a firm that already owns Copilot and does mostly inbound review of supplier paper against a house position, the Legal Agent is the obvious first tool. For drafting from scratch, non-US law or anything needing authority, it is not.

What the benchmarks say about AI redlining

Redlining is the one contract task where lawyers still won the last independent head-to-head. In Vals.ai's February 2025 study lawyers scored 79.7% on redlining against Harvey's 65.0%, while AI won document Q&A and summarisation by wide margins (Vals.ai). The models have moved since: GC AI's RedlineBench rose from 47.9 to 57.0 with Claude Fable 5.1 in September 2026 (GC AI), and Anthropic claims Opus 5's first-turn NDA redlines score "nearly double Opus 4.8" (Anthropic). None of those figures is a pass mark. The contract review use case sets out how firms measure their own error rates; the practical rule is that the agent produces the first pass on inbound paper and a lawyer produces every outbound mark-up.

One legal quirk favours Microsoft in US litigation practices: Judges Newman (S.D. Ohio) and Boyko (N.D. Ohio) prohibit generative AI in their courtrooms' filings except legal search engines and Microsoft products, as the standing orders page records. That is not an endorsement of accuracy. It is an accident of drafting that makes a Copilot-based tool easier to use in some courts than a better one.

A rollout plan for a contracts team

  1. Write the playbook first. For one contract type, the table above, extended to every clause you negotiate. Two weeks of a senior associate's time; the playbook guide has the structure.
  2. Resolve the sub-processor question. IT enables Anthropic models in the tenant only after the four vendor-file items are answered and, where outside-counsel guidelines require it, clients are told.
  3. Run it on your own template. Ten completed negotiations you already know the outcome of; compare the agent's comments with what was actually agreed.
  4. Measure checking time. If a lawyer needs longer to review the agent's comments than to review the contract, the playbook is too vague; tighten positions before blaming the model.
  5. Go live on inbound paper only. Supplier and counterparty drafts against your playbook; outbound mark-ups remain lawyer-drafted for at least a quarter.
  6. Log and revisit quarterly. Which comments were dismissed most often tells you where the playbook, not the agent, is wrong.

Module 2 of the AI Lab for Lawyers course covers drafting and document analysis, including running a playbook review with a general model, which is the quickest way to see where a Word-native agent stops and a lawyer starts. The Legal Agent is a $30 first step into agentic contract review for the three-quarters of firms that already pay for Copilot; it is only worth taking once the playbook exists and the sub-processor question has been answered in writing.

Frequently asked questions

How much does the Microsoft Word Legal Agent cost?

There is no separate charge as of September 2026. It is part of the Microsoft 365 Copilot add-on, which costs $30 per user per month on an annual commitment, with access through Microsoft's Frontier preview programme. That makes it roughly a sixth of Spellbook's typical price and a small fraction of Harvey's reported seat cost, though it does far less than either.

Does the Word Legal Agent use Anthropic or OpenAI models?

It requires Anthropic models to be enabled in your Microsoft 365 tenant, which makes Anthropic a sub-processor of any document you review with it. Microsoft's Copilot runs under its enterprise data protection terms and does not train on your prompts, but your vendor file should record the Anthropic sub-processing, its data location and retention, before the agent is switched on for client documents.

Can the Legal Agent redline a contract with tracked changes?

Yes. It first returns comments on each clause with the issue, the risk and proposed replacement wording, each with a scroll-to-clause link and apply or dismiss buttons. A follow-up prompt converts accepted suggestions into tracked changes. Microsoft says its insertion algorithm preserves numbered lists, tables and any tracked changes already in the document, which is the part earlier Word add-ins most often broke.

Is the Word Legal Agent available outside the US?

At launch on 30 April 2026 it was limited to US tenants on Windows. A worldwide expansion of the Frontier preview was reported for mid-June 2026 and general availability was targeted for August, but both rest on secondary reports. Check the Frontier programme status in your own tenant's admin centre rather than relying on any date published elsewhere.

Put this into practice

AI Lab for Lawyers: a live, hands-on course for practising lawyers

4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion