AI Agents for Lawyers

Tools

Spellbook Review: Contract Drafting Inside Word for Smaller Teams

Spellbook review for lawyers: the Word add-in at about $180 a month, 4,000 customers, grounded frontier models and an autonomous agent in early access.

By Dr. Niklas SchmidtUpdated 13 November 20266 min read
On this page
  1. The verdict first: who should buy it
  2. What Spellbook does inside Word
  3. Built on grounding, not fine-tuning
  4. The business behind the product
  5. A 30-minute playbook test before you buy
  6. Limits you should price in
  7. A setup that works

Spellbook is a Word add-in that drafts, reviews and redlines contracts against your playbook, priced at roughly $180 per user per month with no enterprise seat minimum. As of September 2026 it has about 4,000 customers in 80+ countries, more than half of its revenue from in-house teams, and an autonomous review agent in early access. For a small firm or lean legal department that lives in Word, it is the sensible default; for a large firm it is a point solution.

The verdict first: who should buy it

Your situationSpellbook fitBetter alternative
Solo or 2–10 lawyer firm doing commercial contracts in WordStrong; no seat minimum, Word-native, playbook reviewMicrosoft's Legal Agent if you already pay for Copilot and only need review
In-house team of 1–5 handling NDAs, MSAs, vendor paperStrong; this is now Spellbook's core marketGC AI at $500 per user per month if you want research and advice too
Mid-size firm with a contracts group of 20+Workable, but you will bump into the limits belowLegora or Harvey, reportedly from ~$3,000 per user per year and ~$1,000–1,200 per seat per month respectively
Firm already on Harvey or LegoraRedundant; both ship Word add-insUse the platform's add-in
Team that wants a free first stepNot thisClaude for Word, included in Pro, Max, Team and Enterprise plans

The pricing anchors come from a July 2026 comparison of Word add-ins, which put Spellbook at about $99–350 per user per month depending on tier, LegalOn at about $550 per month and Harvey at an estimated $1,200 per seat (Vaquill). The commonly reported figure for a typical seat is about $180 per month. Harvey's and Legora's minimums are secondary-source estimates; neither publishes a price list, as the pricing guide explains.

What Spellbook does inside Word

Spellbook is a Toronto company led by Scott Stevenson. The product lives in the Word task pane. Highlight a clause and it suggests a redraft; run a review and it walks the document against a playbook, flags deviations, proposes tracked-change language and explains the risk in a comment. A second product, "Spellbook Associate", takes multi-step drafting instructions ("turn this term sheet into a first-draft SPA using our precedent") and works through the document rather than one clause at a time. Nestlé is a named customer.

On 30 June 2026 the company opened early access to "Autonomous Contract Management", described as agents that "work autonomously to review and redline documents" through DMS and email integrations, using several LLMs (Legaltech News). That moves Spellbook from an assistant you invoke to a workflow agent that picks up inbound paper on its own. The distinction matters for supervision: an add-in cannot send anything, an agent connected to your inbox can.

Built on grounding, not fine-tuning

Spellbook's architecture is a deliberate bet. Rather than training a legal model, it grounds GPT-5-class frontier models on your document, your playbook and its clause libraries. Stevenson has argued publicly that fine-tuning "encourages hallucinations" (Nonbillable). The trade-off is real in both directions. Grounding means Spellbook improves the day a new frontier model ships and does not carry a stale fine-tune; it also means Spellbook's quality ceiling is whichever lab's model it routes to, and that a competitor with the same model access can catch up. Harvey went the other way in August 2026 with Tenet, its own post-trained open-weight model, and Litera argues its deterministic 30-year-old redline engine beats probabilistic ones for tracked changes. The fine-tuning entry explains what each approach can and cannot do.

The independent evidence on redlining is humbling for all of them. In the February 2025 Vals Legal AI Report, lawyers beat every tool tested on redlining, 79.7% to Harvey's 65.0%, even as the tools won document Q&A and summarisation by wide margins (Vals). Spellbook was not among the four tools tested. GC AI's own RedlineBench score rose from 47.9 to 57.0 when it moved to Claude Fable 5.1 in September 2026, which suggests the underlying models are improving quickly and that redlining remains the hardest contract task for them.

The business behind the product

Spellbook raised a $50 million Series B at a $350 million post-money valuation in October 2025 and, on 4 March 2026, added a $40 million debt facility from RBCx explicitly to fund acquisitions in what it called a consolidating legal AI market (Business Wire). Two things follow for a buyer. First, a vendor with acquisition capital is more likely to survive the shake-out than one hunting for a buyer, as Robin AI was in October 2025 before Microsoft acqui-hired its team. Second, a vendor on an acquisition path will change its product; ask how your playbooks and clause libraries export, and put it in the contract.

The in-house tilt is the other signal. More than half of revenue now comes from legal departments, which explains why the roadmap leans toward intake, triage and autonomous review rather than litigation. If you are a disputes lawyer, this is not your tool.

A 30-minute playbook test before you buy

Vendors will demo on their own NDA. Use yours. Take a live counterparty NDA you have already negotiated, so you know the answer, and a five-line playbook. The contract playbook entry explains the format; the playbook automation guide has fuller examples.

Review the attached NDA against this playbook. For each deviation, insert a
tracked change with our fallback language and add a comment that states
(a) the issue, (b) the risk in one sentence, and (c) whether it is a
walk-away point. Do not change clauses that comply. Do not invent defined
terms. If a clause is ambiguous, comment rather than edit.

Playbook:
1. Term of confidentiality: preferred 3 years; fallback 5 years; never
   perpetual for non-trade-secret information.
2. Governing law: preferred England and Wales; fallback Ireland; walk away
   on any US state other than New York or Delaware.
3. Residual knowledge clause: must be deleted; no fallback.
4. Non-solicitation: strike; fallback 6 months limited to named individuals.
5. Return or destruction: must allow retention of one archival copy and
   automatic backups.

Score it against what your associate actually did on that NDA. The pass criteria: every playbook deviation found; no edits to compliant clauses; no invented defined terms; comments a partner could send to the client unedited. If it misses the residual knowledge clause or rewrites a compliant governing-law clause, that is the limit you will live with every day. Run the same test on Microsoft's Word Legal Agent if you have Copilot, and on Claude's commercial-legal plugin if you have a paid Claude plan; the Spellbook vs Harvey comparison covers the enterprise end.

Limits you should price in

  • Redlining quality tracks the underlying model. Grounding keeps Spellbook current but ties its ceiling to the labs. Expect step changes with each model release, and re-test after them.
  • It is a contracts tool. No research grounding, no citator, no litigation workflows. Do not ask it for case law; the contract drafting use case explains where drafting ends and research begins.
  • Playbooks are only as good as yours. A team without written positions gets generic redlines. Write the playbook first, then buy the tool.
  • Autonomous review needs a gate. The early-access agent can pick up inbound paper and produce redlines; until you have measured error rates on your own contract types, no redline should leave the building without a lawyer's review. Vals' 2025 finding that lawyers still win redlining is the reason.
  • Data terms deserve a read. The legal-tech press has reported "give-to-get" data arrangements in Spellbook's history; ask whether any apply to your plan, get the sub-processor list for each model it routes to, and treat the answers as contract terms. The confidentiality checklist sets out what to ask every vendor.

A setup that works

For a small firm: buy seats for the two or three lawyers who draft most, write a one-page playbook for each of your three most common contract types, and spend the first month using Spellbook only on inbound paper where a lawyer will mark up the output anyway. Measure minutes per review before and after; Spellbook's own audience, small firms and in-house teams, tends to see the gain on high-volume, low-variance paper such as NDAs and vendor terms, which is also where the contract review use case shows agents doing best. Keep the autonomous features switched off until the manual results are boringly consistent. The small-firm AI guide covers what else to run alongside it, and Module 2 of the AI Lab for Lawyers course works through playbook-driven review with Word-native tools so you can run this test with your own paper.

Frequently asked questions

How much does Spellbook cost?

Spellbook is priced per user at roughly $180 per month as of September 2026, with reported plans spanning about $99 to $350 per user per month depending on tier and volume. It is sold by subscription with no published seat minimum, which is why it suits solo practitioners, small firms and lean in-house teams that cannot meet enterprise platform minimums.

Is Spellbook better than Harvey for contracts?

They are different products. Spellbook is a Word add-in focused on drafting, redlining and playbook review, priced for small teams. Harvey is an enterprise platform with research, diligence, Vault and agent building, reportedly around $1,000 to $1,200 per seat per month with a 20-seat minimum. For a small contracts practice Spellbook is usually the better fit; for a large firm running many matter types, Harvey or Legora is.

Does Spellbook train on my contracts?

Spellbook says it grounds frontier models on your documents rather than fine-tuning on them, and it sells primarily to in-house teams under commercial terms. Ask for the data processing agreement, the list of model sub-processors, retention periods and whether any data-sharing arrangement applies to your plan before uploading client documents, and record the answers in your vendor file.

Can Spellbook redline a contract automatically?

Yes, within Word it suggests tracked-change redlines against your playbook and explains each one. Since 30 June 2026 an early-access Autonomous Contract Management product goes further, with agents that pick up contracts from a DMS or inbox, review and redline them across several models, and hand the result to a lawyer. Treat that output as a first pass, not a finished negotiation position.

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